def read(stack):
print('Process to read: %s' % os.getpid())
yolo = YOLO()
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# deep_sort
d = os.path.dirname(__file__)
model_filename = d + '/model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
max_boxs = 0
face = ['17208019']
#目标上一帧的点
history = {}
#id和标签的字典
person = {}
#赋予新标签的id列表
change = []
while True:
if len(stack) != 0:
frame = stack.pop()
t1 = time.time()
frame_count = 0
localtime = time.asctime(time.localtime(time.time()))
utils.draw(frame, line.readline())
# 获取警戒线
transboundaryline = line.readline()
utils.draw(frame, transboundaryline)
img = Image.fromarray(frame)
#img.save('frame.jpg')
'''
cv2.line(frame, (837, 393), (930, 300), (0, 255, 255), 3)
transboundaryline = t.line_detect_possible_demo(frame)
'''
# image = Image.fromarray(frame)
image = Image.fromarray(frame[..., ::-1]) # bgr to rgb
boxs = yolo.detect_image(image)
# print("box_num",len(boxs))
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(
boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
if len(boxs) > max_boxs:
max_boxs = len(boxs)
# Call the tracker
tracker.predict()
tracker.update(detections)
#一帧信息
info = {}
target = []
for track in tracker.tracks:
#一帧中的目标
per_info = {}
if not track.is_confirmed() or track.time_since_update > 1:
continue
if track.track_id not in person:
person[track.track_id] = str(track.track_id)
bbox = track.to_tlbr()
PointX = bbox[0] + ((bbox[2] - bbox[0]) / 2)
PointY = bbox[3]
dis = int(PointX) - 1200
try:
if dis < 15:
if track.track_id not in change:
person[track.track_id] = face.pop(0)
change.append(track.track_id)
except:
print('非法入侵')
#当前目标
if track.track_id not in change:
per_info['worker_id'] = 'unknow' + str(track.track_id)
else:
per_info['worker_id'] = person[track.track_id]
#当前目标坐标
yoloPoint = (int(PointX), int(PointY))
per_info['current_point'] = yoloPoint
# 卡尔曼滤波预测
if per_info['worker_id'] not in utils.KalmanNmae:
utils.myKalman(per_info['worker_id'])
if per_info['worker_id'] not in utils.lmp:
utils.setLMP(per_info['worker_id'])
cpx, cpy = utils.predict(yoloPoint[0], yoloPoint[1],
per_info['worker_id'])
if cpx[0] == 0.0 or cpy[0] == 0.0:
cpx[0] = yoloPoint[0]
cpy[0] = yoloPoint[1]
if frame_count > 20:
per_info['next_point'] = (int(cpx), int(cpy))
else:
per_info['next_point'] = yoloPoint
# 写入越线情况
if per_info['worker_id'] in history:
per_info['transboundary'] = 'no'
#print(transboundaryline)
line1 = [
per_info['next_point'], history[per_info['worker_id']]
]
a = line.IsIntersec2(transboundaryline, line1)
if a == '有交点':
print('越线提醒')
per_info['transboundary'] = 'yes'
history[per_info['worker_id']] = per_info['current_point']
frame_count = frame_count + 1
#print(per_info)
#画目标框
#cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
cv2.putText(frame, per_info['worker_id'],
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 255, 0), 2)
target.append(per_info)
info['time'] = localtime
#info['frame'] = str(img.tolist()).encode('base64')
info['frame'] = 'frame'
info['target'] = target
#写入josn
info_json = json.dumps(info)
info_queue.put(info_json)
getInfo(info_queue)
cv2.imshow("img", frame)
key = cv2.waitKey(1) & 0xFF
if key == ord('q'):
break
return tracker
class_names = [c.strip() for c in open('./data/labels/coco.names').readlines()]
yolo = YoloV3(classes=len(class_names))
yolo.load_weights('./weights/yolov3.tf')
max_cosine_distance = 0.5
nn_budget = None
nms_max_overlap = 0.8
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric('cosine',
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
vid = cv2.VideoCapture('./data/video/' + name + '.mp4')
# get video dimensions
width = 0
height = 0
if vid.isOpened():
width = vid.get(3) # float `width`
height = vid.get(4) # float `height`
multiTracker = cv2.MultiTracker_create()
def main():
def main(yolo, video_path=0, save_path=None):
# Definition of the parameters
max_cosine_distance = 0.5
nn_budget = None
# openpose
w, h = model_wh('0x0')
model = 'mobilenet_thin'
config = tf.ConfigProto(device_count={'gpu': 0})
config.gpu_options.per_process_gpu_memory_fraction = 0.3
e = TfPoseEstimator(get_graph_path(model),
target_size=(64, 64),
tf_config=config)
# deep_sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
video = Video(video_path)
# video_capture = cv2.VideoCapture() # 'data/person.mp4'
w = video.w
h = video.h
tracker = Tracker(metric,
img_shape=(w, h),
max_eu_dis=0.1 * np.sqrt((w**2 + h**2)))
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
out = cv2.VideoWriter(save_path, fourcc, 7, (w, h))
list_file = open(output_dir + 'detection.txt', 'w')
frame_index = -1
fps = 0.0
while True:
ret, frame = video.video_capture.read() # frame shape 640*480*3
if ret is not True or frame is None:
break
t1 = time.time()
image = Image.fromarray(frame)
boxes, scores, _ = yolo.detect_image(image)
# start = time.time()
features = encoder(frame, boxes) # 提取到每个框的特征
# end = time.time()
# print(end-start)
detections = [
Detection(bbox_and_feature[0], scores[idx], bbox_and_feature[1])
for idx, bbox_and_feature in enumerate(zip(boxes, features))
] # 保存到一个类中
# Call the tracker
tracker.predict()
tracker.update(detections, np.asarray(image))
humans = get_keypoints(image, e)
frame = draw_humans(image, humans, imgcopy=False)
for index, track in enumerate(tracker.tracks):
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
cv2.putText(frame, "{}".format(track.track_id),
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 255, 0), 2)
# for det in detections:
# bbox = det.to_tlbr()
# cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), (255, 0, 0), 2)
cv2.imshow('', frame)
if save_path is not None:
# Define the codec and create VideoWriter object
# save a frame
out.write(frame)
frame_index = frame_index + 1
list_file.write(str(frame_index) + ' ')
if len(boxes) != 0:
for i in range(0, len(boxes)):
list_file.write(
str(boxes[i][0]) + ' ' + str(boxes[i][1]) + ' ' +
str(boxes[i][2]) + ' ' + str(boxes[i][3]) + ' ')
list_file.write('\n')
fps = (fps + (1. / (time.time() - t1))) / 2
# print("fps= %f" % fps)
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
video.video_capture.release()
if save_path is not None:
out.release()
list_file.close()
cv2.destroyAllWindows()
os.path.dirname(__file__))) + os.sep + "logs" + os.sep
make_dir(log_file_folder)
log_file_str = log_file_folder + os.sep + "text.log"
th = handlers.TimedRotatingFileHandler(filename=log_file_str,
when='H',
encoding='utf-8')
th.setFormatter(format_str)
logger.addHandler(sh)
logger.addHandler(th)
if __name__ == '__main__':
# Deep SORT 跟踪器
encoder = generate_detections.create_box_encoder(ARGS.model_feature,
batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
ARGS.min_score, None)
tracker = Tracker(metric)
# 载入模型
yolo = YOLO4(ARGS.model_yolo, ARGS.min_score)
# 读取摄像头实时图像数据、或视频文件
try:
video = cv2.VideoCapture(int(ARGS.video))
except:
video = cv2.VideoCapture(ARGS.video)
# 输出保存视频
fourcc = cv2.VideoWriter_fourcc(*'XVID')
fps = video.get(cv2.CAP_PROP_FPS)
size = (int(video.get(cv2.CAP_PROP_FRAME_WIDTH)),
def main(yolo):
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# deep_sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
writeVideo_flag = True
video_capture = cv2.VideoCapture(0)
if writeVideo_flag:
# Define the codec and create VideoWriter object
w = int(video_capture.get(3))
h = int(video_capture.get(4))
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
out = cv2.VideoWriter('output.avi', fourcc, 15, (w, h))
list_file = open('detection.txt', 'w')
frame_index = -1
fps = 0.0
while True:
ret, frame = video_capture.read() # frame shape 640*480*3
if ret != True:
break
t1 = time.time()
# image = Image.fromarray(frame)
image = Image.fromarray(frame[..., ::-1]) #bgr to rgb
boxs = yolo.detect_image(image)
# print("box_num",len(boxs))
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 255, 0), 2)
for det in detections:
bbox = det.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 0, 0), 2)
cv2.imshow('', frame)
if writeVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
list_file.write(str(frame_index) + ' ')
if len(boxs) != 0:
for i in range(0, len(boxs)):
list_file.write(
str(boxs[i][0]) + ' ' + str(boxs[i][1]) + ' ' +
str(boxs[i][2]) + ' ' + str(boxs[i][3]) + ' ')
list_file.write('\n')
fps = (fps + (1. / (time.time() - t1))) / 2
print("fps= %f" % (fps))
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
video_capture.release()
if writeVideo_flag:
out.release()
list_file.close()
cv2.destroyAllWindows()
def main(yolo):
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# deep_sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
#video_path ="C:/Users/Rashid Ali/Desktop/Person counting/V05.mp4"
writeVideo_flag = True
video_capture = cv2.VideoCapture(0)
if writeVideo_flag:
# Define the codec and create VideoWriter object
video_width = int(video_capture.get(3))
video_height = int(video_capture.get(4))
video_fps = int(video_capture.get(5))
video_size = (int(video_width), int(video_height))
fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
out = cv2.VideoWriter('output_V09.mp4', fourcc, video_fps, video_size)
list_file = open('detection_v2 .txt', 'w')
frame_index = -1
fps = 0.0
while True:
ret, frame = video_capture.read() # frame shape 640*480*3
if ret != True:
break
t1 = time.time()
# image = Image.fromarray(frame)
image = Image.fromarray(frame[..., ::-1]) #bgr to rgb
boxs = yolo.detect_image(image)
# print("box_num",len(boxs))
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
'''
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])),(255,255,255), 2)
cv2.putText(frame, str(track.track_id),(int(bbox[0]), int(bbox[1])),0, 5e-3 * 200, (0,255,0),2)
'''
person_count = 0
count1 = 0
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
person_count = person_count + 1
track_id = '{} {:.1f}'.format('Track_ID', track.track_id)
count1 = '{} {:.1f}'.format('Total Persons Count', person_count)
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
cv2.putText(frame, str(track_id), (int(bbox[0]), int(bbox[1])), 0,
5e-3 * 100, (0, 255, 0), 2)
cv2.putText(frame, str(count1), (20, 50), 0, 5e-3 * 100, (247, 7, 7),
2)
cv2.putText(frame,
'{:.2f}ms'.format((time.time() - t1) * 1000), (20, 20),
fontFace=cv2.FONT_HERSHEY_COMPLEX,
fontScale=0.5,
color=(247, 7, 7),
thickness=1)
cv2.imshow('Detections Window', frame)
'''
#cv2.namedWindow("Detections Window", cv2.WINDOW_AUTOSIZE)
#cv2.rectangle(frame,(int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])),(247, 7, 7), 1)
#cv2.putText(frame, str(count), (int(bbox[0]), int(bbox[1])), fontFace=cv2.FONT_HERSHEY_COMPLEX,
# fontScale=0.5, color=(247, 7, 7), thickness=1)
person_count=0
count1=0
for det in detections:
person_count = person_count+1
count = '{} {:.1f}'.format('Count', person_count)
count1 = '{} {:.1f}'.format('Total Count', person_count)
bbox = det.to_tlbr()
cv2.rectangle(frame,(int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])),(247, 7, 7), 2)
cv2.putText(frame, str(count), (int(bbox[0]), int(bbox[1])), fontFace=cv2.FONT_HERSHEY_COMPLEX,
fontScale=0.5, color=(247, 7, 7), thickness=2)
'''
if writeVideo_flag:
# save a frame
print("Writing detections in file")
out.write(frame)
frame_index = frame_index + 1
list_file.write(str(frame_index) + ' ')
if len(boxs) != 0:
for i in range(0, len(boxs)):
list_file.write(
str(boxs[i][0]) + ' ' + str(boxs[i][1]) + ' ' +
str(boxs[i][2]) + ' ' + str(boxs[i][3]) + ' ')
list_file.write(str(person_count) + '\n')
fps = (fps + (1. / (time.time() - t1))) / 2
print("fps= %f" % (fps))
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
video_capture.release()
if writeVideo_flag:
out.release()
list_file.close()
cv2.destroyAllWindows()
def Object_tracking(Yolo, video_path, output_path, input_size=416, show=False, CLASSES=YOLO_COCO_CLASSES, score_threshold=0.3, iou_threshold=0.45, rectangle_colors='', Track_only = [], custom_yolo=None, custom_classes=YOLO_CUSTOM_CLASSES, Custom_track_only=[]):
# Definition of the parameters
max_cosine_distance = 0.7
nn_budget = None
#initialize deep sort object
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
times, times_2 = [], []
if video_path:
vid = cv2.VideoCapture(video_path) # detect on video
else:
vid = cv2.VideoCapture(0) # detect from webcam
# by default VideoCapture returns float instead of int
width = int(vid.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(vid.get(cv2.CAP_PROP_FPS))
codec = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(output_path, codec, fps, (width, height)) # output_path must be .mp4
NUM_CLASS = read_class_names(CLASSES)
key_list = list(NUM_CLASS.keys())
val_list = list(NUM_CLASS.values())
# set a bunch of flags and variables for made baskets and possessions
possession = None
possession_list = []
combined_possession_avg = 0.5
total_basket_count=0
basket_frame_list = []
baskets_dict = {"Dark": 0, "Light": 0}
made_basket_first_frame = 0
made_basket_frames = 0
basket_marked = False
if custom_yolo:
NUM_CUSTOM_CLASS = read_class_names(custom_classes)
custom_key_list = list(NUM_CUSTOM_CLASS.keys())
custom_val_list = list(NUM_CUSTOM_CLASS.values())
frame_counter = 0
# loop through each frame in video
while True:
_, frame = vid.read()
try:
first_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
original_frame = cv2.cvtColor(first_frame, cv2.COLOR_BGR2RGB)
frame_counter += 1
except:
break
image_data = image_preprocess(np.copy(first_frame), [input_size, input_size])
#image_data = tf.expand_dims(image_data, 0)
image_data = image_data[np.newaxis, ...].astype(np.float32)
t1 = time.time()
# CUSTOM BLOCK FOR BASKETBALL
if custom_yolo:
if YOLO_FRAMEWORK == "tf":
# use yolo model to make prediction on the image data
custom_pred_bbox = custom_yolo.predict(image_data)
# reshape our data to be in correct form for processing
custom_pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in custom_pred_bbox]
custom_pred_bbox = tf.concat(custom_pred_bbox, axis=0)
# get boxes based on threshhold
custom_bboxes = postprocess_boxes(custom_pred_bbox, original_frame, input_size, 0.3)
# custom_bboxes = nms(custom_bboxes, iou_threshold, method='nms')
# extract bboxes to boxes (x, y, width, height), scores and names
custom_boxes, custom_scores, custom_names = [], [], []
for bbox in custom_bboxes:
if len(Custom_track_only) !=0 and NUM_CUSTOM_CLASS[int(bbox[5])] in Custom_track_only or len(Custom_track_only) == 0:
custom_boxes.append([bbox[0].astype(int), bbox[1].astype(int), bbox[2].astype(int)-bbox[0].astype(int), bbox[3].astype(int)-bbox[1].astype(int)])
custom_scores.append(bbox[4])
custom_names.append(NUM_CUSTOM_CLASS[int(bbox[5])])
# Obtain all the detections for the given frame.
custom_boxes = np.array(custom_boxes)
custom_names = np.array(custom_names)
custom_scores = np.array(custom_scores)
# take note of the highest "scoring" made basket and basketball obj in each frame
highest_scoring_basketball = 0
basketball_box = None
basketball_center = None
highest_scoring_made_basket = 0
made_basket_box = None
for i, bbox in enumerate(custom_bboxes):
# loop through each bounding box to get the "best one" of the frame
# we do this because sometimes our model will detect two, and we know there can only be one
name = custom_names[i]
score = round(custom_scores[i], 3)
if name == 'basketball':
if score > highest_scoring_basketball:
highest_scoring_basketball = score
basketball_box = bbox
if name == 'made-basket':
if score > .85 and score > highest_scoring_made_basket:
highest_scoring_made_basket = score
made_basket_box = bbox
# if it sees a basketball, put a box on it and note the center (for possession)
if basketball_box is not None:
cv2.rectangle(original_frame, (int(basketball_box[0]), int(basketball_box[1])), (int(basketball_box[2]), int(basketball_box[3])), (0,0,255), 1)
cv2.rectangle(original_frame, (int(basketball_box[0]), int(basketball_box[1]-30)), (int(basketball_box[0])+(10)*17, int(basketball_box[1])), (0,0,255), -1)
cv2.putText(original_frame, "basketball" + "-" + str(highest_scoring_basketball),(int(basketball_box[0]), int(basketball_box[1]-10)),0, 0.5, (255,255,255),1)
basketball_center = ( (basketball_box[2]+basketball_box[0])/2, (basketball_box[3]+basketball_box[1])/2 )
if made_basket_box is not None:
# if theres a made basket put the box on it
cv2.rectangle(original_frame, (int(made_basket_box[0]), int(made_basket_box[1])), (int(made_basket_box[2]), int(made_basket_box[3])), (0,255,0), 1)
cv2.rectangle(original_frame, (int(made_basket_box[0]), int(made_basket_box[1]-30)), (int(made_basket_box[0])+(15)*17, int(made_basket_box[1])), (0,255,0), -1)
cv2.putText(original_frame, "made-basket" + " " + str(highest_scoring_made_basket),(int(made_basket_box[0]), int(made_basket_box[1]-10)),0, 0.6, (0,0,0),1)
if made_basket_frames == 0:
# if this is the first frame in the sequence
made_basket_first_frame = frame_counter
# increment a counter for made basket frames
made_basket_frames += 1
# if there were 3 consecuative frames AND we havnt marked the basket yet then lets count it!
if made_basket_frames >= 3 and not basket_marked:
basket_marked = True
basket_frame_list.append(made_basket_first_frame)
if possession:
# record which "team" scored the basket
baskets_dict[possession] += 1
# if no made basket make sure the made basket counter is at zero
else:
# no made basket
made_basket_frames = 0
# 60 frames after a made basket we can reset the "marked basket" flag to False
# in essence this means we start looking for made baskets again
if basket_marked and frame_counter > basket_frame_list[-1] + 60:
basket_marked = False
# END CUSTOM BLOCK
# PRESON PREDICTION and TRACKING BLOCK
if YOLO_FRAMEWORK == "tf":
pred_bbox = Yolo.predict(image_data)
elif YOLO_FRAMEWORK == "trt":
batched_input = tf.constant(image_data)
result = Yolo(batched_input)
pred_bbox = []
for key, value in result.items():
value = value.numpy()
pred_bbox.append(value)
t2 = time.time()
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = postprocess_boxes(pred_bbox, original_frame, input_size, score_threshold)
bboxes = nms(bboxes, iou_threshold, method='nms')
# extract bboxes to boxes (x, y, width, height), scores and names
boxes, scores, names = [], [], []
for bbox in bboxes:
if len(Track_only) !=0 and NUM_CLASS[int(bbox[5])] in Track_only or len(Track_only) == 0:
w = bbox[2].astype(int)-bbox[0].astype(int)
h = bbox[3].astype(int)-bbox[1].astype(int)
if h < height/3 and w < width/4:
if h > 120:
boxes.append([bbox[0].astype(int), bbox[1].astype(int), w, h])
scores.append(bbox[4])
names.append(NUM_CLASS[int(bbox[5])])
# Obtain all the detections for the given frame.
boxes = np.array(boxes)
names = np.array(names)
scores = np.array(scores)
# detect jersey color using the tracked persons bounding box
patches = [gdet.extract_image_patch(frame, box, [box[3], box[2]]) for box in boxes]
color_ratios = [find_color(patch) for patch in patches]
features = np.array(encoder(original_frame, boxes))
# mark the detection
detections = [Detection(bbox, score, class_name, feature, color_ratio) for bbox, score, class_name, feature, color_ratio in zip(boxes, scores, names, features, color_ratios)]
# Pass detections to the deepsort object and obtain the track information.
tracker.predict()
tracker.update(detections)
# Obtain info from the tracks
tracked_bboxes = []
color_ratio_list = []
check_possession = False
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 5:
continue
color_ratio = track.get_color_ratio()
color_ratio_list.append(color_ratio)
bbox = track.to_tlbr() # Get the corrected/predicted bounding box
class_name = track.get_class() #Get the class name of particular object
tracking_id = track.track_id # Get the ID for the particular track
index = key_list[val_list.index(class_name)] # Get predicted object index by object name
tracked_bboxes.append(bbox.tolist() + [tracking_id, index]) # Structure data, that we could use it with our draw_bbox function
# if there is a basketball in the frame, and its "in" a ersons bounding box, check what box it is in for psosession
if basketball_center:
if basketball_center[0] >= bbox[0] and basketball_center[0] <= bbox[2]:
if basketball_center[1] >= bbox[1] and basketball_center[1] <= bbox[3]:
check_possession = True
if color_ratio <= .2:
# light team
possession_list.append(0)
else:
# dark team
possession_list.append(1)
else:
# no basketball in frame
# possession_list.append(-1)
# test_list.pop(0)
pass
# if the ball is in a bounding box, update out possession tracker
if check_possession:
if len(possession_list) > 60:
# this function takes an average of the last 60 posessions marked to determine current position
# it weights the most recent detections more
# this algo is a WIP
possession_list = possession_list[-60:]
# full_avg = sum(possession_list)/len(possession)
last_60_avg = sum(possession_list[-60:])/60
last_30_avg = sum(possession_list[-30:])/30
last_15_avg = sum(possession_list[-15:])/15
last_5_avg = sum(possession_list[-5:])/5
combined_possession_avg = round((last_60_avg + last_30_avg + last_15_avg + last_5_avg)/4,3)
#most_common_possession = stats.mode(possession_list)[0]
else:
combined_possession_avg = round(sum(possession_list)/len(possession_list),3)
# use our possession average to determine who has the ball right now
if combined_possession_avg < 0.5:
possession = "Light"
elif combined_possession_avg > 0.5:
possession = "Dark"
# draw detection on frame
image = draw_bbox(original_frame, tracked_bboxes, color_ratios=color_ratio_list, CLASSES=CLASSES, tracking=True)
t3 = time.time()
times.append(t2-t1)
times_2.append(t3-t1)
times = times[-20:]
times_2 = times_2[-20:]
ms = sum(times)/len(times)*1000
fps = 1000 / ms
fps2 = 1000 / (sum(times_2)/len(times_2)*1000)
if possession == "Light":
image = cv2.putText(image, "Posession: {}".format(possession), (width-400, 30), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (50, 255, 255), 2)
else:
image = cv2.putText(image, "Posession: {}".format(possession), (width-400, 30), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2)
# image = cv2.putText(image, "Light: {} Dark: {} None: {}".format(possession_list.count(0), possession_list.count(1), possession_list.count(-1)), (400, 30), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2)
image = cv2.putText(image, "Posession Avg: {}".format(combined_possession_avg), (400, 30), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2)
image = cv2.putText(image, "Time: {:.1f} FPS".format(fps), (0, 30), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2)
# draw original yolo detection
#image = draw_bbox(image, bboxes, CLASSES=CLASSES, show_label=False, rectangle_colors=rectangle_colors, tracking=True)
print("Time: {:.2f}ms, Detection FPS: {:.1f}, total FPS: {:.1f}".format(ms, fps, fps2))
if output_path != '': out.write(image)
if show:
cv2.imshow('output', image)
if cv2.waitKey(25) & 0xFF == ord("q"):
cv2.destroyAllWindows()
break
cv2.destroyAllWindows()
return_data = {"baskets_dict": baskets_dict, "basket_frame_list": basket_frame_list}
print("video saved to {}".format(output_path))
return(return_data)
# opencv 2
if cv2.__version__.split(".")[0] == "2":
frame_count = vcap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT)
else:
# opencv 3/4
frame_count = vcap.get(cv2.CAP_PROP_FRAME_COUNT)
# initialize tracking module
if args.get_tracking:
tracking_objs = args.tracking_objs.split(",")
tracker_dict = {}
tracking_results_dict = {}
tmp_tracking_results_dict = {}
for tracking_obj in tracking_objs:
metric = metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", args.max_cosine_distance, args.nn_budget)
tracker_dict[tracking_obj] = Tracker(
metric, max_iou_distance=args.max_iou_distance)
tracking_results_dict[tracking_obj] = []
tmp_tracking_results_dict[tracking_obj] = {}
# videoname = os.path.splitext(os.path.basename(videofile))[0]
videoname = os.path.basename(videofile)
video_obj_out_path = None
if args.obj_out_dir is not None: # not saving box json to save time
video_obj_out_path = os.path.join(args.obj_out_dir, videoname)
if not os.path.exists(video_obj_out_path):
os.makedirs(video_obj_out_path)
video_queuer = VideoEnqueuer(args,
vcap,
def main(yolo):
global p1_flag, p2_flag, vline1, mid_line, vline2, lane_1, lane_2, lane_3, lane_4, lane_5, lane_6, mask, preset
global cnt_lane_1, cnt_lane_2, cnt_lane_3, cnt_lane_4, cnt_lane_5, cnt_lane_6, global_point
global speed_lane_1, speed_lane_2, speed_lane_3, speed_lane_4, speed_lane_5, speed_lane_6
################# parameters ######################
track_len = 2
detect_interval = 4
of_track = []
preset = 0
alpha = 0.3
mm1, mm2, mm3, mm4, mm5, mm6 = 0, 0, 0, 0, 0, 0
v1, v2, v3, v4, v5, v6 = 0, 0, 0, 0, 0,0
ptn1, ptn2, ptn3, ptn4, ptn5, ptn6 = 0, 0, 0, 0, 0,0
prv1, prv2, prv3, prv4, prv5, prv6 = 0, 0, 0, 0, 0, 0
ms2kmh = 3.6
fps = 30
max_cosine_distance = 0.5
nn_budget = None
nms_max_overlap = 0.3
counter = []
##################################################
# Deep SORT
model_filename = 'model_data/market1501.pb'
encoder = gdet.create_box_encoder(model_filename,batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
writeVideo_flag = True
#video_path = "./output/output.avi"
video_capture = cv2.VideoCapture(args["input"])
if writeVideo_flag:
# Define the codec and create VideoWriter object
w = int(video_capture.get(3))
h = int(video_capture.get(4))
total_frame = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))
fourcc = cv2.VideoWriter_fourcc(*'FMP4')
video_fps = video_capture.get(cv2.CAP_PROP_FPS)
out = cv2.VideoWriter('output/CH4_output.avi', fourcc, video_fps, (w, h))
list_file = open('detection.txt', 'w')
frame_index = -1
fps = 0.0
frame_idx = 0
speed_dict = OrderedDict()
ret, first_frame = video_capture.read()
cal_mask1 = np.zeros_like(first_frame[:, :, 0])
cal_mask2 = np.zeros_like(first_frame[:, :, 0])
while True:
ret, frame = video_capture.read()
glob = frame.copy()
cmask = frame.copy()
# Channel and preset setting
if ch == 8:
if 0 <= frame_idx <= 480 or 2415 <= frame_idx <= 4203 or 6140 <=frame_idx<=7925 or 9864 <=frame_idx<=11648 or 13585 <= frame_idx <= 15370 or frame_idx >=17306:
preset = 1
elif 559 <= frame_idx <= 2340 or 4275 <= frame_idx<=6064 or 8000 <= frame_idx<=9787 or 11730 <=frame_idx<=13513 or 15450 <=frame_idx<=17237:
preset = 2
else:
preset = 0
if preset == 1:
vline1 = p1_vline1
mid_line = p1_mid_line
vline2 = p1_vline2
lane_1 = p1_lane_1
lane_2 = p1_lane_2
lane_3 = p1_lane_3
lane_4 = p1_lane_4
lane_5 = p1_lane_5
lane_6 = p1_lane_6
global_point = p1_global_point
mask = p1_mask
# Polyline
cv2.polylines(frame, [lane_1], True, (153,255,255))
cv2.polylines(frame, [lane_2], True, (255,204,204))
cv2.polylines(frame, [lane_3], True, (204,255,204))
cv2.polylines(frame, [lane_4], True, (255,204,255))
cv2.polylines(frame, [lane_5], True, (153,153,255))
cv2.polylines(frame, [lane_6], True, (102,255,153))
frame = cv2.line(frame, vline1[0], vline1[1], (0,0,255), 1)
frame = cv2.line(frame, mid_line[0], mid_line[1], (0,0,255), 1)
frame = cv2.line(frame, vline2[0], vline2[1], (0,0,255), 1)
p1_flag = True
view_polygon = np.array([[10, 1920], [380, 250], [680, 250], [1080, 480], [1080, 1920]])
cal_polygon = np.array([[361, 304], [755, 293], [1076, 480], [1077, 1067], [163, 1051]])
pg1 = np.array([[382, 347], [359, 347], [236, 833], [272, 832]]) # RT, LT, LB, RB
pg2 = np.array([[460, 347], [434, 346], [456, 833], [505, 833]]) # LB, RT, LT, RB
pg3 = np.array([[544, 345], [514, 345], [686, 833], [755, 832]]) # LB, LT, RT, RB
pg4 = np.array([[630, 342], [598, 343], [924, 829], [991, 829]]) # LB, LT, LB, RB
pg5 = np.array([[725, 343], [696, 345], [996, 650], [1056, 646]]) # RT, LB, LT, RB
pg6 = np.array([[798, 340], [761, 340], [1037, 535], [1070, 530]]) # RT, LB, LT, RB
cv2.fillConvexPoly(cal_mask1, cal_polygon, 1)
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
frame_gray = cv2.bitwise_and(frame_gray, frame_gray, mask=cal_mask1)
if pg1.size>0:
cv2.fillPoly(cmask, [pg1], (120, 0, 120), cv2.LINE_AA)
if pg2.size>0:
cv2.fillPoly(cmask, [pg2], (120, 120, 0), cv2.LINE_AA)
if pg3.size>0:
cv2.fillPoly(cmask, [pg3], (0, 120, 120), cv2.LINE_AA)
if pg4.size>0:
cv2.fillPoly(cmask, [pg4], (80, 0, 255), cv2.LINE_AA)
if pg5.size>0:
cv2.fillPoly(cmask, [pg5], (255, 0, 80), cv2.LINE_AA)
if pg6.size>0:
cv2.fillPoly(cmask, [pg6], (120, 0, 0), cv2.LINE_AA)
elif preset == 2:
vline1 = p2_vline1
mid_line = p2_mid_line
vline2 = p2_vline2
lane_1 = p2_lane_1
lane_2 = p2_lane_2
lane_3 = p2_lane_3
lane_4 = p2_lane_4
lane_5 = p2_lane_5
lane_6 = p2_lane_6
global_point = p2_global_point
mask = p2_mask
# Polyline
cv2.polylines(frame, [lane_1], True, (153,255,255))
cv2.polylines(frame, [lane_2], True, (255,204,204))
cv2.polylines(frame, [lane_3], True, (204,255,204))
cv2.polylines(frame, [lane_4], True, (255,204,255))
cv2.polylines(frame, [lane_5], True, (153,153,255))
cv2.polylines(frame, [lane_6], True, (102,255,153))
frame = cv2.line(frame, vline1[0], vline1[1], (0,0,255), 1)
frame = cv2.line(frame, mid_line[0], mid_line[1], (0,0,255), 1)
frame = cv2.line(frame, vline2[0], vline2[1], (0,0,255), 1)
p2_flag = True
view_polygon = np.array([[284, 649], [0, 1629], [1076, 1574], [1079, 888], [676, 634]])
cal_polygon = np.array([[896, 778], [244, 794], [105, 1271], [1077, 1245], [1077, 879]])
pg1 = np.array([[276, 846], [234, 847], [134, 1200], [199, 1198]]) # RT, LT, LB, RB
pg2 = np.array([[418, 844], [375, 844], [384, 1196], [442, 1198]]) # LB, RT, LT, RB
pg3 = np.array([[553, 843], [508, 844], [637, 1194], [706, 1194]]) # LB, LT, RT, RB
pg4 = np.array([[686, 841], [637, 843], [886, 1190], [968, 1189]]) # LB, LT, LB, RB
pg5 = np.array([[817, 837], [773, 841], [1005, 1051], [1060, 1047]]) # RT, LB, LT, RB
pg6 = np.array([[966, 837], [919, 840], [1043, 929], [1087, 927]]) # RT, LT, LB, RB
cv2.fillConvexPoly(cal_mask2, cal_polygon, 1)
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
frame_gray = cv2.bitwise_and(frame_gray, frame_gray, mask=cal_mask2)
if pg1.size > 0:
cv2.fillPoly(cmask, [pg1], (120, 0, 120), cv2.LINE_AA)
if pg2.size > 0:
cv2.fillPoly(cmask, [pg2], (120, 120, 0), cv2.LINE_AA)
if pg3.size > 0:
cv2.fillPoly(cmask, [pg3], (0, 120, 120), cv2.LINE_AA)
if pg4.size > 0:
cv2.fillPoly(cmask, [pg4], (80, 0, 255), cv2.LINE_AA)
if pg5.size > 0:
cv2.fillPoly(cmask, [pg5], (255, 0, 80), cv2.LINE_AA)
if pg6.size > 0:
cv2.fillPoly(cmask, [pg6], (120, 0, 0), cv2.LINE_AA)
elif ch == 4:
if 0 <= frame_idx <= 1751 or frame_idx >= 3655:
preset = 1
elif 1797 <= frame_idx <= 3600:
preset = 2
else:
preset = 0
if preset == 1:
lane_1 = p1_lane_1
lane_2 = p1_lane_2
lane_3 = p1_lane_3
lane_4 = p1_lane_4
lane_5 = p1_lane_5
lane_6 = p1_lane_6
global_point = p1_global_point
mask = p1_mask
# Polyline
# cv2.polylines(frame, [lane_1], True, (153,255,255))
# cv2.polylines(frame, [lane_2], True, (255,204,204))
# cv2.polylines(frame, [lane_3], True, (204,255,204))
# cv2.polylines(frame, [lane_4], True, (255,204,255))
# cv2.polylines(frame, [lane_5], True, (153,153,255))
# cv2.polylines(frame, [lane_6], True, (102,255,153))
p1_flag = True
view_polygon = np.array([[731, 563], [385, 567], [33, 1260], [1077, 1254], [1078, 812]])
cal_polygon = np.array([[914, 669],[286, 675], [89, 1083], [1078, 1083], [1078, 772]])
pg6 = np.array([[346, 686], [313, 686], [163, 992], [244, 996]]) # RT, LT, LB, RB
pg5 = np.array([[430, 684], [401, 685], [338, 998], [420, 1000]]) # LB, RT, LT, RB
pg4 = np.array([[534, 685], [506, 685], [547, 999], [631, 999]]) # LB, LT, RT, RB
pg3 = np.array([[654, 685], [609, 684], [760, 1000], [839, 999]]) # LB, LT, LB, RB
pg2 = np.array([[770, 685], [723, 684], [979, 999], [1051, 998]]) # RT, LB, LT, RB
pg1 = np.array([[858, 683], [815, 683], [1031, 860], [1077, 857]]) # RT, LB, LT, RB
cv2.fillConvexPoly(cal_mask1, cal_polygon, 1)
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
frame_gray = cv2.bitwise_and(frame_gray, frame_gray, mask=cal_mask1)
if pg1.size > 0:
cv2.fillPoly(cmask, [pg1], (120, 0, 120), cv2.LINE_AA)
if pg2.size > 0:
cv2.fillPoly(cmask, [pg2], (120, 120, 0), cv2.LINE_AA)
if pg3.size > 0:
cv2.fillPoly(cmask, [pg3], (0, 120, 120), cv2.LINE_AA)
if pg4.size > 0:
cv2.fillPoly(cmask, [pg4], (80, 0, 255), cv2.LINE_AA)
if pg5.size > 0:
cv2.fillPoly(cmask, [pg5], (255, 0, 80), cv2.LINE_AA)
if pg6.size > 0:
cv2.fillPoly(cmask, [pg6], (120, 0, 0), cv2.LINE_AA)
elif preset == 2:
lane_1 = p2_lane_1
lane_2 = p2_lane_2
lane_3 = p2_lane_3
lane_4 = p2_lane_4
lane_5 = p2_lane_5
lane_6 = p2_lane_6
global_point = p2_global_point
mask = p2_mask
# Polyline
# cv2.polylines(frame, [lane_1], True, (153,255,255))
# cv2.polylines(frame, [lane_2], True, (255,204,204))
# cv2.polylines(frame, [lane_3], True, (204,255,204))
# cv2.polylines(frame, [lane_4], True, (255,204,255))
# cv2.polylines(frame, [lane_5], True, (153,153,255))
# cv2.polylines(frame, [lane_6], True, (102,255,153))
p2_flag = True
view_polygon = np.array([[547, 609], [0, 1109], [1, 1271], [1078, 1278], [1079, 594]])
cal_polygon = np.array([[529, 611], [8, 1105], [1077, 1110], [1078, 599]])
pg6 = np.array([[556, 609], [493, 607], [108, 1033], [190, 1030]]) # RT, LT, LB, RB
pg5 = np.array([[693, 604], [642, 602], [356, 1020], [455, 1020]]) # LB, RT, LT, RB
pg4 = np.array([[812, 633], [765, 633], [604, 1026], [702, 1026]]) # LB, LT, RT, RB
pg3 = np.array([[932, 638], [882, 636], [883, 1007], [953, 1001]]) # LB, LT, LB, RB
pg2 = np.array([[1059, 641], [978, 638], [1028, 941], [1079, 916]]) # RT, LB, LT, RB
pg1 = np.array([])
cv2.fillConvexPoly(cal_mask2, cal_polygon, 1)
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
frame_gray = cv2.bitwise_and(frame_gray, frame_gray, mask=cal_mask2)
if pg1.size > 0:
cv2.fillPoly(cmask, [pg1], (120, 0, 120), cv2.LINE_AA)
if pg2.size > 0:
cv2.fillPoly(cmask, [pg2], (120, 120, 0), cv2.LINE_AA)
if pg3.size > 0:
cv2.fillPoly(cmask, [pg3], (0, 120, 120), cv2.LINE_AA)
if pg4.size > 0:
cv2.fillPoly(cmask, [pg4], (80, 0, 255), cv2.LINE_AA)
if pg5.size > 0:
cv2.fillPoly(cmask, [pg5], (255, 0, 80), cv2.LINE_AA)
if pg6.size > 0:
cv2.fillPoly(cmask, [pg6], (120, 0, 0), cv2.LINE_AA)
if ret != True:
break
t1 = time.time()
cnt_lane_1 = cnt_lane_2 = cnt_lane_3 = cnt_lane_4 = cnt_lane_5 = cnt_lane_6 = 0
image = Image.fromarray(frame[...,::-1]) #bgr to rgb
boxs,class_names = yolo.detect_image(image)
#features is 128-dimension vector for each bounding box
features = encoder(frame,boxs)
# score to 1.0 here.
# score 1.0 means that bbox's class score after yolo is 100%
detections = [Detection(bbox, 1.0, feature) for bbox, feature in zip(boxs, features)]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap, scores)
# detections is object detection result at current frame
# it may contain many bbox or no bbox
detections = [detections[i] for i in indices]
# calculate lane by lane avg speed
## swhan method (optical flow)
mm1, mm2, mm3, mm4, mm5, mm6 = 0, 0, 0, 0, 0,0
if len(of_track) > 0:
img0, img1 = prev_gray, frame_gray
p0 = np.float32([tr[-1] for tr in of_track]).reshape(-1, 1, 2)
p1, _st, _err = cv2.calcOpticalFlowPyrLK(img0, img1, p0, None, **lk_params)
p0r, _st, _err = cv2.calcOpticalFlowPyrLK(img1, img0, p1, None, **lk_params)
d = abs(p0-p0r).reshape(-1, 2).max(-1)
good = d < 1
new_of_tracks = []
for tr, (x, y), good_flag in zip(of_track, p1.reshape(-1, 2), good):
if not good_flag:
continue
tr.append((x, y))
if len(tr) > track_len:
del tr[0]
new_of_tracks.append(tr)
cv2.circle(frame, (x, y), 3, (0, 255, 0), -1)
of_track = new_of_tracks
for idx, tr in enumerate(of_track):
#print(frame_idx, tr)
if pg1.size > 0:
result_pg1 = cv2.pointPolygonTest(pg1, tr[0],True)
else:
result_pg1 = -999
if pg2.size > 0:
result_pg2 = cv2.pointPolygonTest(pg2, tr[0], True)
else:
result_pg2 = -999
if pg3.size > 0:
result_pg3 = cv2.pointPolygonTest(pg3, tr[0], True)
else:
result_pg3 = -999
if pg4.size > 0:
result_pg4 = cv2.pointPolygonTest(pg4, tr[0], True)
else:
result_pg5 = -999
if pg5.size > 0:
result_pg5 = cv2.pointPolygonTest(pg5, tr[0], True)
else:
result_pg5 = -999
if pg6.size > 0:
result_pg6 = cv2.pointPolygonTest(pg6, tr[0], True)
else:
result_pg6 = -999
if frame_idx % detect_interval == 0:
if result_pg1 > 0:
ptn1 += 1
if preset == 1:
mm1 += convPtoR_1(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
elif preset == 2:
mm1 += convPtoR_2(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
mmm1 = mm1/ptn1
v1 = mmm1*fps*ms2kmh*6
if result_pg2 > 0:
ptn2 += 1
if preset == 1:
mm2 += convPtoR_1(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
elif preset == 2:
mm2 += convPtoR_2(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
mmm2 = mm2/ptn2
v2 = mmm2*fps*ms2kmh*6
if result_pg3 > 0:
ptn3 += 1
if preset == 1:
mm3 += convPtoR_1(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
elif preset == 2:
mm3 += convPtoR_2(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
mmm3 = mm3/ptn3
v3 = mmm3*fps*ms2kmh*6
if result_pg4 > 0:
ptn4 += 1
if preset == 1:
mm4 += convPtoR_1(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
elif preset == 2:
mm4 += convPtoR_2(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
mmm4 = mm4/ptn4
v4 = mmm4*fps*ms2kmh*6
if result_pg5 > 0:
ptn5 += 1
if preset == 1:
mm5 += convPtoR_1(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
elif preset == 2:
mm5 += convPtoR_2(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
mmm5 = mm5/ptn5
v5 = mmm5*fps*ms2kmh*6
if result_pg6 > 0:
ptn6 += 1
if preset == 1:
mm6 += convPtoR_1(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
elif preset == 2:
mm6 += convPtoR_2(tr[0][0],tr[0][1],tr[1][0],tr[1][1])
mmm6 = mm6/ptn6
v6 = mmm6*fps*ms2kmh*6
mask = np.zeros_like(frame_gray)
mask[:] = 255
for x, y in [np.int32(tr[-1]) for tr in of_track]:
cv2.circle(mask, (x, y), 3, 0, -1)
p = cv2.goodFeaturesToTrack(frame_gray, mask = mask, **feature_params)
if p is not None:
for x, y in np.float32(p).reshape(-1, 2):
of_track.append([(x, y)])
prev_gray = frame_gray
## swhan method
if frame_idx % detect_interval == 0:
if ptn1 > 0:
avg_speed_lane_1 = v1
prv1=v1
elif ptn1 == 0:
avg_speed_lane_1 = 0
prv1 = 0
if ptn2 > 0:
avg_speed_lane_2 = v2
prv2=v2
elif ptn2 == 0:
avg_speed_lane_2 = 0
prv2 = 0
if ptn3 > 0:
avg_speed_lane_3 = v3
prv3=v3
elif ptn3 == 0:
avg_speed_lane_3 = 0
prv3 = 0
if ptn4 > 0:
avg_speed_lane_4 = v4
prv4=v4
elif ptn4 == 0:
avg_speed_lane_4 = 0
prv4 = 0
if ptn5 > 0:
avg_speed_lane_5 = v5
prv5=v5
elif ptn5 == 0:
avg_speed_lane_5 = 0
prv5 = 0
if ptn6 > 0:
avg_speed_lane_6 = v6
prv6=v6
elif ptn6 == 0:
avg_speed_lane_6 = 0
prv6 = 0
else:
avg_speed_lane_1 = prv1
avg_speed_lane_2 = prv2
avg_speed_lane_3 = prv3
avg_speed_lane_4 = prv4
avg_speed_lane_5 = prv5
avg_speed_lane_6 = prv6
ptn1, ptn2, ptn3, ptn4, ptn5, ptn6 = 0, 0, 0, 0, 0, 0
# Call the tracker
tracker.predict()
tracker.update(detections)
i = int(0)
indexIDs = []
c = []
boxes = []
for det in detections:
bbox = det.to_tlbr()
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
indexIDs.append(int(track.track_id))
counter.append(int(track.track_id))
bbox = track.to_tlbr()
color = [int(c) for c in COLORS[indexIDs[i] % len(COLORS)]]
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])),(color), 3)
if len(class_names) > 0:
class_name = class_names[0]
i += 1
center = (int(((bbox[0])+(bbox[2]))/2),int(((bbox[1])+(bbox[3]))/2))
# global point matching
track.matching_point[0] = center[0]
track.matching_point[1] = center[1]
temp_list = [track.matching_point[0],track.matching_point[1],frame_idx]
if len(track.matching_point_list) == 4:
track.matching_point_list.append(temp_list)
x1 = track.matching_point_list[0][0]
y1 = track.matching_point_list[0][1]
x2 = track.matching_point_list[-1][0]
y2 = track.matching_point_list[-1][1]
# If the pixel don't change, the speed should be zero.
if x1 == x2 and y1 == y2:
track.matching_point_list.pop(0)
else:
time1 = track.matching_point_list[0][2]
time2 = track.matching_point_list[-1][2]
if preset == 1:
R_dist1 = convPtoR_1(x1,y1,x2,y2)
t_time1 = (time2-time1)/30
speed = int(3.6*R_dist1 //t_time1)
#print("time1 : ",time1, "time2 : ",time2)
elif preset == 2:
R_dist1 = convPtoR_2(x1,y1,x2,y2)
t_time1 = (time2-time1)/30
speed = int(3.6*R_dist1 //t_time1)
#print("time1 : ",time1, "time2 : ",time2)
track.matching_point_list.pop(0)
if frame_idx % 6 ==1 :
track.speed = speed
cv2.putText(frame, str(int(track.speed))+'km/h', (int(bbox[0]), int(bbox[1]+((bbox[3]-bbox[1])/2))),cv2.FONT_HERSHEY_SIMPLEX, 1, (255,255,255), 4)
elif len(track.matching_point_list) == 0 :
track.matching_point_list.append(temp_list)
elif track.matching_point_list[-1][0] != track.matching_point[0] and track.matching_point_list[-1][1] != track.matching_point[1]:
track.matching_point_list.append(temp_list)
cv2.circle(frame, (track.matching_point[0], track.matching_point[1]), 5, (0,0,255),-1)
# traffic lane by lane
if cv2.pointPolygonTest(lane_1, center, False) >= 0:
cnt_lane_1 += 1
track.driving_lane = 1
elif cv2.pointPolygonTest(lane_2, center, False) >= 0:
cnt_lane_2 += 1
track.driving_lane = 2
elif cv2.pointPolygonTest(lane_3, center, False) >= 0:
cnt_lane_3 += 1
track.driving_lane = 3
elif cv2.pointPolygonTest(lane_4, center, False) >= 0:
cnt_lane_4 += 1
track.driving_lane = 4
elif cv2.pointPolygonTest(lane_5, center, False) >= 0:
cnt_lane_5 += 1
track.driving_lane = 5
elif cv2.pointPolygonTest(lane_6, center, False) >= 0:
cnt_lane_6 += 1
track.driving_lane = 6
cv2.putText(frame, "ID:"+str(track.track_id) + "/"+ str(track.driving_lane), (int(bbox[0]), int(bbox[1])-20),cv2.FONT_HERSHEY_SIMPLEX, 1, color, 2)
pts[track.track_id].append(center)
ctime[track.time].append(time.time())
thickness = 5
#center point
cv2.circle(frame, (center), 1, color, thickness)
#draw motion path
for j in range(1, len(pts[track.track_id])):
if pts[track.track_id][j - 1] is None or pts[track.track_id][j] is None:
continue
thickness = int(np.sqrt(64 / float(j + 1)) * 2)
cv2.line(frame,(pts[track.track_id][j-1]), (pts[track.track_id][j]),(color),thickness)
# cv2.putText(frame, str(class_names[j]),(int(bbox[0]), int(bbox[1] -20)),0, 5e-3 * 150, (255,255,255),2)
cv2.putText(frame, "Current Object Counter: "+str(i),(int(20), int(80)),0, 5e-3 * 200, (0,255,0),2)
cv2.putText(frame, "FPS: %f"%(fps),(int(20), int(40)),0, 5e-3 * 200, (0,255,0),3)
cv2.putText(frame, "Traffic, Avg_speed", (570,40), 0, 1, (0,255,0), 2)
cv2.putText(frame, "Lane_1: " + str(cnt_lane_1)+', '+str(int(avg_speed_lane_1)), (500,80), 0, 1, (0,255,0), 2)
cv2.putText(frame, "Lane_2: " + str(cnt_lane_2)+', '+str(int(avg_speed_lane_2)), (500,120), 0, 1, (0,255,0), 2)
cv2.putText(frame, "Lane_3: " + str(cnt_lane_3)+', '+str(int(avg_speed_lane_3)), (500,160), 0, 1, (0,255,0), 2)
cv2.putText(frame, "Lane_4: " + str(cnt_lane_4)+', '+str(int(avg_speed_lane_4)), (500,200), 0, 1, (0,255,0), 2)
cv2.putText(frame, "Lane_5: " + str(cnt_lane_5)+', '+str(int(avg_speed_lane_5)), (500,240), 0, 1, (0,255,0), 2)
cv2.putText(frame, "Lane_6: " + str(cnt_lane_6)+', '+str(int(avg_speed_lane_6)), (500,280), 0, 1, (0,255,0), 2)
frame_idx += 1
if writeVideo_flag:
#save a frame
# cv2.fillPoly(frame, [mask], (0,0,0))
# for i,v in enumerate(global_point):
# cv2.circle(frame, v, 1, (0,255,255),-1)
cv2.addWeighted(cmask, alpha, frame, 1 - alpha, 0, frame)
out.write(frame)
frame_index = frame_index + 1
list_file.write(str(frame_index)+' ')
if len(boxs) != 0:
for i in range(0,len(boxs)):
list_file.write(str(boxs[i][0]) + ' '+str(boxs[i][1]) + ' '+str(boxs[i][2]) + ' '+str(boxs[i][3]) + ' ')
list_file.write('\n')
fps = ( fps + (1./(time.time()-t1)) ) / 2
print(" ")
print("[Finish]")
end = time.time()
video_capture.release()
if writeVideo_flag:
out.release()
list_file.close()
cv2.destroyAllWindows()
def main(yolo):
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# Deep SORT
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
np.random.seed(42)
COLORS = np.random.randint(0, 255, size=(200, 3),
dtype="uint8")
writeVideo_flag = True
asyncVideo_flag = False
file_path = 'testvideo2.avi'
# load the COCO class labels our YOLO model was trained on 加载我们的YOLO模型经过培训的COCO类标签
labelsPath = os.path.sep.join(["model_data", "coco_classes.txt"])
LABELS = open(labelsPath).read().strip().split("\n")
# print(str(len(LABELS))+"load successfully")
# print(LABELS)
class_nums = np.zeros(80)
counter = np.zeros(80)
track_id_max = -1
if asyncVideo_flag:
video_capture = VideoCaptureAsync(file_path)
else:
video_capture = cv2.VideoCapture(file_path)
if asyncVideo_flag:
video_capture.start()
if writeVideo_flag:
if asyncVideo_flag:
w = int(video_capture.cap.get(3))
h = int(video_capture.cap.get(4))
else:
w = int(video_capture.get(3))
h = int(video_capture.get(4))
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter('testvideo2_out.avi', fourcc, 30, (w, h))
frame_index = -1
fps = 0.0
fps_imutils = imutils.video.FPS().start()
img_num = 0
frame_cnt = 0
while True:
ret, frame = video_capture.read() # frame shape 640*480*3
frame_copy = frame.copy()
if ret != True or frame_cnt > 30:
break
t1 = time.time()
image = Image.fromarray(frame[..., ::-1]) # bgr to rgb
boxs, confidence, class_names = yolo.detect_image(image)
features = encoder(frame, boxs)
detections = [Detection(bbox, confidence, feature) for bbox, confidence, feature in
zip(boxs, confidence, features)]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap, scores)
# print("print indices!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
# print(indices)
# for i in indices:
# print(str(i)+class_names[i][0])
# print(indices.shape)
detections = [detections[i] for i in indices]
class_names = [class_names[i] for i in indices]
print("class_name:" + str(class_names))
class_IDs = []
current_nums = np.zeros(80)
# class_IDs=[]
for class_name in class_names:
for i, LABEL in enumerate(LABELS):
if class_name[0] == LABEL:
current_nums[i] += 1
class_IDs.append(i)
# print("person:"+str(current_nums[0]))
cv2.putText(frame, 'Current', (20, 70), cv2.FONT_HERSHEY_DUPLEX, 0.75, (255, 255, 255), 2)
cv2.putText(frame, 'Total', (180, 70), cv2.FONT_HERSHEY_DUPLEX, 0.75, (0, 255, 255), 2)
x1 = 20
y1 = 100
for i, cl in enumerate(current_nums):
if cl > 0:
cv2.putText(frame, LABELS[i] + "=" + str(cl), (x1, y1), cv2.FONT_HERSHEY_DUPLEX, 0.6, (255, 255, 255),
1) # 当前帧各类别数量
y1 = y1 + 20
for i, det in enumerate(detections):
bbox = det.to_tlbr()
score = "%.2f" % round(det.confidence * 100, 2)
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), (255, 255, 255), 1)
cv2.putText(frame, score + '%', (int(bbox[0]), int(bbox[1]) + 10), cv2.FONT_HERSHEY_DUPLEX, 0.3,
(255, 255, 255), 1)
# cv2.putText(frame, class_names[i],(int(bbox[0]), int(bbox[1])-5), 0, 5e-3 * 130, (0, 255, 0), 2)
# cv2.putText(frame, class_names[i], (int(bbox[0]), int(bbox[1])), 0, 5e-3 * 130,(0, 255, 255), 2)
print("Total of detections:" + str(len(detections)))
# Call the tracker
tracker.predict()
tracker.update(detections, class_IDs)
# for i, cl in enumerate(class_nums):
# if cl > 0:
# print("add: " + LABELS[i] + str(cl - class_last_nums[i]))
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
print("not track.is_confirmed() or track.time_since_update > 1: " + str(track.track_id))
continue
bbox = track.to_tlbr()
color = [int(c) for c in COLORS[track.class_id % len(COLORS)]]
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), color, 2)
cv2.putText(frame, str(track.track_id) + ' ' + LABELS[track.class_id], (int(bbox[0]), int(bbox[1]) - 5),
cv2.FONT_HERSHEY_DUPLEX, 0.4, color, 1)
if track.track_id > track_id_max:
counter[track.class_id] = counter[track.class_id] + 1
track_id_max = track.track_id
dest = frame_copy[int(bbox[1]):int(bbox[3]), int(bbox[0]):int(bbox[2])]
pdest = "./output_image/" + str(LABELS[track.class_id]) + str(int(counter[track.class_id])) + ".png"
cv2.imwrite(pdest, dest)
img_num += 1
# class_nums[track.class_id].append(track.track_id)
# print(str(LABELS[track.class_id])+":"+class_nums[track.class_id])
# print("track.id: " + str(track.track_id))
# print("track.class_name: " + str(LABELS[track.class_id]))
print(str(counter))
print("--------------------------该帧输出完毕！--------------------------------------")
# cv2.putText(frame, 'Total', (200, 60), cv2.FONT_HERSHEY_DUPLEX, 0.75, (255, 0, 0), 2)
# x2 = 200
# y2 = 100
# for i, cl in enumerate(class_nums):
# if cl > 0:
# cv2.putText(frame, LABELS[i] + "=" + str(cl), (x2, y2), cv2.FONT_HERSHEY_DUPLEX, 0.5, (255, 0, 0), 2)
# y2 = y2 + 20
cv2.putText(frame, "FPS: %f" % (fps), (int(20), int(40)), 0, 5e-3 * 200, (0, 255, 0), 3) # ！！！！！！！！！输出FPS
x2 = 180
y2 = 100
for i, cl in enumerate(counter):
if cl > 0:
cv2.putText(frame, LABELS[i] + "=" + str(cl), (x2, y2), cv2.FONT_HERSHEY_DUPLEX, 0.6, (0, 255, 255), 1)
y2 = y2 + 20
# cv2.imshow('', frame)
if writeVideo_flag: # and not asyncVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
fps_imutils.update()
fps = (fps + (1. / (time.time() - t1))) / 2
# print("FPS = %f"%(fps))
frame_cnt += 1
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
fps_imutils.stop()
print('imutils FPS: {}'.format(fps_imutils.fps()))
if asyncVideo_flag:
video_capture.stop()
else:
video_capture.release()
if writeVideo_flag:
out.release()
cv2.destroyAllWindows()
def main(yolo):
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# Deep SORT
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
writeVideo_flag = True
asyncVideo_flag = False
file_path = 'input_video.avi'
if asyncVideo_flag:
video_capture = VideoCaptureAsync(file_path)
else:
video_capture = cv2.VideoCapture(file_path)
if asyncVideo_flag:
video_capture.start()
if writeVideo_flag:
if asyncVideo_flag:
w = int(video_capture.cap.get(3))
h = int(video_capture.cap.get(4))
else:
w = int(video_capture.get(3))
h = int(video_capture.get(4))
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter('output_yolov3.mp4', fourcc, 25, (w, h))
frame_index = -1
fps = 0.0
fps_imutils = imutils.video.FPS().start()
while True:
ret, frame = video_capture.read() # frame shape 640*480*3
if ret != True:
break
t1 = time.time()
image = Image.fromarray(frame[..., ::-1]) # bgr to rgb
boxes, confidences = yolo.detect_image(image)
boxes = np.array(boxes)
confidences = np.array(confidences)
# Run non-maxima suppression.
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
confidences)
boxes = [boxes[i] for i in indices]
confidences = [confidences[i] for i in indices]
time1 = time.time()
features = encoder(frame, boxes)
time2 = time.time()
detections = [
Detection(bbox, confidence, feature)
for bbox, confidence, feature in zip(boxes, confidences, features)
]
# Call the tracker
tracker.predict()
tracker.update(detections)
time3 = time.time()
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 255, 0), 2)
for det in detections:
bbox = det.to_tlbr()
score = "%.2f" % round(det.confidence * 100, 2)
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 0, 0), 2)
cv2.putText(frame, score + '%', (int(bbox[0]), int(bbox[3])), 0,
5e-3 * 130, (0, 255, 0), 2)
cv2.imshow('', frame)
if writeVideo_flag: # and not asyncVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
fps_imutils.update()
fps = (fps + (1. / (time.time() - t1))) / 2
print("FPS = %f" % (fps))
time4 = time.time()
time_sum = time4 - t1
print("time:", (time1 - t1) / time_sum, (time2 - time1) / time_sum,
(time3 - time2) / time_sum, (time4 - time3) / time_sum)
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
fps_imutils.stop()
print('imutils FPS: {}'.format(fps_imutils.fps()))
if asyncVideo_flag:
video_capture.stop()
else:
video_capture.release()
if writeVideo_flag:
out.release()
cv2.destroyAllWindows()
def main(_argv):
with open("./config_birdview.yml", "r") as ymlfile:
bird_view_cfg = yaml.load(ymlfile)
width_og, height_og = 0, 0
corner_points = []
for section in bird_view_cfg:
corner_points.append(bird_view_cfg["image_parameters"]["p1"])
corner_points.append(bird_view_cfg["image_parameters"]["p2"])
corner_points.append(bird_view_cfg["image_parameters"]["p3"])
corner_points.append(bird_view_cfg["image_parameters"]["p4"])
width_og = int(bird_view_cfg["image_parameters"]["width_og"])
height_og = int(bird_view_cfg["image_parameters"]["height_og"])
img_path = bird_view_cfg["image_parameters"]["img_path"]
size_height = bird_view_cfg["image_parameters"]["size_height"]
size_width = bird_view_cfg["image_parameters"]["size_width"]
tr = np.array([
bird_view_cfg["image_parameters"]["p4"][0],
bird_view_cfg["image_parameters"]["p4"][1],
])
tl = np.array([
bird_view_cfg["image_parameters"]["p2"][0],
bird_view_cfg["image_parameters"]["p2"][1],
])
br = np.array([
bird_view_cfg["image_parameters"]["p3"][0],
bird_view_cfg["image_parameters"]["p3"][1],
])
bl = np.array([
bird_view_cfg["image_parameters"]["p1"][0],
bird_view_cfg["image_parameters"]["p1"][1],
])
widthA = np.sqrt(((br[0] - bl[0])**2) + ((br[1] - bl[1])**2))
widthB = np.sqrt(((tr[0] - tl[0])**2) + ((tr[1] - tl[1])**2))
maxWidth = max(int(widthA), int(widthB))
heightA = np.sqrt(((tr[0] - br[0])**2) + ((tr[1] - br[1])**2))
heightB = np.sqrt(((tl[0] - bl[0])**2) + ((tl[1] - bl[1])**2))
maxHeight = max(int(heightA), int(heightB))
matrix, imgOutput = compute_perspective_transform(corner_points, maxWidth,
maxHeight,
cv2.imread(img_path))
height, width, _ = imgOutput.shape
dim = (width, height)
# Definition of the parameters
max_cosine_distance = 0.4
nn_budget = None
nms_max_overlap = 1.0
# initialize deep sort
model_filename = "model_data/mars-small128.pb"
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
# calculate cosine distance metric
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
# initialize tracker
tracker = Tracker(metric)
# load configuration for object detector
config = ConfigProto()
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)
STRIDES, ANCHORS, NUM_CLASS, XYSCALE = utils.load_config(FLAGS)
input_size = FLAGS.size
video_path = FLAGS.video
# load tflite model if flag is set
if FLAGS.framework == "tflite":
interpreter = tf.lite.Interpreter(model_path=FLAGS.weights)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
print(input_details)
print(output_details)
# otherwise load standard tensorflow saved model
else:
saved_model_loaded = tf.saved_model.load(FLAGS.weights,
tags=[tag_constants.SERVING])
infer = saved_model_loaded.signatures["serving_default"]
# begin video capture
try:
vid = cv2.VideoCapture(int(video_path))
except:
vid = cv2.VideoCapture(video_path)
output_video_1, output_video_2 = None, None
# get video ready to save locally if flag is set
if FLAGS.output:
# by default VideoCapture returns float instead of int
"""
width = int(vid.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))
"""
fps = int(vid.get(cv2.CAP_PROP_FPS))
codec = cv2.VideoWriter_fourcc(*FLAGS.output_format)
out = cv2.VideoWriter(FLAGS.output, codec, fps, (width, height))
frame_num = 0
# while video is running
while True:
black_img = cv2.imread("./black_bg.png")
black_img = cv2.resize(black_img, dim, interpolation=cv2.INTER_AREA)
return_value, frame = vid.read()
if return_value:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
image = Image.fromarray(frame)
else:
print("Video has ended or failed, try a different video format!")
break
frame_num += 1
print("Frame #: ", frame_num)
frame_size = frame.shape[:2]
image_data = cv2.resize(frame, (input_size, input_size))
image_data = image_data / 255.0
image_data = image_data[np.newaxis, ...].astype(np.float32)
start_time = time.time()
# run detections on tflite if flag is set
if FLAGS.framework == "tflite":
interpreter.set_tensor(input_details[0]["index"], image_data)
interpreter.invoke()
pred = [
interpreter.get_tensor(output_details[i]["index"])
for i in range(len(output_details))
]
# run detections using yolov3 if flag is set
if FLAGS.model == "yolov3" and FLAGS.tiny == True:
boxes, pred_conf = filter_boxes(
pred[1],
pred[0],
score_threshold=0.25,
input_shape=tf.constant([input_size, input_size]),
)
else:
boxes, pred_conf = filter_boxes(
pred[0],
pred[1],
score_threshold=0.25,
input_shape=tf.constant([input_size, input_size]),
)
else:
batch_data = tf.constant(image_data)
pred_bbox = infer(batch_data)
for key, value in pred_bbox.items():
boxes = value[:, :, 0:4]
pred_conf = value[:, :, 4:]
(
boxes,
scores,
classes,
valid_detections,
) = tf.image.combined_non_max_suppression(
boxes=tf.reshape(boxes, (tf.shape(boxes)[0], -1, 1, 4)),
scores=tf.reshape(
pred_conf,
(tf.shape(pred_conf)[0], -1, tf.shape(pred_conf)[-1])),
max_output_size_per_class=50,
max_total_size=50,
iou_threshold=FLAGS.iou,
score_threshold=FLAGS.score,
)
# convert data to numpy arrays and slice out unused elements
num_objects = valid_detections.numpy()[0]
bboxes = boxes.numpy()[0]
bboxes = bboxes[0:int(num_objects)]
scores = scores.numpy()[0]
scores = scores[0:int(num_objects)]
classes = classes.numpy()[0]
classes = classes[0:int(num_objects)]
# format bounding boxes from normalized ymin, xmin, ymax, xmax ---> xmin, ymin, width, height
original_h, original_w, _ = frame.shape
bboxes = utils.format_boxes(bboxes, original_h, original_w)
# store all predictions in one parameter for simplicity when calling functions
pred_bbox = [bboxes, scores, classes, num_objects]
# read in all class names from config
class_names = utils.read_class_names(cfg.YOLO.CLASSES)
# by default allow all classes in .names file
# allowed_classes = list(class_names.values())
# custom allowed classes (uncomment line below to customize tracker for only people)
allowed_classes = ["person"]
# loop through objects and use class index to get class name, allow only classes in allowed_classes list
names = []
deleted_indx = []
for i in range(num_objects):
class_indx = int(classes[i])
class_name = class_names[class_indx]
if class_name not in allowed_classes:
deleted_indx.append(i)
else:
names.append(class_name)
names = np.array(names)
count = len(names)
if FLAGS.count:
cv2.putText(
frame,
"Objects being tracked: {}".format(count),
(5, 35),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
2,
(0, 255, 0),
2,
)
print("Objects being tracked: {}".format(count))
# delete detections that are not in allowed_classes
bboxes = np.delete(bboxes, deleted_indx, axis=0)
scores = np.delete(scores, deleted_indx, axis=0)
# encode yolo detections and feed to tracker
features = encoder(frame, bboxes)
detections = [
Detection(bbox, score, class_name, feature)
for bbox, score, class_name, feature in zip(
bboxes, scores, names, features)
]
# initialize color map
cmap = plt.get_cmap("tab20b")
colors = [cmap(i)[:3] for i in np.linspace(0, 1, 20)]
# run non-maxima supression
boxs = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
classes = np.array([d.class_name for d in detections])
indices = preprocessing.non_max_suppression(boxs, classes,
nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
bbox_array = []
# update tracks
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
bbox_array.append(
(int(bbox[0]), int(bbox[1]), int(bbox[2]), int(bbox[3])))
class_name = track.get_class()
# draw bbox on screen
color = colors[int(track.track_id) % len(colors)]
color = [i * 255 for i in color]
cv2.rectangle(
frame,
(int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])),
color,
2,
)
cv2.rectangle(
frame,
(int(bbox[0]), int(bbox[1] - 30)),
(
int(bbox[0]) +
(len(class_name) + len(str(track.track_id))) * 17,
int(bbox[1]),
),
color,
-1,
)
cv2.putText(
frame,
class_name + "-" + str(track.track_id),
(int(bbox[0]), int(bbox[1] - 10)),
0,
0.75,
(255, 255, 255),
2,
)
# if enable info flag then print details about each track
if FLAGS.info:
print(
"Tracker ID: {}, Class: {}, BBox Coords (xmin, ymin, xmax, ymax): {}"
.format(
str(track.track_id),
class_name,
(int(bbox[0]), int(bbox[1]), int(bbox[2]), int(
bbox[3])),
))
if len(bbox_array) >= 1:
array_centroids, array_groundpoints = get_centroids_and_groundpoints(
bbox_array)
transformed_downoids = compute_point_perspective_transformation(
matrix, array_centroids)
# Show every point on the top view image
for point in transformed_downoids:
x, y = point
cv2.circle(black_img, (x, y), 60, (0, 255, 0), 2)
cv2.circle(black_img, (x, y), 3, (0, 255, 0), -1)
# calculate frames per second of running detections
fps = 1.0 / (time.time() - start_time)
print("FPS: %.2f" % fps)
result = np.asarray(frame)
# result = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
if not FLAGS.dont_show:
cv2.imshow("Output Video", result)
# if output flag is set, save video file
if FLAGS.output:
if output_video_1 is None and output_video_2 is None:
fourcc1 = cv2.VideoWriter_fourcc(*"MJPG")
output_video_1 = cv2.VideoWriter(
"./video.avi", fourcc1, 25,
(frame.shape[1], frame.shape[0]), True)
fourcc2 = cv2.VideoWriter_fourcc(*"MJPG")
output_video_2 = cv2.VideoWriter(
"./bird_view.avi",
fourcc2,
25,
(black_img.shape[1], black_img.shape[0]),
True,
)
elif output_video_1 is not None and output_video_2 is not None:
output_video_1.write(frame)
output_video_2.write(black_img)
if cv2.waitKey(1) & 0xFF == ord("q"):
break
cv2.destroyAllWindows()
def main(_argv):
# Definition of the parameters
max_cosine_distance = 0.5
nn_budget = None
nms_max_overlap = 1
#initialize deep sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric, max_age=40)
physical_devices = tf.config.experimental.list_physical_devices('GPU')
if len(physical_devices) > 0:
tf.config.experimental.set_memory_growth(physical_devices[0], True)
if FLAGS.tiny:
yolo = YoloV3Tiny(classes=FLAGS.num_classes)
else:
yolo = YoloV3(classes=FLAGS.num_classes)
yolo.load_weights(FLAGS.weights)
logging.info('weights loaded')
class_names = [c.strip() for c in open(FLAGS.classes).readlines()]
logging.info('classes loaded')
try:
vid = cv2.VideoCapture(int(FLAGS.video))
except:
vid = cv2.VideoCapture(FLAGS.video)
out = None
if FLAGS.output:
# by default VideoCapture returns float instead of int
width = int(vid.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(vid.get(cv2.CAP_PROP_FPS))
codec = cv2.VideoWriter_fourcc(*FLAGS.output_format)
out = cv2.VideoWriter(FLAGS.output, codec, fps, (width, height))
list_file = open('detection.txt', 'w')
frame_index = -1
fps = 0.0
count = 0
while True:
_, img = vid.read()
if img is None:
logging.warning("Empty Frame")
time.sleep(0.1)
count += 1
if count < 3:
continue
else:
break
img_in = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img_in = tf.expand_dims(img_in, 0)
img_in = transform_images(img_in, FLAGS.size)
t1 = time.time()
boxes, scores, classes, nums = yolo.predict(img_in)
classes = classes[0]
names = []
for i in range(len(classes)):
names.append(class_names[int(classes[i])])
names = np.array(names)
converted_boxes = convert_boxes(img, boxes[0])
features = encoder(img, converted_boxes)
detections = [
Detection(bbox, score, class_name, feature)
for bbox, score, class_name, feature in zip(
converted_boxes, scores[0], names, features)
]
#initialize color map
cmap = plt.get_cmap('tab20b')
colors = [cmap(i)[:3] for i in np.linspace(0, 1, 20)]
# run non-maxima suppresion
boxs = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
classes = np.array([d.class_name for d in detections])
indices = preprocessing.non_max_suppression(boxs, classes,
nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
objects = 0
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
class_name = track.get_class()
if (FLAGS.class_1 == 'all'):
objects += 1
color = colors[int(track.track_id) % len(colors)]
color = [i * 255 for i in color]
cv2.rectangle(img, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), color, 2)
cv2.rectangle(
img, (int(bbox[0]), int(bbox[1] - 30)),
(int(bbox[0]) +
(len(class_name) + len(str(track.track_id))) * 17,
int(bbox[1])), color, -1)
cv2.putText(img, class_name + "-" + str(track.track_id),
(int(bbox[0]), int(bbox[1] - 10)), 0, 0.75,
(255, 255, 255), 2)
elif (FLAGS.class_1 != 'all'):
if (class_name == FLAGS.class_1):
objects += 1
color = colors[int(track.track_id) % len(colors)]
color = [i * 255 for i in color]
cv2.rectangle(img, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), color, 2)
cv2.rectangle(
img, (int(bbox[0]), int(bbox[1] - 30)),
(int(bbox[0]) +
(len(class_name) + len(str(track.track_id))) * 17,
int(bbox[1])), color, -1)
cv2.putText(img, class_name + "-" + str(track.track_id),
(int(bbox[0]), int(bbox[1] - 10)), 0, 0.75,
(255, 255, 255), 2)
# print("Objetos filtrados:{}".format(objects))
# print N_objects on screen
cv2.putText(img, "# Objetos: {}".format(objects), (0, 30),
cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (255, 0, 0), 2)
### UNCOMMENT BELOW IF YOU WANT CONSTANTLY CHANGING YOLO DETECTIONS TO BE SHOWN ON SCREEN
for det in detections:
bbox = det.to_tlbr()
cv2.rectangle(img, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 0, 0), 2)
# print fps on screen
#fps = ( fps + (1./(time.time()-t1)) ) / 2
#cv2.putText(img, "FPS: {:.2f}".format(fps), (0, 30),
# cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2)
cv2.imshow('output', img)
if FLAGS.output:
out.write(img)
frame_index = frame_index + 1
list_file.write(str(frame_index) + ' ')
if len(converted_boxes) != 0:
for i in range(0, len(converted_boxes)):
list_file.write(
str(converted_boxes[i][0]) + ' ' +
str(converted_boxes[i][1]) + ' ' +
str(converted_boxes[i][2]) + ' ' +
str(converted_boxes[i][3]) + ' ')
list_file.write('\n')
# press q to quit
if cv2.waitKey(1) == ord('q'):
break
vid.release()
if FLAGS.ouput:
out.release()
list_file.close()
cv2.destroyAllWindows()
def main(yolo):
# Definition of the parameters
max_cosine_distance = 0.2
nn_budget = None
nms_max_overlap = 1.0
# deep_sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
video_capture = cv2.VideoCapture(0)
fps = 0.0
while True:
_, frame = video_capture.read() # frame shape 640*480*3
t1 = time.time()
image = Image.fromarray(frame)
boxs = yolo.detect_image(image)
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
for track, det in zip(tracker.tracks, detections):
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 4)
bbox = det.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 0, 0), 4)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 480,
(124, 252, 0), 4)
cv2.imshow('', frame)
fps = (fps + (1. / (time.time() - t1))) / 2
print("fps= %f" % (fps))
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
video_capture.release()
cv2.destroyAllWindows()
def main(yolo, input):
#拡張子ありのファイル名
basename = os.path.basename(input)
print(" START YOLOv4 + DeepSort input file is ", basename)
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# Deep SORT
model_filename = '../model/mars-small128.pb'
cencoder = gdet.create_box_encoder(model_filename, batch_size=1)
pencoder = gdet.create_box_encoder(model_filename, batch_size=1)
cmetric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
pmetric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
ctracker = Tracker(cmetric)
ptracker = Tracker(pmetric)
tracking = True
writeVideo_flag = True
#推論したいカテゴリを設定
cl_list = ['Pedestrian', 'Car']
video_capture = cv2.VideoCapture(input)
fps = 0.0
fps_imutils = imutils.video.FPS().start()
if writeVideo_flag:
basename_without_ext = os.path.splitext(os.path.basename(input))[0]
fname = basename_without_ext + 'output_yolov4.mp4'
output_path = '../output/' + fname
video_FourCC = int(video_capture.get(cv2.CAP_PROP_FOURCC))
video_fps = video_capture.get(cv2.CAP_PROP_FPS)
video_size = (int(video_capture.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(video_capture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
out = cv2.VideoWriter(output_path, video_FourCC, video_fps, video_size)
frame_index = -1
Nm_fr = 0
all_result = []
while True:
Car_result_ALL = []
Pedestrian_result_ALL = []
Nm_fr = Nm_fr + 1
ret, frame = video_capture.read() # frame shape 1920*1216*3
if ret != True:
break
print("Frame no. = ", Nm_fr)
t1 = time.time()
image = Image.fromarray(frame[..., ::-1])
cboxes, cconfidence, cclasses = yolo.detect_image(image, cl_list[1])
pboxes, pconfidence, pclasses = yolo.detect_image(image, cl_list[0])
if tracking:
cfeatures = cencoder(frame, cboxes)
pfeatures = pencoder(frame, pboxes)
cdetections = [Detection(cbbox, cconfidence, ceach_class, cfeature) for cbbox, cconfidence, ceach_class, cfeature in \
zip(cboxes, cconfidence, cclasses, cfeatures)]
pdetections = [Detection(pbbox, pconfidence, peach_class, pfeature) for pbbox, pconfidence, peach_class, pfeature in \
zip(pboxes, pconfidence, pclasses, pfeatures)]
#else:
# detections = [Detection_YOLO(bbox, confidence, each_class) for bbox, confidence, each_class in \
# zip(boxes, confidence, classes)]
# Run non-maxima suppression.
cboxes = np.array([d.tlwh for d in cdetections])
cscores = np.array([d.confidence for d in cdetections])
cindices = preprocessing.non_max_suppression(
cboxes, nms_max_overlap, cscores)
cdetections = [cdetections[i] for i in cindices]
pboxes = np.array([d.tlwh for d in pdetections])
pscores = np.array([d.confidence for d in pdetections])
pindices = preprocessing.non_max_suppression(
pboxes, nms_max_overlap, pscores)
pdetections = [pdetections[i] for i in pindices]
if tracking:
# Call the tracker
ctracker.predict()
ctracker.update(cdetections)
ptracker.predict()
ptracker.update(pdetections)
for ctrack in ctracker.tracks:
if not ctrack.is_confirmed(
) or ctrack.time_since_update > 1:
continue
cbbox = ctrack.to_tlbr()
cv2.rectangle(frame, (int(cbbox[0]), int(cbbox[1])),
(int(cbbox[2]), int(cbbox[3])), (0, 0, 255),
3)
cv2.putText(frame, "ID: " + str(ctrack.track_id), (int(cbbox[0]), int(cbbox[1])), 0, \
1.5e-3 * frame.shape[0], (0, 0, 255), 3)
#OUTPUT TRACKING
ID = int(ctrack.track_id)
left = int(cbbox[0])
top = int(cbbox[1])
right = int(cbbox[2])
bottom = int(cbbox[3])
Car_result = {
'id': ID,
'box2d': [left, top, right, bottom]
} #予測結果
print("Car_result = ", Car_result)
Car_result_ALL.append(Car_result)
for ptrack in ptracker.tracks:
if not ptrack.is_confirmed(
) or ptrack.time_since_update > 1:
continue
pbbox = ptrack.to_tlbr()
cv2.rectangle(frame, (int(pbbox[0]), int(pbbox[1])),
(int(pbbox[2]), int(pbbox[3])), (255, 0, 0),
3)
cv2.putText(frame, "ID: " + str(ptrack.track_id), (int(pbbox[0]), int(pbbox[1])), 0, \
1.5e-3 * frame.shape[0], (255, 0, 0), 3)
#OUTPUT TRACKING
ID = int(ptrack.track_id)
left = int(pbbox[0])
top = int(pbbox[1])
right = int(pbbox[2])
bottom = int(pbbox[3])
Pedestrian_result = {
'id': ID,
'box2d': [left, top, right, bottom]
} #予測結果
print("Pedestrian_result = ", Pedestrian_result)
Pedestrian_result_ALL.append(Pedestrian_result)
#YOLOv4 output to frame for Car
for cdet in cdetections:
cbbox = cdet.to_tlbr()
cscore = "%.2f" % round(cdet.confidence * 100, 2) + "%"
cv2.rectangle(frame, (int(cbbox[0]), int(cbbox[1])),
(int(cbbox[2]), int(cbbox[3])), (255, 255, 255),
2)
if len(cclasses) > 0:
ceach_class = cdet.cls
cv2.putText(frame, str(ceach_class) + " " + cscore, (int(cbbox[0]), int(cbbox[3])), 0, \
1.5e-3 * frame.shape[0], (255, 255, 255), 2)
#YOLOv4 output to frame for Pedestrian
for pdet in pdetections:
pbbox = pdet.to_tlbr()
pscore = "%.2f" % round(pdet.confidence * 100, 2) + "%"
cv2.rectangle(frame, (int(pbbox[0]), int(pbbox[1])),
(int(pbbox[2]), int(pbbox[3])), (127, 127, 127),
2)
if len(pclasses) > 0:
peach_class = pdet.cls
cv2.putText(frame, str(peach_class) + " " + pscore, (int(pbbox[0]), int(pbbox[3])), 0, \
1.5e-3 * frame.shape[0], (127, 127, 127), 2)
# Each frame result
all_result.append({
'Car': Car_result_ALL,
'Pedestrian': Pedestrian_result_ALL
})
if writeVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
fps_imutils.update()
fps = (fps + (1. / (time.time() - t1))) / 2
print(" FPS = %f" % (fps))
if writeVideo_flag:
out.release()
video_capture.release()
fps_imutils.stop()
print('imutils FPS: {}'.format(fps_imutils.fps()))
return {basename: all_result}
def Object_tracking(Yolo,
video_path,
output_path,
input_size=416,
show=False,
CLASSES=YOLO_COCO_CLASSES,
score_threshold=0.3,
iou_threshold=0.45,
rectangle_colors='',
Track_only=[]):
# Definition of the parameters
max_cosine_distance = 0.7
nn_budget = None
#initialize deep sort object
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
times, times_2 = [], []
if video_path:
vid = cv2.VideoCapture(video_path) # detect on video
else:
vid = cv2.VideoCapture(0) # detect from webcam
# by default VideoCapture returns float instead of int
width = int(vid.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(vid.get(cv2.CAP_PROP_FPS))
codec = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(output_path, codec, fps,
(width, height)) # output_path must be .mp4
NUM_CLASS = read_class_names(CLASSES)
key_list = list(NUM_CLASS.keys())
val_list = list(NUM_CLASS.values())
while True:
_, frame = vid.read()
try:
original_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
original_frame = cv2.cvtColor(original_frame, cv2.COLOR_BGR2RGB)
except:
break
image_data = image_preprocess(np.copy(original_frame),
[input_size, input_size])
#image_data = tf.expand_dims(image_data, 0)
image_data = image_data[np.newaxis, ...].astype(np.float32)
t1 = time.time()
if YOLO_FRAMEWORK == "tf":
pred_bbox = Yolo.predict(image_data)
elif YOLO_FRAMEWORK == "trt":
batched_input = tf.constant(image_data)
result = Yolo(batched_input)
pred_bbox = []
for key, value in result.items():
value = value.numpy()
pred_bbox.append(value)
#t1 = time.time()
#pred_bbox = Yolo.predict(image_data)
t2 = time.time()
pred_bbox = [tf.reshape(x, (-1, tf.shape(x)[-1])) for x in pred_bbox]
pred_bbox = tf.concat(pred_bbox, axis=0)
bboxes = postprocess_boxes(pred_bbox, original_frame, input_size,
score_threshold)
bboxes = nms(bboxes, iou_threshold, method='nms')
# extract bboxes to boxes (x, y, width, height), scores and names
boxes, scores, names = [], [], []
for bbox in bboxes:
if len(Track_only) != 0 and NUM_CLASS[int(
bbox[5])] in Track_only or len(Track_only) == 0:
boxes.append([
bbox[0].astype(int), bbox[1].astype(int),
bbox[2].astype(int) - bbox[0].astype(int),
bbox[3].astype(int) - bbox[1].astype(int)
])
scores.append(bbox[4])
names.append(NUM_CLASS[int(bbox[5])])
# Obtain all the detections for the given frame.
boxes = np.array(boxes)
names = np.array(names)
scores = np.array(scores)
features = np.array(encoder(original_frame, boxes))
detections = [
Detection(bbox, score, class_name, feature)
for bbox, score, class_name, feature in zip(
boxes, scores, names, features)
]
# Pass detections to the deepsort object and obtain the track information.
tracker.predict()
tracker.update(detections)
# Obtain info from the tracks
tracked_bboxes = []
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 5:
continue
bbox = track.to_tlbr() # Get the corrected/predicted bounding box
class_name = track.get_class(
) #Get the class name of particular object
tracking_id = track.track_id # Get the ID for the particular track
index = key_list[val_list.index(
class_name)] # Get predicted object index by object name
tracked_bboxes.append(
bbox.tolist() + [tracking_id, index]
) # Structure data, that we could use it with our draw_bbox function
# draw detection on frame
image = draw_bbox(original_frame,
tracked_bboxes,
CLASSES=CLASSES,
tracking=True)
t3 = time.time()
times.append(t2 - t1)
times_2.append(t3 - t1)
times = times[-20:]
times_2 = times_2[-20:]
ms = sum(times) / len(times) * 1000
fps = 1000 / ms
fps2 = 1000 / (sum(times_2) / len(times_2) * 1000)
image = cv2.putText(image, "Time: {:.1f} FPS".format(fps), (0, 30),
cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2)
# draw original yolo detection
#image = draw_bbox(image, bboxes, CLASSES=CLASSES, show_label=False, rectangle_colors=rectangle_colors, tracking=True)
print(
"Time: {:.2f}ms, Detection FPS: {:.1f}, total FPS: {:.1f}".format(
ms, fps, fps2))
if output_path != '': out.write(image)
if show:
cv2.imshow('output', image)
if cv2.waitKey(25) & 0xFF == ord("q"):
cv2.destroyAllWindows()
break
cv2.destroyAllWindows()
def main(_argv):
# Definition of the parameters
max_cosine_distance = 0.4
nn_budget = None
nms_max_overlap = 1.0
# initialize deep sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
# calculate cosine distance metric
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
# initialize tracker
tracker = Tracker(metric)
# load configuration for object detector
config = ConfigProto()
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)
STRIDES, ANCHORS, NUM_CLASS, XYSCALE = utils.load_config(FLAGS)
input_size = FLAGS.size
video_path = FLAGS.video
# load tflite model if flag is set
if FLAGS.framework == 'tflite':
interpreter = tf.lite.Interpreter(model_path=FLAGS.weights)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
print(input_details)
print(output_details)
# otherwise load standard tensorflow saved model
else:
saved_model_loaded = tf.saved_model.load(FLAGS.weights,
tags=[tag_constants.SERVING])
infer = saved_model_loaded.signatures['serving_default']
# begin video capture
try:
vid = cv2.VideoCapture(int(video_path))
except:
vid = cv2.VideoCapture(video_path)
out = None
# get video ready to save locally if flag is set
if FLAGS.output:
# by default VideoCapture returns float instead of int
width = int(vid.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(vid.get(cv2.CAP_PROP_FPS))
codec = cv2.VideoWriter_fourcc(*FLAGS.output_format)
out = cv2.VideoWriter(FLAGS.output, codec, fps, (width, height))
frame_num = 0
# while video is running
while True:
return_value, frame = vid.read()
if return_value:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
image = Image.fromarray(frame)
else:
print('Video has ended or failed, try a different video format!')
break
frame_num += 1
print('Frame #: ', frame_num)
frame_size = frame.shape[:2]
image_data = cv2.resize(frame, (input_size, input_size))
image_data = image_data / 255.
image_data = image_data[np.newaxis, ...].astype(np.float32)
start_time = time.time()
# run detections on tflite if flag is set
if FLAGS.framework == 'tflite':
interpreter.set_tensor(input_details[0]['index'], image_data)
interpreter.invoke()
pred = [
interpreter.get_tensor(output_details[i]['index'])
for i in range(len(output_details))
]
# run detections using yolov3 if flag is set
if FLAGS.model == 'yolov3' and FLAGS.tiny == True:
boxes, pred_conf = filter_boxes(pred[1],
pred[0],
score_threshold=0.25,
input_shape=tf.constant(
[input_size, input_size]))
else:
boxes, pred_conf = filter_boxes(pred[0],
pred[1],
score_threshold=0.25,
input_shape=tf.constant(
[input_size, input_size]))
else:
batch_data = tf.constant(image_data)
pred_bbox = infer(batch_data)
for key, value in pred_bbox.items():
boxes = value[:, :, 0:4]
pred_conf = value[:, :, 4:]
boxes, scores, classes, valid_detections = tf.image.combined_non_max_suppression(
boxes=tf.reshape(boxes, (tf.shape(boxes)[0], -1, 1, 4)),
scores=tf.reshape(
pred_conf,
(tf.shape(pred_conf)[0], -1, tf.shape(pred_conf)[-1])),
max_output_size_per_class=50,
max_total_size=50,
iou_threshold=FLAGS.iou,
score_threshold=FLAGS.score)
# convert data to numpy arrays and slice out unused elements
num_objects = valid_detections.numpy()[0]
bboxes = boxes.numpy()[0]
bboxes = bboxes[0:int(num_objects)]
scores = scores.numpy()[0]
scores = scores[0:int(num_objects)]
classes = classes.numpy()[0]
classes = classes[0:int(num_objects)]
# format bounding boxes from normalized ymin, xmin, ymax, xmax ---> xmin, ymin, width, height
original_h, original_w, _ = frame.shape
bboxes = utils.format_boxes(bboxes, original_h, original_w)
# store all predictions in one parameter for simplicity when calling functions
pred_bbox = [bboxes, scores, classes, num_objects]
# read in all class names from config
class_names = utils.read_class_names(cfg.YOLO.CLASSES)
# by default allow all classes in .names file
allowed_classes = list(class_names.values())
# custom allowed classes (uncomment line below to customize tracker for only people)
#allowed_classes = ['person']
# loop through objects and use class index to get class name, allow only classes in allowed_classes list
names = []
deleted_indx = []
for i in range(num_objects):
class_indx = int(classes[i])
class_name = class_names[class_indx]
if class_name not in allowed_classes:
deleted_indx.append(i)
else:
names.append(class_name)
names = np.array(names)
count = len(names)
if FLAGS.count:
cv2.putText(frame, "Objects being tracked: {}".format(count),
(5, 35), cv2.FONT_HERSHEY_COMPLEX_SMALL, 2,
(0, 255, 0), 2)
print("Objects being tracked: {}".format(count))
# delete detections that are not in allowed_classes
bboxes = np.delete(bboxes, deleted_indx, axis=0)
scores = np.delete(scores, deleted_indx, axis=0)
# encode yolo detections and feed to tracker
features = encoder(frame, bboxes)
detections = [
Detection(bbox, score, class_name, feature)
for bbox, score, class_name, feature in zip(
bboxes, scores, names, features)
]
# initialize color map
cmap = plt.get_cmap('tab20b')
colors = [cmap(i)[:3] for i in np.linspace(0, 1, 20)]
# run non-maxima supression
boxs = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
classes = np.array([d.class_name for d in detections])
indices = preprocessing.non_max_suppression(boxs, classes,
nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
# update tracks
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
class_name = track.get_class()
# draw bbox on screen
# names = {'6_d': 'Thomas Delaney',
# '10_b': 'Leroy Sane',
# '18_b': 'Leon Goretzka',
# '25_b': 'Thomas Muller',
# '5_d': 'Dan-Axel Zagadou',
# '12_d': 'Zaragoza',
# '4_b': 'Niklas Sule',
# '14_d': 'Nico Schulz',
# '11_d': 'Marco Reus',
# 'Referee': 'Referee',
# 'ball': 'ball',
# '10_d': 'Thorgan Hazard',
# '6_b': 'Joshua Kimmich ',
# 'gk_b': 'Ron-Thorben Hoffmann(GK)',
# '17_b': 'Jérôme Boateng',
# '27_b': 'David Alaba',
# '9_d': 'Erling Haaland',
# '8_d': 'Mahmoud Dahoud',
# 'gk_d': 'Luca Unbehaun(GK)',
# '19_b': 'Alphonso Davies',
# '29_b': 'Kingsley Coman',
# '24_d': 'Marcel Schmelzer',
# '9_b': 'Robert Lewandowski',
# "23_d": 'Emre Can',
# }
# if class_name == 'Referee':
# color = (0, 0, 0)
if class_name == 'ball':
# color = (255, 255, 255)
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 1)
# else:
# try:
# colors = {'b': (252, 3, 78), 'd': (250, 247, 80)}
# color = colors[str(class_name.split('_')[-1])]
# except KeyError:
# pass
# class_name = names[str(class_name)]
# color = (250, 247, 80)
# color = colors[int(track.track_id) % len(colors)]
# color = [i * 255 for i in color]
# cv2.rectangle(frame, (int(bbox[0]), int(
# bbox[1])), (int(bbox[2]), int(bbox[3])), color, 1)
# cv2.rectangle(frame, (int(bbox[0]), int(
# bbox[1]-30)), (int(bbox[0])+(len(str(class_name)))*17, int(bbox[1])), color, -1)
cv2.putText(frame, class_name, (int(bbox[0]), int(bbox[1] - 10)),
0, 0.75, (255, 251, 46), 2)
# if enable info flag then print details about each track
if FLAGS.info:
print(
"Tracker ID: {}, Class: {}, BBox Coords (xmin, ymin, xmax, ymax): {}"
.format(str(track.track_id), class_name, (int(
bbox[0]), int(bbox[1]), int(bbox[2]), int(bbox[3]))))
# calculate frames per second of running detections
fps = 1.0 / (time.time() - start_time)
print("FPS: %.2f" % fps)
result = np.asarray(frame)
result = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
if not FLAGS.dont_show:
cv2.imshow("Output Video", result)
# if output flag is set, save video file
if FLAGS.output:
out.write(result)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cv2.destroyAllWindows()
def run_cam_mode(detection_model, args):
'''
Run the tracker on camera stream
'''
cam = cv2.VideoCapture(0)
cam.set(cv2.CAP_PROP_FRAME_WIDTH, FRAME_WIDTH)
cam.set(cv2.CAP_PROP_FRAME_HEIGHT, FRAME_HEIGHT)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", args.max_cosine_distance, NN_BUDGET)
tracker = Tracker(metric)
encoder = gen_det.create_box_encoder(HUMAN_ENCODER_PATH, batch_size=32)
# Cam loop
while True:
ret, frame = cam.read()
t1 = time.time()
# Detect humans in the frame
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
bboxes, detection_scores = detect_humans(detection_model, frame_rgb)
detection_list = cvt_to_detection_objects(frame_rgb, bboxes, detection_scores, encoder)
# Run non-maxima suppression.
detection_list = [d for d in detection_list if d.confidence >= args.min_confidence]
boxes = np.array([d.tlwh for d in detection_list])
scores = np.array([d.confidence for d in detection_list])
indices = dsutil_prep.non_max_suppression(boxes, NMS_MAX_OVERLAP, scores)
detection_list = [detection_list[i] for i in indices]
# Update tracker
tracker.predict()
tracker.update(detection_list)
# FPS counter
fps = 1/(time.time()-t1)
# Update visualization
output_img = frame.copy()
## Visualize all detections
for i, detection in enumerate(detection_list):
x,y,w,h = detection.tlwh
pt1 = int(x), int(y)
pt2 = int(x + w), int(y + h)
cv2.rectangle(output_img, pt1, pt2, (0, 0, 255), 2)
## Visualize confirmed tracks
tracks = tracker.tracks
for track in tracks:
if not track.is_confirmed() or track.time_since_update > 0:
continue
color = dsutil_viz.create_unique_color_uchar(track.track_id)
x,y,w,h = track.to_tlwh()
pt1 = int(x), int(y)
pt2 = int(x + w), int(y + h)
cv2.rectangle(output_img, pt1, pt2, color, 2)
text_size = cv2.getTextSize(str(track.track_id), cv2.FONT_HERSHEY_PLAIN, 1, 2)
center = pt1[0] + 5, pt1[1] + 5 + text_size[0][1]
pt2 = pt1[0] + 10 + text_size[0][0], pt1[1] + 10 + text_size[0][1]
cv2.rectangle(output_img, pt1, pt2, color, -1)
cv2.putText(output_img, str(track.track_id), center, cv2.FONT_HERSHEY_PLAIN, 1, (255, 255, 255), 2)
cv2.putText(output_img, 'FPS: {:.1f}'.format(fps), (50,50), cv2.FONT_HERSHEY_SIMPLEX, 1, (0,255,0), 2, cv2.LINE_AA)
cv2.imshow('detection output', output_img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cv2. destroyAllWindows()
cam.release()
def Display():
print("Start Displaying")
yolo = YOLO()
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# deep_sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
writeVideo_flag = True
w = 768
h = 432
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
out = cv2.VideoWriter('output.avi', fourcc, 15, (w, h))
list_file = open('detection.txt', 'w')
frame_index = -1
fps = 0.0
max_boxs = 0
person_track = {}
yolo2 = YOLO2()
while True:
if q.empty() != True:
#读取打卡信息
face = []
cur1 = conn.cursor() # 获取一个游标
sql1 = "select * from worker"
cur1.execute(sql1)
data = cur1.fetchall()
for d in data:
# 注意int类型需要使用str函数转义
name = str(d[1]) + '_' + d[2]
face.append(name)
cur1.close() # 关闭游标
#获取队列帧
frame = q.get()
t1 = time.time()
#进行安全措施检测
image = Image.fromarray(frame[..., ::-1]) # bgr to rgb
img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
frame, wear = yolo2.detect_image(img)
frame = np.array(frame)
frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
# 获取警戒线
#cv2.line(frame, (132,368), (229, 368), (0, 255, 255), 3)
cv2.line(frame, (275,360), (378, 360), (0, 255, 255), 1)
transboundaryline = t.line_detect_possible_demo(frame)
#yolo目标检测
boxs = yolo.detect_image(image)
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [Detection(bbox, 1.0, feature) for bbox, feature in zip(boxs, features)]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
if len(boxs) > max_boxs:
max_boxs = len(boxs)
# Call the tracker
tracker.predict()
tracker.update(detections)
for track in tracker.tracks:
if max_boxs < track.track_id:
tracker.tracks.remove(track)
tracker._next_id = max_boxs + 1
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
PointX = bbox[0] + ((bbox[2] - bbox[0]) / 2)
PointY = bbox[3]
if track.track_id not in person_track:
track2 = copy.deepcopy(track)
person_track[track.track_id] = track2
else:
track2 = copy.deepcopy(track)
bbox2 = person_track[track.track_id].to_tlbr()
PointX2 = bbox2[0] + ((bbox2[2] - bbox2[0]) / 2)
PointY2 = bbox2[3]
distance = math.sqrt(pow(PointX - PointX2, 2) + pow(PointY - PointY2, 2))
#修正
if distance < 120:
person_track[track.track_id] = track2
else:
# print('last',track.track_id)
dis = {}
for key in person_track:
bbox3 = person_track[key].to_tlbr()
PointX3 = bbox3[0] + ((bbox3[2] - bbox3[0]) / 2)
PointY3 = bbox3[3]
d = math.sqrt(pow(PointX3 - PointX, 2) + pow(PointY3 - PointY, 2))
dis[key] = d
dis = sorted(dis.items(), key=operator.itemgetter(1), reverse=False)
track2.track_id = dis[0][0]
person_track[dis[0][0]] = track2
tracker.tracks.remove(track)
tracker.tracks.append(person_track[track.track_id])
# 写入class
try:
box_title = face[track2.track_id - 1]
except Exception as e:
box_title = str(track2.track_id) + "_" + "unknow"
if box_title not in workers:
wid = box_title.split('_')[0]
localtime = time.asctime(time.localtime(time.time()))
workers[box_title] = wk.Worker()
workers[box_title].set(box_title, localtime, (int(PointX), int(PointY)))
cur2 = conn.cursor() # 获取一个游标
sql2 = "UPDATE worker SET in_time='" + localtime + "' WHERE worker_id= '" + wid + "'"
cur2.execute(sql2)
cur2.close() # 关闭游标
else:
localtime = time.asctime(time.localtime(time.time()))
yoloPoint = (int(PointX), int(PointY))
wear_dic = {}
workers[box_title].current_point = yoloPoint
workers[box_title].track_point.append(workers[box_title].current_point)
mytrack = str(workers[box_title].track_point)
wid = box_title.split('_')[0]
# 卡尔曼滤波预测
if wid not in utils.KalmanNmae:
utils.myKalman(wid)
if wid not in utils.lmp:
utils.setLMP(wid)
cpx, cpy = utils.predict(workers[box_title].current_point[0], workers[box_title].current_point[1], wid)
if cpx[0] == 0.0 or cpy[0] == 0.0:
cpx[0] = workers[box_title].current_point[0]
cpy[0] = workers[box_title].current_point[1]
workers[box_title].next_point = (int(cpx), int(cpy))
cur3 = conn.cursor() # 获取一个游标
sql3 = "UPDATE worker SET current_point= '" + str(workers[box_title].current_point) + "' ,track_point = '" + mytrack + "',next_point = '" + str(workers[box_title].next_point) + "' WHERE worker_id= '" + wid + "'"
cur3.execute(sql3)
cur3.close()
# 写入安全措施情况
if len(wear) > 0:
for w in wear:
wear_dis = int(math.sqrt(pow(w[0] - yoloPoint[0], 2) + pow(w[1] - yoloPoint[1], 2)))
wear_dic[wear_dis] = w
wear_dic = sorted(wear_dic.items(), key=operator.itemgetter(0), reverse=False)
if wear_dic[0][0] < 120:
cur4 = conn.cursor() # 获取一个游标
if wear[wear_dic[0][1]] == 1:
if len(workers[box_title].wear['no helmet']) == 0:
workers[box_title].wear['no helmet'].append(localtime)
sql = "INSERT INTO wear SET worker_id = '" + wid + "', type = 'no_helmet',abnormal_time = '" + localtime + "'"
cur4.execute(sql)
cur4.close() # 关闭游标
else:
print(box_title,workers[box_title].wear['no helmet'])
if localtime not in workers[box_title].wear['no helmet']:
workers[box_title].wear['no helmet'].append(localtime)
sql = "INSERT INTO wear SET worker_id = '" + wid + "', type = 'no_helmet',abnormal_time = '" + localtime + "'"
cur4.execute(sql)
cur4.close() # 关闭游标
elif wear[wear_dic[0][1]] == 2:
if len(workers[box_title].wear['no work cloths']) == 0:
workers[box_title].wear['no work cloths'].append(localtime)
sql = "INSERT INTO wear SET worker_id = '" + wid + "', type = 'no work cloths',abnormal_time = '" + localtime + "'"
cur4.execute(sql)
cur4.close() # 关闭游标
else:
if localtime not in workers[box_title].wear['no work cloths']:
workers[box_title].wear['no work cloths'].append(localtime)
sql = "INSERT INTO wear SET worker_id = '" + wid + "', type = 'no work cloths',abnormal_time = '" + localtime + "'"
cur4.execute(sql)
cur4.close() # 关闭游标
elif wear[wear_dic[0][1]] == 3:
if len(workers[box_title].wear['unsafe wear']) == 0:
workers[box_title].wear['unsafe wear'].append(localtime)
sql = "INSERT INTO wear SET worker_id = '" + wid + "', type = 'unsafe wear',abnormal_time = '" + localtime + "'"
cur4.execute(sql)
cur4.close() # 关闭游标
else:
if localtime not in workers[box_title].wear['unsafe wear']:
workers[box_title].wear['unsafe wear'].append(localtime)
sql = "INSERT INTO wear SET worker_id = '" + wid + "', type = 'unsafe wear',abnormal_time = '" + localtime + "'"
cur4.execute(sql)
cur4.close() # 关闭游标
# 写入越线情况
if len(workers[box_title].track_point) > 4:
for i in range(len(transboundaryline)):
p1 = (transboundaryline[i][0], transboundaryline[i][1])
p2 = (transboundaryline[i][2], transboundaryline[i][3])
p3 = workers[box_title].track_point[-2]
p4 = workers[box_title].track_point[-1]
a = t.IsIntersec(p1, p2, p3, p4)
if a == '有交点':
cur5 = conn.cursor() # 获取一个游标
cur6 = conn.cursor() # 获取一个游标
cur5.execute(
"select time from transboundary where worker_id = '" + wid + "' ")
qurrytime = cur5.fetchone()
cur5.close() # 关闭游标
if qurrytime == None:
print('越线')
sql = "INSERT INTO transboundary SET worker_id = '" + wid + "',time = '" + localtime + "'"
cur6.execute(sql)
cur6.close() # 关闭游标
else:
temp1 = 0
for qt in qurrytime:
if qt == localtime:
temp1 = 1
if temp1 == 0:
print('越线')
sql = "INSERT INTO transboundary SET worker_id = '" + wid + "',time = '" + localtime + "'"
cur6.execute(sql)
cur6.close() # 关闭游标
if len(workers[box_title].track_point) >= 20:
workers[box_title].previous_point = workers[box_title].track_point[-5]
conn.commit()
try:
cv2.putText(frame, face[track2.track_id - 1], (int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200, (0, 255, 0),
2)
except Exception as e:
cv2.putText(frame, "unknow", (int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 255, 0), 2)
cv2.imshow('', frame)
if writeVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
list_file.write(str(frame_index) + ' ')
if len(boxs) != 0:
for i in range(0, len(boxs)):
list_file.write(str(boxs[i][0]) + ' ' + str(boxs[i][1]) + ' ' + str(boxs[i][2]) + ' ' + str(
boxs[i][3]) + ' ')
list_file.write('\n')
fps = (fps + (1. / (time.time() - t1))) / 2
if cv2.waitKey(1) & 0xFF == ord('q'):
break
def run_eval_mode(detection_model, args):
'''
Evaluate the tracker on MOT-16 training set
'''
train_sequence_names = ['MOT16-02', 'MOT16-04', 'MOT16-05', 'MOT16-09',
'MOT16-10', 'MOT16-11', 'MOT16-13']
# Print parameters before starting ----------------
logger.info("\nConfig ---- \n----Detector: {} \n----Online detection: {}".format(args.detector, args.online_detection))
for sequence_name in train_sequence_names:
logger.info("Processing sequence: {}".format(sequence_name))
sequence_dir = MOT16_TRAIN_DATA_DIR + sequence_name + '/'
if args.online_detection == 1:
if args.detector == 'DPM':
raise Exception(" Online detection is possible only when using SSD")
detection_file = None
else:
detection_dir = MOT16_DETECTION_DIR_BASE + args.detector + "/MOT16_POI_train/"
detection_file = detection_dir + sequence_name + '.npy'
detections = np.load(detection_file, allow_pickle=True)
#print("detections shape:", detections.shape)
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", args.max_cosine_distance, NN_BUDGET)
tracker = Tracker(metric)
results = []
encoder = gen_det.create_box_encoder(HUMAN_ENCODER_PATH, batch_size=32)
n_frames = len(seq_info['image_filenames'])
logger.info("----Total frames:{}".format(n_frames))
logger.info("----Detections file:{}".format(detection_file))
def frame_callback(vis, frame_idx):
logger.info("Processing Sequence {}, Frame {:05d}" .format(sequence_name, frame_idx))
frame = cv2.imread(seq_info['image_filenames'][frame_idx])
t1 = time.time()
if args.online_detection == 0: # If not online detection- Use pre-computed detections
# Load image and generate detections.
detection_list = create_detections(detections, frame_idx, MIN_DETECTION_HEIGHT)
detection_list = [d for d in detection_list if d.confidence >= args.min_confidence]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detection_list])
scores = np.array([d.confidence for d in detection_list])
indices = dsutil_prep.non_max_suppression(boxes, NMS_MAX_OVERLAP, scores)
detection_list = [detection_list[i] for i in indices]
else: # Use Mobilenet-SSD on the fly
# Detect humans
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
bboxes, detection_scores = detect_humans(detection_model, frame_rgb)
detection_list = cvt_to_detection_objects(frame, bboxes, detection_scores, encoder)
# Update tracker.
tracker.predict()
tracker.update(detection_list)
# FPS counter
fps = 1/(time.time()-t1)
# Update visualization.
if args.display == 1:
vis.set_image(frame.copy())
vis.draw_detections(detection_list)
vis.draw_trackers(tracker.tracks)
# Store results.
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlwh()
results.append([frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3], fps])
# Visualize
if args.display == 1:
visualizer = dsutil_viz.Visualization(seq_info, update_ms=5)
else:
visualizer = dsutil_viz.NoVisualization(seq_info)
visualizer.run(frame_callback)
# Store results.
if args.online_detection == 0: # If not online detection
output_dir = RESULTS_DIR + 'Pedestrian Tracking/' + 'EVAL_' + args.detector + '/trackOutput-{}minConf/'.format(args.min_confidence)
output_file_path = output_dir + sequence_name + '.txt'
else:
output_file_path = './Results/temp_hypotheses.txt'
with open(output_file_path, 'w') as output_file:
avg_fps = 0
for row in results:
output_file.write("{:d},{:d},{:.2f},{:.2f},{:.2f},{:.2f}\n".format(row[0], row[1], row[2], row[3], row[4], row[5]))
avg_fps += row[6]
avg_fps /= n_frames
def main(args):
logger.info('Start Tracking...')
ctx = mx.gpu(0) if args.gpu else mx.cpu()
fps = max(0, min(BASE_FPS, args.fps))
net = model_zoo.get_model(args.network, pretrained=True, ctx=ctx)
net.reset_class(classes=['person'], reuse_weights=['person'])
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# feature extractor for deepsort re-id
encoder = gdet.BoxEncoder()
metric = nn_matching.NearestNeighborDistanceMetric('cosine',
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
capture = cv2.VideoCapture(args.src)
frame_interval = BASE_FPS // fps if fps > 0 else 0
frame_index = 0
while True:
ret, frame = capture.read()
if ret != True:
break
if 0 < fps and frame_index % frame_interval != 0:
frame_index += 1
continue
x, img = gcv.data.transforms.presets.yolo.transform_test(
mx.nd.array(frame).astype('uint8'),
short=args.short,
)
class_IDs, det_scores, det_boxes = net(x.as_in_context(ctx))
boxs = []
person = mx.nd.array([0])
score_threshold = mx.nd.array([0.5])
for i, class_ID in enumerate(class_IDs[0]):
if class_ID == person and det_scores[0][i] >= score_threshold:
boxs.append(det_boxes[0][i].asnumpy())
if boxs:
features = encoder(img, boxs)
else:
features = np.array([])
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
frame_index += 1
# store original scene
cv2.imwrite(os.path.join(args.out_dir, f'{frame_index}.jpg'), img)
show_img = img.copy()
# check missed
for track in tracker.tracks:
bbox = [max(0, int(x)) for x in track.to_tlbr()]
if not track.is_confirmed() or track.time_since_update > 1:
if 2 <= track.time_since_update < 10:
try:
cv2.imwrite(
os.path.join(
args.out_dir,
f'missed-{frame_index}-{track.track_id}.jpg'),
img,
)
except:
traceback.print_exc()
logger.info('Skipped by time_since_update')
continue
logger.info(f'Frame #{frame_index} - Id: {track.track_id}')
cv2.rectangle(show_img, (bbox[0], bbox[1]), (bbox[2], bbox[3]),
(255, 255, 255), 2)
cv2.putText(show_img, str(track.track_id), (bbox[0], bbox[1] + 30),
0, 5e-3 * 200, (0, 255, 0), 2)
# show image
cv2.imwrite(os.path.join(args.out_dir, f'anno-{frame_index}.jpg'),
show_img)
cv2.imshow('', show_img)
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
logger.info(
f'Missed obj: {tracker.missed_obj}, Missed frame: {tracker.missed_frame}'
)
capture.release()
cv2.destroyAllWindows()
def process_frame(dataset):
# Definition of the parameters
to_process = {}
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
producer = KafkaProducer(
bootstrap_servers='master:6667',
value_serializer=lambda m: json.dumps(m).encode('utf8'))
data = dataset.collect()
# deep_sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
dt_now = dt.datetime.now()
for datum in data:
event = json.loads(datum[1])
dt_event = dt.datetime.strptime(event['timestamp'],
'%Y-%m-%dT%H:%M:%S.%f')
delta = dt_now - dt_event
print("timestamp = " + str(dt_event))
if delta.seconds > 5:
continue
to_process[event['camera_id']] = event
for key, event in to_process.items():
t1 = time.time()
event = json.loads(datum[1])
try:
decoded = base64.b64decode(event['image'])
except TypeError:
return
filename = '/home/haohsiang/Vigilancia-Distributed/codev1frame.jpg' # I assume you have a way of picking unique filenames
with open(filename, 'wb') as f:
f.write(decoded)
frame = cv2.imread(filename)
#ret, frame = video_capture.read() # frame shape 640*480*3
image = Image.fromarray(frame[..., ::-1]) #bgr to rgb
boxs = yolo.detect_image(image)
print("box_num", len(boxs))
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 255, 0), 2)
for det in detections:
bbox = det.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 0, 0), 2)
#sent result to kafka
if len(boxs) > 0:
print(
str(track.track_id) + ' :' + str(bbox[0]) + ' ' +
str(bbox[1]) + ' ' + str(bbox[2]) + ' ' + str(bbox[3]))
print(dt.datetime.now().time())
result = {
'ID': str(track.track_id),
'timestamp': dt.datetime.now().isoformat(),
'location_x': str(bbox[0]),
'w': str(bbox[2])
}
producer.send('position', result)
producer.flush()
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
def run(sequence_dir, detection_file, output_file, min_confidence, nms_max_overlap, min_detection_height, max_cosine_distance,
nn_budget, display):
"""
Run multi-target tracker on a particular sequence.
Parameters
----------
sequence_dir : str
Path to the MOTChallenge sequence directory.
detection_file : str
Path to the detections file.
output_file : str
Path to the tracking output file. This file will contain the tracking
results on completion.
min_confidence : float
Detection confidence threshold. Disregard all detections that have
a confidence lower than this value.
nms_max_overlap: float
Maximum detection overlap (non-maxima suppression threshold).
min_detection_height : int
Detection height threshold. Disregard all detections that have
a height lower than this value.
max_cosine_distance : float
Gating threshold for cosine distance metric (object appearance).
nn_budget : Optional[int]
Maximum size of the appearance descriptor gallery. If None, no budget
is enforced.
display : bool
If True, show visualization of intermediate tracking results.
"""
# 收集流的信息，包括图片名称，检测结果以及置信度
seq_info = gather_sequence_info(sequence_dir, detection_file)
# metric 实例化 nn_matching 的 NearestNeighborDistanceMetric 类，输入的初始距离度量函数是 cosine，此时可以传入 euclidean
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
# 创建一个追踪器对象
tracker = Tracker(metric)
results = []
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
# 根据之前的参数生成检测框，并且只保留置信度 confidence 大于 min_confidence 的检测框
detections = create_detections(seq_info["detections"], frame_idx, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# Run non-maximal suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Update tracker.
tracker.predict()
tracker.update(detections)
# Update visualization.
if display:
image = cv2.imread(seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
vis.set_image(image.copy())
vis.draw_detections(detections)
vis.draw_trackers(tracker.tracks)
# Store results.
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlwh()
results.append([frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3]])
# Run tracker.
if display:
visualizer = visualization.Visualization(seq_info, update_ms=5)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
# Store results.
f = open(output_file, 'w')
for row in results:
print('%d,%d,%.2f,%.2f,%.2f,%.2f,1,-1,-1,-1' % (
row[0], row[1], row[2], row[3], row[4], row[5]), file=f)
def Display(self, yolo):
self.ui.Open.setEnabled(False)
self.ui.Close.setEnabled(True)
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# deep_sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
writeVideo_flag = True
if writeVideo_flag:
# Define the codec and create VideoWriter object
w = int(self.cap.get(3))
h = int(self.cap.get(4))
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
out = cv2.VideoWriter("output.avi", fourcc, 15, (w, h))
list_file = open('detection.txt', 'w')
frame_index = -1
fps = 0.0
while self.cap.isOpened():
ret, frame = self.cap.read()
if ret != True:
break
t1 = time.time()
image = Image.fromarray(frame)
boxs = yolo.detect_image(image)
# print("box_num",len(boxs))
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(
boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
str1 = "Identification List:" + '\n\n'
for track in tracker.tracks:
if track.is_confirmed() and track.time_since_update > 1:
continue
bbox = track.to_tlbr()
import random
if track.track_id == 2 and self.fileName.split(
"/")[-1] == "reid.mp4":
str1 = str1 + " person" + str(
track.track_id) + " —— LaiSiyu (" + str(
random.random() * 0.1 + 0.85) + ")" + '\n\n'
# str1 += " person{0} - LaiSiyu ({:%.2f})".format(track.track_id, random.random() * 0.13 + 0.85)
ui.label_2.setText(str1)
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (0, 0, 255), 2)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 0, 255), 2)
else:
str1 = str1 + " person" + str(track.track_id) + '\n\n'
ui.label_2.setText(str1)
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])),
(255, 255, 255), 2)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1])), 0, 5e-3 * 200,
(0, 255, 0), 2)
for det in detections:
bbox = det.to_tlbr()
# cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), (255, 0, 0), 2)
cv2.imshow('', frame) # TODO
# RGB转BGR
frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
img = QImage(frame.data, frame.shape[1], frame.shape[0],
QImage.Format_RGB888)
self.ui.DisplayLabel.setPixmap(QPixmap.fromImage(img))
if writeVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
list_file.write(str(frame_index) + ' ')
if len(boxs) != 0:
for i in range(0, len(boxs)):
list_file.write(
str(boxs[i][0]) + ' ' + str(boxs[i][1]) + ' ' +
str(boxs[i][2]) + ' ' + str(boxs[i][3]) + ' ')
list_file.write('\n')
fps = (fps + (1. / (time.time() - t1))) / 2
print("fps= %f" % (fps))
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if self.isCamera:
cv2.waitKey(1)
else:
cv2.waitKey(int(1000 / self.frameRate))
# 判断关闭事件是否已触发
if True == self.stopEvent.is_set():
# 关闭事件置为未触发，清空显示label
self.stopEvent.clear()
self.ui.DisplayLabel.clear()
self.ui.label_2.clear()
self.ui.Close.setEnabled(False)
self.ui.Open.setEnabled(True)
break
self.cap.release()
if writeVideo_flag:
out.release()
list_file.close()
cv2.destroyAllWindows()
def main(yolo):
# Determining the FPS of a video having variable frame rate
# cv2.CAP_PROP_FPS is not used since it returns 'infinity' for variable frame rate videos
filename = "clip1.mp4"
# Determining the total duration of the video
clip = VideoFileClip(filename)
cap2 = cv2.VideoCapture(filename)
co = 0
ret2 = True
while ret2:
ret2, frame2 = cap2.read()
# Determining the total number of frames
co += 1
cap2.release()
# Computing the average FPS of the video
Input_FPS = co / clip.duration
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
frame_count = 0
# Implementing Deep Sort algorithm
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
# Cosine distance is used as the metric
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
video_capture = cv2.VideoCapture(filename)
# Define the codec and create a VideoWriter object to save the output video
out = cv2.VideoWriter(
'output.mp4', cv2.VideoWriter_fourcc(*'MP4V'), Input_FPS,
(int(video_capture.get(3)), int(video_capture.get(4))))
# To calculate the frames processed by the deep sort algorithm per second
fps = 0.0
# Initializing empty variables for counting and tracking purpose
queue_track_dict = {} # Count time in queue
queue_track_dict_2 = {}
alley_track_dict = {} # Count time in alley
store_track_dict = {} # Count total time in store
latest_frame = {} # Track the last frame in which a person was identified
reidentified = {
} # Yes or No : whether the person has been re-identified at a later point in time
plot_head_count_store = [] # y-axis for Footfall Analysis
plot_head_count_queue = [] # y-axis for Footfall Analysis
plot_time = [] # x-axis for Footfall Analysis
# Loop to process each frame and track people
while True:
ret, frame = video_capture.read()
if ret != True:
break
maxIntensity = 255.0
phi = 1
theta = 1
newImage1 = (maxIntensity / phi) * (frame /
(maxIntensity / theta))**1.3
frame = array(newImage1, dtype=uint8)
cv2.imwrite('testing.jpg', frame)
if frame_count == 5000:
break
head_count_store = 0
head_count_queue = 0
t1 = time.time()
image = Image.fromarray(frame[..., ::-1]) # BGR to RGB conversion
boxs = yolo.detect_image(image)
features = encoder(frame, boxs)
# Getting the detections having score of 0.0 to 1.0
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression on the bounding boxes
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker to associate tracking boxes to detection boxes
tracker.predict()
tracker.update(detections)
# Defining the co-ordinates of the area of interest
pts3 = np.array([[133, 120], [303, 120], [297, 175], [127, 175]],
np.int32)
pts3 = pts3.reshape((-1, 1, 2))
pts2 = np.array(
[[380, 250], [380, 365], [175, 480], [0, 480], [0, 380]], np.int32)
pts2 = pts2.reshape((-1, 1, 2)) # Queue Area - 1
pts = np.array([[0, 380], [0, 0], [640, 0], [640, 480], [170, 480],
[380, 360], [380, 250]], np.int32)
pts = pts.reshape((-1, 1, 2)) # Alley Region
cv2.polylines(frame, [pts], True, (255, 0, 255), thickness=1)
cv2.polylines(frame, [pts3], True, (248, 197, 39), thickness=2)
cv2.polylines(frame, [pts2], True, (0, 255, 255), thickness=2)
# Drawing tracker boxes and frame count for people inside the areas of interest
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
# Checking if the person is within an area of interest
queue_point_test_2 = center_point_inside_polygon(bbox, pts3)
queue_point_test = center_point_inside_polygon(bbox, pts2)
alley_point_test = center_point_inside_polygon(bbox, pts)
# Checking if a person has been reidentified in a later frame
if queue_point_test == 'inside' or alley_point_test == 'inside' or queue_point_test_2 == 'inside':
if track.track_id in latest_frame.keys():
if latest_frame[track.track_id] != frame_count - 1:
reidentified[track.track_id] = 1
# Initializing variables incase a new person has been seen by the model
if queue_point_test == 'inside' or alley_point_test == 'inside' or queue_point_test_2 == 'inside':
head_count_store += 1
if track.track_id not in store_track_dict.keys():
store_track_dict[track.track_id] = 0
queue_track_dict[track.track_id] = 0
queue_track_dict_2[track.track_id] = 0
alley_track_dict[track.track_id] = 0
reidentified[track.track_id] = 0
# Processing for people inside the Queue Area - 1
if queue_point_test == 'inside':
head_count_queue += 1
queue_track_dict[track.track_id] += 1
latest_frame[track.track_id] = frame_count
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (0, 255, 255), 2)
cv2.rectangle(frame, (int(bbox[0]) - 1, int(bbox[1]) - 15),
(int(bbox[2]) + 1, int(bbox[1])), (0, 255, 255),
-1)
wait_time = round(
(queue_track_dict[track.track_id] / Input_FPS), 2)
cv2.putText(frame,
str(wait_time) + "s", (int(bbox[0]), int(bbox[1])),
5, 0.9, (0, 0, 0), 1)
# Processing for people inside the Queue Area - 2
if queue_point_test_2 == 'inside':
head_count_queue += 1
queue_track_dict_2[track.track_id] += 1
latest_frame[track.track_id] = frame_count
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (248, 197, 39), 2)
cv2.rectangle(frame, (int(bbox[0]) - 1, int(bbox[1]) - 15),
(int(bbox[2]) + 1, int(bbox[1])), (248, 197, 39),
-1)
wait_time = round(
(queue_track_dict_2[track.track_id] / Input_FPS), 2)
cv2.putText(frame,
str(wait_time) + "s", (int(bbox[0]), int(bbox[1])),
5, 0.9, (0, 0, 0), 1)
# Processing for people inside the Alley Region
if alley_point_test == 'inside':
alley_track_dict[track.track_id] += 1
latest_frame[track.track_id] = frame_count
# cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])),(255,255,255), 2)
# cv2.putText(frame, str(track.track_id), (int(bbox[0]), int(bbox[1])), 0, 0.8, (0, 0, 0), 4)
# cv2.putText(frame, str(track.track_id), (int(bbox[0]), int(bbox[1])), 0, 0.8, (0, 255, 77), 2)
# Getting the total Store time for a person
if track.track_id in store_track_dict.keys():
store_track_dict[track.track_id] = queue_track_dict[
track.track_id] + alley_track_dict[track.track_id]
# Drawing bounding box detections for people inside the store
for det in detections:
bbox = det.to_tlbr()
# Checking if the person is within an area of interest
queue_point_test = center_point_inside_polygon(bbox, pts)
alley_point_test = center_point_inside_polygon(bbox, pts2)
# if queue_point_test == 'inside' or alley_point_test == 'inside':
# cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])), (255,0,0), 2)
# Video Overlay - Head Count Data at that instant
cv2.putText(frame, "Count: " + str(head_count_store), (30, 610),
cv2.FONT_HERSHEY_COMPLEX_SMALL, 1.5, (0, 0, 0), 3,
cv2.LINE_AA, False)
cv2.putText(frame, "Count: " + str(head_count_store), (30, 610),
cv2.FONT_HERSHEY_COMPLEX_SMALL, 1.5, (0, 255, 77), 2,
cv2.LINE_AA, False)
# Calculating the average wait time in queue
total_people = len([v for v in queue_track_dict.values() if v > 0])
total_queue_frames = sum(v for v in queue_track_dict.values() if v > 0)
avg_queue_frames = 0
if total_people != 0:
avg_queue_frames = total_queue_frames / total_people
avg_queue_time = round((avg_queue_frames / Input_FPS), 2)
# Video Overlay - Average Wait Time in Queue
cv2.putText(frame, "Avg Queue Time: " + str(avg_queue_time) + 's',
(30, 690), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1.5, (0, 0, 0),
3, cv2.LINE_AA, False)
cv2.putText(frame, "Avg Queue Time: " + str(avg_queue_time) + 's',
(30, 690), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1.5,
(0, 255, 77), 2, cv2.LINE_AA, False)
# Calculating the average wait time in the store
total_people = len(store_track_dict)
total_store_frames = sum(store_track_dict.values())
avg_store_frames = 0
if total_people != 0:
avg_store_frames = total_store_frames / total_people
avg_store_time = round((avg_store_frames / Input_FPS), 2)
# Video Overlay - Average Store time
cv2.putText(frame, "Avg Store Time: " + str(avg_store_time) + 's',
(30, 650), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1.5, (0, 0, 0),
3, cv2.LINE_AA, False)
cv2.putText(frame, "Avg Store Time: " + str(avg_store_time) + 's',
(30, 650), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1.5,
(0, 255, 77), 2, cv2.LINE_AA, False)
# Write the frame onto the VideoWriter object
out.write(frame)
# Calculating the frames processed per second by the model
fps = (fps + (1. / (time.time() - t1))) / 2
frame_count += 1
# Printing processing status to track completion
op = "FPS_" + str(frame_count) + "/" + str(co) + ": " + str(
round(fps, 2))
print("\r" + op, end="")
# Adding plot values for Footfall Analysis every 2 seconds (hard coded for now)
if frame_count % 50 == 0:
plot_time.append(round((frame_count / Input_FPS), 2))
plot_head_count_store.append(head_count_store)
plot_head_count_queue.append(head_count_queue)
# Press Q to stop the video
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# Defining data to be written into the csv file - Detailed Report
csv_columns = ['Unique Person ID', 'Queue Time in AOI', 'Total Store Time']
csv_data = []
csv_row = {}
detailed_csv_file = 'Detailed_Store_Report.csv'
queue_unique_count = 0
for k, v in store_track_dict.items():
csv_row = {}
csv_row = {
csv_columns[0]: k,
csv_columns[1]: round((queue_track_dict[k] / Input_FPS), 2),
csv_columns[2]: round((v / Input_FPS), 2)
}
if round((queue_track_dict[k] / Input_FPS), 2) > 1:
queue_unique_count += 1
csv_data.append(csv_row)
# Writing the data into the csv file - Detailed Report
with open(detailed_csv_file, 'w') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=csv_columns)
writer.writeheader()
for data in csv_data:
writer.writerow(data)
# Defining data to be written into the csv file - Detailed Report
csv_columns = ['Unique Person ID', 'Queue Time in AOI', 'Total Store Time']
csv_data = []
csv_row = {}
detailed_csv_file = 'Detailed_Store_Report_2.csv'
queue_unique_count = 0
for k, v in store_track_dict.items():
csv_row = {}
csv_row = {
csv_columns[0]: k,
csv_columns[1]: round((queue_track_dict_2[k] / Input_FPS), 2),
csv_columns[2]: round((v / Input_FPS), 2)
}
if round((queue_track_dict_2[k] / Input_FPS), 2) > 1:
queue_unique_count += 1
csv_data.append(csv_row)
# Writing the data into the csv file - Detailed Report
with open(detailed_csv_file, 'w') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=csv_columns)
writer.writeheader()
for data in csv_data:
writer.writerow(data)
# Defining data to be written into the csv file - Brief Report
csv_columns_brief = [
'Total Head Count', 'Total Queue Time', 'Average Queue Time',
'Total Store Time', 'Average Store Time'
]
brief_csv_file = 'Brief_Store_Report.csv'
csv_data_brief = {
csv_columns_brief[0]:
queue_unique_count,
csv_columns_brief[1]:
round((sum(queue_track_dict.values()) / Input_FPS), 2),
csv_columns_brief[2]:
avg_queue_time,
csv_columns_brief[3]:
round((sum(store_track_dict.values()) / Input_FPS), 2),
csv_columns_brief[4]:
avg_store_time
}
# Writing the data into the csv file - Brief Report
with open(brief_csv_file, 'w') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=csv_columns_brief)
writer.writeheader()
writer.writerow(csv_data_brief)
# Releasing objects created
video_capture.release()
out.release()
cv2.destroyAllWindows()
def main(yolo):
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# Deep SORT
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
tracking = True
writeVideo_flag = False
asyncVideo_flag = False
file_path = 'D:\\video/[Sala Outside][2020-05-28T16-01-39][2020-05-28T18-02-09].mp4'
if asyncVideo_flag:
video_capture = VideoCaptureAsync(file_path)
else:
video_capture = cv2.VideoCapture(file_path)
if asyncVideo_flag:
video_capture.start()
if writeVideo_flag:
if asyncVideo_flag:
w = int(video_capture.cap.get(3))
h = int(video_capture.cap.get(4))
else:
w = int(video_capture.get(3))
h = int(video_capture.get(4))
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter('output_yolov4.avi', fourcc, 30, (w, h))
frame_index = -1
fps = 0.0
fps_imutils = imutils.video.FPS().start()
while True:
ret, frame = video_capture.read() # frame shape 640*480*3
if ret != True:
break
t1 = time.time()
image = Image.fromarray(frame[..., ::-1]) # bgr to rgb
boxes, confidence, classes = yolo.detect_image(image)
if tracking:
features = encoder(frame, boxes)
detections = [
Detection(bbox, confidence, cls, feature)
for bbox, confidence, cls, feature in zip(
boxes, confidence, classes, features)
]
else:
detections = [
Detection_YOLO(bbox, confidence, cls)
for bbox, confidence, cls in zip(boxes, confidence, classes)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
if tracking:
# Call the tracker
tracker.predict()
tracker.update(detections)
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
cv2.putText(frame, "ID: " + str(track.track_id),
(int(bbox[0]), int(bbox[1])), 0,
1.5e-3 * frame.shape[0], (0, 255, 0), 1)
for det in detections:
bbox = det.to_tlbr()
score = "%.2f" % round(det.confidence * 100, 2) + "%"
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 0, 0), 2)
if len(classes) > 0:
cls = det.cls
cv2.putText(frame,
str(cls) + " " + score,
(int(bbox[0]), int(bbox[3])), 0,
1.5e-3 * frame.shape[0], (0, 255, 0), 1)
cv2.imshow('', frame)
if writeVideo_flag: # and not asyncVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
fps_imutils.update()
if not asyncVideo_flag:
fps = (fps + (1. / (time.time() - t1))) / 2
print("FPS = %f" % (fps))
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
fps_imutils.stop()
print('imutils FPS: {}'.format(fps_imutils.fps()))
if asyncVideo_flag:
video_capture.stop()
else:
video_capture.release()
if writeVideo_flag:
out.release()
cv2.destroyAllWindows()
def main(yolo):
start = time.time()
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
counter = []
#deep_sort
model_filename = 'model_data/market1501.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
find_objects = ['person']
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
writeVideo_flag = True
video_capture = cv2.VideoCapture(args["input"])
if writeVideo_flag:
# Define the codec and create VideoWriter object
w = int(video_capture.get(3))
h = int(video_capture.get(4))
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
out = cv2.VideoWriter('./output/output.avi', fourcc, 15, (w, h))
list_file = open('detection_rslt.txt', 'w')
frame_index = -1
fps = 0.0
while True:
ret, frame = video_capture.read() # frame shape 640*480*3
if ret != True:
break
t1 = time.time()
#image = Image.fromarray(frame)
image = Image.fromarray(frame[..., ::-1]) #bgr to rgb
boxs, confidence, class_names = yolo.detect_image(image)
# print(class_names,":",boxs)
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature, class_name)
for bbox, feature, class_name in zip(boxs, features, class_names)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
i = int(0)
indexIDs = []
c = []
boxes = []
for det in detections:
bbox = det.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255), 2)
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
class_name = track.get_class()
# boxes.append([track[0], track[1], track[2], track[3]])
indexIDs.append(int(track.track_id))
counter.append(int(track.track_id))
bbox = track.to_tlbr()
color = [int(c) for c in COLORS[indexIDs[i] % len(COLORS)]]
list_file.write(str(frame_index) + ',')
list_file.write(str(track.track_id) + ',')
list_file.write(str(class_name) + ',')
# cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])), (int(bbox[2]), int(bbox[3])),(color), 3)
b0 = str(bbox[0]
) #.split('.')[0] + '.' + str(bbox[0]).split('.')[0][:1]
b1 = str(bbox[1]
) #.split('.')[0] + '.' + str(bbox[1]).split('.')[0][:1]
b2 = str(bbox[2] - bbox[0]
) #.split('.')[0] + '.' + str(bbox[3]).split('.')[0][:1]
b3 = str(bbox[3] - bbox[1])
list_file.write(
str(b0) + ',' + str(b1) + ',' + str(b2) + ',' + str(b3))
list_file.write('\n')
list_file.write(str(track.track_id) + ',')
# print(class_name)
if class_name == "car":
_, image0 = video_capture.read()
print([int(bbox[0]), int(bbox[1]), int(bbox[2]), int(bbox[3])])
cv2.putText(frame,
str(track.track_id) + class_name,
(int(bbox[0]), int(bbox[1] - 10)), 0, 5e-3 * 150,
(color), 2)
i += 1
#bbox_center_point(x,y)
center = (int(
((bbox[0]) + (bbox[2])) / 2), int(((bbox[1]) + (bbox[3])) / 2))
#track_id[center]
pts[track.track_id].append(center)
thickness = 5
#center point
cv2.circle(frame, (center), 1, color, thickness)
# draw motion path
for j in range(1, len(pts[track.track_id])):
if pts[track.track_id][j - 1] is None or pts[
track.track_id][j] is None:
continue
thickness = int(np.sqrt(64 / float(j + 1)) * 2)
cv2.line(frame, (pts[track.track_id][j - 1]),
(pts[track.track_id][j]), (color), thickness)
#cv2.putText(frame, str(class_names[j]),(int(bbox[0]), int(bbox[1] -20)),0, 5e-3 * 150, (255,255,255),2)
count = len(set(counter))
cv2.putText(frame, "Total Pedestrian Counter: " + str(count),
(int(20), int(120)), 0, 5e-3 * 200, (0, 255, 0), 2)
cv2.putText(frame, "Current Pedestrian Counter: " + str(i),
(int(20), int(80)), 0, 5e-3 * 200, (0, 255, 0), 2)
cv2.putText(frame, "FPS: %f" % (fps), (int(20), int(40)), 0,
5e-3 * 200, (0, 255, 0), 3)
cv2.namedWindow("YOLO4_Deep_SORT", 0)
cv2.resizeWindow('YOLO4_Deep_SORT', 1024, 768)
cv2.imshow('YOLO4_Deep_SORT', frame)
if writeVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
fps = (fps + (1. / (time.time() - t1))) / 2
out.write(frame)
frame_index = frame_index + 1
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
print(" ")
print("[Finish]")
end = time.time()
if len(pts[track.track_id]) != None:
print(args["input"][43:57] + ": " + str(count) + " " +
str(class_name) + ' Found')
else:
print("[No Found]")
#print("[INFO]: model_image_size = (960, 960)")
video_capture.release()
if writeVideo_flag:
out.release()
list_file.close()
cv2.destroyAllWindows()
def main(_argv):
# Definition of the parameters
max_cosine_distance = 0.4
nn_budget = None
nms_max_overlap = 1.0
# initialize deep sort
model_filename = 'model_data/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
# calculate cosine distance metric
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
# initialize tracker
tracker = Tracker(metric)
# load configuration for object detector
config = ConfigProto()
config.gpu_options.allow_growth = False
config.gpu_options.per_process_gpu_memory_fraction = 0.1
_ = InteractiveSession(config=config)
utils.load_config(FLAGS)
input_size = FLAGS.size
video_path = FLAGS.video
# load tflite model if flag is set
if FLAGS.framework == 'tflite':
interpreter = tf.lite.Interpreter(
model_path=f'{FLAGS.weights}_{FLAGS.size}')
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
print(input_details)
print(output_details)
# otherwise load standard tensorflow saved model
else:
saved_model_loaded = tf.saved_model.load(
f'{FLAGS.weights}_{FLAGS.size}', tags=[tag_constants.SERVING])
infer = saved_model_loaded.signatures['serving_default']
# begin video capture
try:
vid = cv2.VideoCapture(int(video_path))
except:
vid = cv2.VideoCapture(video_path)
out = None
# get video ready to save locally if flag is set
if FLAGS.output:
# by default VideoCapture returns float instead of int
width = int(vid.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(vid.get(cv2.CAP_PROP_FPS))
codec = cv2.VideoWriter_fourcc(*FLAGS.output_format)
out = cv2.VideoWriter(FLAGS.output, codec, fps, (width, height))
all_start_time = None
frame_num = 0
# while video is running
while True:
return_value, frame = vid.read()
if return_value:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
Image.fromarray(frame)
else:
fps = float(frame_num) / (time.time() - all_start_time)
print("fps=%.2f size=%d frames=%d deep=%s output=%s" %
(fps, FLAGS.size, frame_num,
"true" if FLAGS.deep else "false", FLAGS.output))
break
frame_num += 1
if FLAGS.info:
print("frame_num=%d" % frame_num)
start_time = time.time()
if all_start_time is None:
all_start_time = time.time()
image_data = cv2.resize(frame, (input_size, input_size))
image_data = image_data / 255.
image_data = image_data[np.newaxis, ...].astype(np.float32)
# run detections on tflite if flag is set
if FLAGS.framework == 'tflite':
interpreter.set_tensor(input_details[0]['index'], image_data)
interpreter.invoke()
pred = [
interpreter.get_tensor(output_details[i]['index'])
for i in range(len(output_details))
]
# run detections using yolov3 if flag is set
if FLAGS.model == 'yolov3' and FLAGS.tiny == True:
boxes, pred_conf = filter_boxes(pred[1],
pred[0],
score_threshold=0.25,
input_shape=tf.constant(
[input_size, input_size]))
else:
boxes, pred_conf = filter_boxes(pred[0],
pred[1],
score_threshold=0.25,
input_shape=tf.constant(
[input_size, input_size]))
else:
batch_data = tf.constant(image_data)
pred_bbox = infer(batch_data)
for _, value in pred_bbox.items():
boxes = value[:, :, 0:4]
pred_conf = value[:, :, 4:]
boxes, scores, classes, valid_detections = tf.image.combined_non_max_suppression(
boxes=tf.reshape(boxes, (tf.shape(boxes)[0], -1, 1, 4)),
scores=tf.reshape(
pred_conf,
(tf.shape(pred_conf)[0], -1, tf.shape(pred_conf)[-1])),
max_output_size_per_class=50,
max_total_size=50,
iou_threshold=FLAGS.iou,
score_threshold=FLAGS.score)
# convert data to numpy arrays and slice out unused elements
num_objects = valid_detections.numpy()[0]
bboxes = boxes.numpy()[0]
bboxes = bboxes[0:int(num_objects)]
scores = scores.numpy()[0]
scores = scores[0:int(num_objects)]
classes = classes.numpy()[0]
classes = classes[0:int(num_objects)]
# format bounding boxes from normalized ymin, xmin, ymax, xmax ---> xmin, ymin, width, height
original_h, original_w, _ = frame.shape
bboxes = utils.format_boxes(bboxes, original_h, original_w)
# store all predictions in one parameter for simplicity when calling functions
pred_bbox = [bboxes, scores, classes, num_objects]
# read in all class names from config
class_names = utils.read_class_names(cfg.YOLO.CLASSES)
# by default allow all classes in .names file
allowed_classes = list(class_names.values())
# custom allowed classes (uncomment line below to customize tracker for only people)
allowed_classes = ['person']
# loop through objects and use class index to get class name, allow only classes in allowed_classes list
names = []
deleted_indx = []
for i in range(num_objects):
class_indx = int(classes[i])
class_name = class_names[class_indx]
if class_name not in allowed_classes:
deleted_indx.append(i)
else:
names.append(class_name)
names = np.array(names)
count = len(names)
if FLAGS.count:
cv2.putText(frame, "Objects being tracked: {}".format(count),
(5, 35), cv2.FONT_HERSHEY_COMPLEX_SMALL, 2,
(0, 255, 0), 2)
print("Objects being tracked: {}".format(count))
# delete detections that are not in allowed_classes
bboxes = np.delete(bboxes, deleted_indx, axis=0)
scores = np.delete(scores, deleted_indx, axis=0)
# encode yolo detections and feed to tracker
if FLAGS.deep:
features = encoder(frame, bboxes)
else:
features = np.empty((len(bboxes), 0), np.float32)
detections = [
Detection(bbox, score, class_name, feature)
for bbox, score, class_name, feature in zip(
bboxes, scores, names, features)
]
#initialize color map
cmap = plt.get_cmap('tab20b')
colors = [cmap(i)[:3] for i in np.linspace(0, 1, 20)]
# run non-maxima supression
boxs = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
classes = np.array([d.class_name for d in detections])
indices = preprocessing.non_max_suppression(boxs, classes,
nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
# update tracks
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
class_name = track.get_class()
# draw bbox on screen
color = colors[int(track.track_id) % len(colors)]
color = [i * 255 for i in color]
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), color, 2)
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1] - 30)),
(int(bbox[0]) +
(len(class_name) + len(str(track.track_id))) * 17,
int(bbox[1])), color, -1)
cv2.putText(frame, class_name + "-" + str(track.track_id),
(int(bbox[0]), int(bbox[1] - 10)), 0, 0.75,
(255, 255, 255), 2)
# if enable info flag then print details about each track
if FLAGS.info:
print(
"Tracker ID: {}, Class: {}, BBox Coords (xmin, ymin, xmax, ymax): {}"
.format(str(track.track_id), class_name, (int(
bbox[0]), int(bbox[1]), int(bbox[2]), int(bbox[3]))))
result = np.asarray(frame)
result = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
if not FLAGS.dont_show:
cv2.imshow("Output Video", result)
# if output flag is set, save video file
if FLAGS.output:
out.write(result)
# calculate frames per second of running detections
if FLAGS.info:
fps = 1.0 / (time.time() - start_time)
print("fps=%.2f" % fps)
if not FLAGS.dont_show:
if cv2.waitKey(1) & 0xFF == ord('q'): break
if not FLAGS.dont_show:
cv2.destroyAllWindows()
border_line = [(0, 400), (1200, 400)]
path_track = 20 # how many frames in path are saves
detector = YOLOv4(
input_shape=(HEIGHT, WIDTH, 3),
anchors=YOLOV4_ANCHORS,
num_classes=80,
training=False,
yolo_max_boxes=64,
yolo_iou_threshold=0.3,
yolo_score_threshold=0.4,
)
detector.load_weights("models/yolov4.h5")
max_cosine_distance = 0.2 # 03
encoder = gdet.create_box_encoder("models/mars-small128.pb", batch_size=64)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance, None)
tracker = Tracker(metric)
border_lines = {'border1': [[0, 100], [312, 104]]}
def drawBorderLines(frame):
for b in border_lines:
a = border_lines[b][0]
b = border_lines[b][1]
length = 40
vX0 = b[0] - a[0]
vY0 = b[1] - a[1]
mag = math.sqrt(vX0 * vX0 + vY0 * vY0)
vX = vX0 / mag
vY = vY0 / mag
temp = vX
def main(yolo):
t0 = time.time()
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
counter = []
#deep_sort
model_filename = 'model_data/market1501.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1) #跟踪使用
find_objects = ['person']
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
writeVideo_flag = True
video_capture = cv2.VideoCapture(args["input"])
if writeVideo_flag:
# Define the codec and create VideoWriter object
w = int(video_capture.get(3))
h = int(video_capture.get(4))
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
if args["ids"] == False:
out = cv2.VideoWriter('./output/output%s.avi' % args["camera"][1],
fourcc, 50, (w, h))
else:
out = cv2.VideoWriter(
'./output/output%s_reid.avi' % args["camera"][1], fourcc, 50,
(w, h))
list_file = open('detection_rslt.txt', 'w')
frame_index = -1
nump = 1
#fps = 0.0
while True:
ret, frame = video_capture.read() # frame shape 640*480*3
if ret != True:
break
t1 = time.time()
frame2 = copy.deepcopy(frame)
#image = Image.fromarray(frame)
image = Image.fromarray(frame[..., ::-1]) #bgr to rgb 仅yolo使用
boxs, confidence, class_names = yolo.detect_image(image)
print(boxs)
features = encoder(frame, boxs)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(boxs, features)
]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# Call the tracker
tracker.predict()
tracker.update(detections)
i = int(0)
indexIDs = []
c = []
boxes = []
makequery = True
for det in detections:
bbox = det.to_tlbr()
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
indexIDs.append(int(track.track_id))
counter.append(int(track.track_id))
bbox = track.to_tlbr()
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (255, 255, 255),
2) #跟踪框
color = [int(c) for c in COLORS[indexIDs[i] % len(COLORS)]]
list_file.write(str(frame_index) + ',') #3-5-7-9
list_file.write(str(track.track_id) + ',') #画面内的所有人id
b0 = str(bbox[0]
) #.split('.')[0] + '.' + str(bbox[0]).split('.')[0][:1]
b1 = str(bbox[1]
) #.split('.')[0] + '.' + str(bbox[1]).split('.')[0][:1]
b2 = str(bbox[2] - bbox[0]
) #.split('.')[0] + '.' + str(bbox[3]).split('.')[0][:1]
b3 = str(bbox[3] - bbox[1])
#放置id
list_file.write(
str(b0) + ',' + str(b1) + ',' + str(b2) + ',' + str(b3))
list_file.write('\n')
if len(class_names) > 0:
class_name = class_names[0]
cv2.putText(frame, str(class_names[0]),
(int(bbox[0]), int(bbox[1] - 20)), 0, 5e-3 * 150,
(255, 255, 255), 2) #person
i += 1
center = (int(
((bbox[0]) + (bbox[2])) / 2), int(((bbox[1]) + (bbox[3])) / 2))
pts[track.track_id].append(center)
thickness1 = 5
for j in range(1, len(pts[track.track_id])):
if pts[track.track_id][j - 1] is None or pts[
track.track_id][j] is None:
continue
thickness = int(np.sqrt(64 / float(j + 1)) * 2)
cv2.line(frame, (pts[track.track_id][j - 1]),
(pts[track.track_id][j]), (255, 255, 255), thickness)
if args["ids"] == False:
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (color), 3)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1] - 50)), 0, 5e-3 * 150,
(color), 2) #id
cv2.circle(frame, (center), 1, color, thickness1)
for j in range(1, len(pts[track.track_id])):
if pts[track.track_id][j - 1] is None or pts[
track.track_id][j] is None:
continue
thickness = int(np.sqrt(64 / float(j + 1)) * 2)
cv2.line(frame, (pts[track.track_id][j - 1]),
(pts[track.track_id][j]), (color), thickness)
try:
num = (int(args["camera"][1]) - 1) * 200
path = 'Z:\\pro2\\whole\\person\\gallery\\%04d' % int(
track.track_id + num)
if not os.path.exists(path):
os.makedirs(path)
if len(os.listdir(path)) <= 150: #最多存储150张相片
crop = frame2[int(bbox[1]):int(bbox[3]),
int(bbox[0]):int(bbox[2])]
crop = cv2.resize(crop, (64, 128),
interpolation=cv2.INTER_AREA
) #CUBIC 对扩大图片 area 对缩小图片
filepath = path + '\\' + '%04d' % int(
track.track_id +
num) + '_%s_' % args["camera"] + '%04d' % int(
len(os.listdir(path)) +
1) + '_%.2f' % (video_capture.get(0) /
1000) + '.jpg' #%04d
cv2.imwrite(filepath, crop)
except:
continue
#单独索引
else:
makequery = False
id1 = int(args["ids"])
if int(track.track_id) == id1:
cv2.rectangle(frame, (int(bbox[0]), int(bbox[1])),
(int(bbox[2]), int(bbox[3])), (color), 3)
cv2.putText(frame, str(track.track_id),
(int(bbox[0]), int(bbox[1] - 50)), 0,
5e-3 * 150, (color), 2) #id
cv2.circle(frame, (center), 1, color, thickness1)
for j in range(1, len(pts[track.track_id])):
if pts[track.track_id][j - 1] is None or pts[
track.track_id][j] is None:
continue
thickness = int(np.sqrt(64 / float(j + 1)) * 2)
cv2.line(frame, (pts[track.track_id][j - 1]),
(pts[track.track_id][j]), (color), thickness)
cv2.putText(frame, str(class_names[0]),
(int(bbox[0]), int(bbox[1] - 20)), 0,
5e-3 * 150, (color), 2) #person
else:
continue
count = len(set(counter))
cv2.putText(frame, "Total Pedestrian Counter: " + str(count),
(int(20), int(80)), 0, 5e-3 * 200, (0, 255, 0), 2)
cv2.putText(frame, "Current Pedestrian Counter: " + str(i),
(int(20), int(40)), 0, 5e-3 * 200, (0, 255, 0), 2)
cv2.namedWindow("YOLO4_Deep_SORT", 0)
cv2.resizeWindow('YOLO4_Deep_SORT', 1024, 768)
cv2.imshow('YOLO4_Deep_SORT', frame)
if writeVideo_flag:
# save a frame
out.write(frame)
frame_index = frame_index + 1
out.write(frame)
frame_index = frame_index + 1
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
#makequery
if makequery == True:
root_path = 'Z:\\pro2\\whole\\person\\gallery\\'
copy_path = 'Z:\\pro2\\whole\\person\\query\\'
ids = os.listdir(root_path)
#print(ids)
for i in ids:
img_path = root_path + i
img = os.listdir(img_path)
indeximg = img[int(len(img) / 2)]
old_name = img_path + '\\' + indeximg
new_path = copy_path + i
new_name = new_path + '\\' + indeximg
if not os.path.exists(new_path):
os.makedirs(new_path)
shutil.copyfile(old_name, new_name)
print(" ")
print("[Finish]")
end = time.time()
print("the whole time ", end - t0)
if len(pts[track.track_id]) != None:
print(args["input"][43:57] + ": " + str(count) + " " +
str(class_name) + ' Found')
else:
print("[No Found]")
video_capture.release()
if writeVideo_flag:
out.release()
list_file.close()
cv2.destroyAllWindows()