def __init__(self, class_id, names, setofpoint, multi, MOI, ROI, info,
frame_delay):
self.max_cosine_distance = 0.9
self.nms_max_overlap = 0.5
self.nn_budget = None
self.model_filename = 'model_data/market1501.pb'
self.detections = []
self.id = class_id
self.boxes_tracked = []
self.color = (255, 255, 255)
#self.encoder = gdet.create_box_encoder(self.model_filename,batch_size=1)
self.class_names = names
metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", self.max_cosine_distance, self.nn_budget)
self.obtracker = Tracker(metric, id_cam=int(info[0]))
self.pts = [deque(maxlen=1000) for _ in range(999999)]
#self.point_first = [deque(maxlen=2) for _ in range(999999)]
self.vis = visualization.Visualization(img_shape=(960, 1280, 3),
update_ms=2000)
COLORS = np.random.randint(100, 255, size=(255, 3), dtype="uint8")
self.color = [int(c) for c in COLORS[self.id]]
self.setofpoint = setofpoint
self.multi = multi
self.MOI = MOI
self.ROI = ROI
self.frame_track = 1
self.info = info
self.frame_delay = frame_delay
self.begin_time = 0
self.shape_img = None
def run(img, det, viewer, min_confidence, nms_max_overlap, tracker, display):
def frame_callback(vis, frame_idx):
print("Processing frame {:05d}".format(int(frame_idx) + 1))
detections = create_detections(det, frame_idx)
detections = [d for d in detections if d.confidence >= min_confidence]
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]
tracker.predict()
tracker.update(detections)
if display:
vis.set_image(img.copy())
vis.draw_trackers(tracker.tracks)
min_frame_idx = int(det[:, 0].min())
max_frame_idx = int(det[:, 0].max())
seq_info = {
"image_size": img.shape[:2],
"min_frame_idx": min_frame_idx,
"max_frame_idx": max_frame_idx
}
if display:
visualizer = visualization.Visualization(seq_info, viewer)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback, vedio_file='deep_sort1.avi')
def __init__(self, class_id, names):
self.max_cosine_distance = 0.9
self.nms_max_overlap = 0.5
self.nn_budget = None
self.model_filename = 'model_data/market1501.pb'
self.detections = []
self.id = class_id
self.boxes_tracked = []
self.color = (255, 255, 255)
#self.encoder = gdet.create_box_encoder(self.model_filename,batch_size=1)
self.class_names = names
metric = nn_matching.NearestNeighborDistanceMetric("cosine", self.max_cosine_distance, self.nn_budget)
self.obtracker = Tracker(metric)
self.pts = [deque(maxlen=100) for _ in range(999999)]
self.point_first = [deque(maxlen=2) for _ in range(999999)]
self.vis=visualization.Visualization(img_shape=(960,1280,3), update_ms=2000)
COLORS = np.random.randint(100, 255, size=(255, 3), dtype="uint8")
self.color = [int(c) for c in COLORS[self.id]]
def run(sequence_dir, det, output_file, min_confidence, nms_max_overlap,
max_cosine_distance, nn_budget, display):
seq_info = gather_sequence_info(sequence_dir, det)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
detections = create_detections(seq_info["detections"], frame_idx)
detections = [d for d in detections if d.confidence >= min_confidence]
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]
tracker.predict()
tracker.update(detections)
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)
if display:
viewer = ImageViewer(40, seq_info["image_size"][::-1],
"Figure KITTI-16")
visualizer = visualization.Visualization(seq_info, viewer)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback, vedio_file='deep_sort1.avi')
def run(sequence_dir, det_dir, checkpoint, output_file, max_age,
context_amount, iou, occlusion_thres, association_thres, display):
seq_info = gather_sequence_info(sequence_dir, det_dir)
tracker = MOT_Tracker(max_age, occlusion_thres, association_thres)
conf = configparser.ConfigParser()
conf.read(os.path.join(sequence_dir, 'seqinfo.ini'))
tracker.frame_rate = int(conf.get('Sequence', 'frameRate'))
if display:
gt = seq_info['groundtruth'][np.where(
seq_info['groundtruth'][:, -2] == 1)[0], :]
results = []
runtime = []
detections = seq_info["detections"]
def frame_callback(vis, detections, frame_idx, iou):
# Load image and generate detections.
print("Processing frame %05d" % frame_idx)
frame_path = seq_info['image_filenames'][int(frame_idx)]
detections = create_detections(detections, frame_idx)
# Update tracker.
before_time = time.time()
trackers = tracker.update(frame_path, checkpoint, context_amount,
detections, iou) # tracking
runtime.append(time.time() - before_time)
# Store results.
for d in trackers:
results.append([frame_idx, d[4], d[0], d[1], d[2], d[3]])
# 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_groundtruth(gt[np.where(gt[:,0] == frame_idx)[0], 1], \
gt[np.where(gt[:,0] == frame_idx)[0], 2: 6])
record_trks = [
t for t in tracker.tracks
if (t.is_tracked() and t.time_since_update <= 5)
]
vis.draw_trackers(record_trks)
# Run tracker.
if display:
visualizer = visualization.Visualization(seq_info,
detections,
iou,
update_ms=100)
else:
visualizer = visualization.NoVisualization(
seq_info,
detections,
iou,
)
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)
return sum(runtime) / len(runtime) * 1000
def run(sequence_dir,
result_file,
show_false_alarms=False,
detection_file=None,
update_ms=None,
video_filename=None):
"""Run tracking result visualization.
Parameters
----------
sequence_dir : str
Path to the MOTChallenge sequence directory.
result_file : str
Path to the tracking output file in MOTChallenge ground truth format.
show_false_alarms : Optional[bool]
If True, false alarms are highlighted as red boxes.
detection_file : Optional[str]
Path to the detection file.
update_ms : Optional[int]
Number of milliseconds between cosecutive frames. Defaults to (a) the
frame rate specifid in the seqinfo.ini file or DEFAULT_UDPATE_MS ms if
seqinfo.ini is not available.
video_filename : Optional[Str]
If not None, a video of the tracking results is written to this file.
"""
seq_info = deep_sort_app.gather_sequence_info(sequence_dir, detection_file)
results = np.loadtxt(result_file, delimiter=',')
if show_false_alarms and seq_info["groundtruth"] is None:
raise ValueError("No groundtruth available. Cannot show false alarms.")
def frame_callback(vis, frame_idx):
print("Frame idx", frame_idx)
image = cv2.imread(seq_info["image_filenames"][frame_idx],
cv2.IMREAD_COLOR)
vis.set_image(image.copy())
if seq_info["detections"] is not None:
detections = deep_sort_app.create_detections(
seq_info["detections"], frame_idx)
vis.draw_detections(detections)
mask = results[:, 0].astype(np.int) == frame_idx
track_ids = results[mask, 1].astype(np.int)
boxes = results[mask, 2:6]
vis.draw_groundtruth(track_ids, boxes)
if show_false_alarms:
groundtruth = seq_info["groundtruth"]
mask = groundtruth[:, 0].astype(np.int) == frame_idx
gt_boxes = groundtruth[mask, 2:6]
for box in boxes:
# NOTE(nwojke): This is not strictly correct, because we don't
# solve the assignment problem here.
min_iou_overlap = 0.5
if iou(box, gt_boxes).max() < min_iou_overlap:
vis.viewer.color = 0, 0, 255
vis.viewer.thickness = 4
vis.viewer.rectangle(*box.astype(np.int))
if update_ms is None:
update_ms = seq_info["update_ms"]
if update_ms is None:
update_ms = DEFAULT_UPDATE_MS
visualizer = visualization.Visualization(seq_info, update_ms)
if video_filename is not None:
visualizer.viewer.enable_videowriter(video_filename)
visualizer.run(frame_callback)
def run(video_file, min_confidence,
nms_max_overlap, min_detection_height, max_cosine_distance,
nn_budget, display):
"""Run multi-target tracker on a particular sequence.
Parameters
----------
video_file : str
Path to the video file.
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.
"""
cfg_file = "yolo3/cfg/yolov3.cfg"
weight_file = "yolo3/yolov3.weights"
#weight_file = 'yolo3/backup/MOT17Det/yolov3-mot17det_10000.weights'
use_cuda = 1
det_model = create_model(cfg_file, weight_file, use_cuda)
seq_info = gather_video_info(video_file)
metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
# just for warming up
img = cv2.imread('./000001.jpg')
sized = cv2.resize(img, (det_model.width, det_model.height))
sized = cv2.cvtColor(sized, cv2.COLOR_BGR2RGB)
boxes = do_detect(det_model, sized, 0.5, 0.4, use_cuda)
def frame_callback(vis, frame_idx):
#print("Processing frame %05d" % frame_idx)
#global total_frames, total_time
ret, img = seq_info['video_cap'].read()
if not ret:
print('there is no frame!')
sys.exit(1)
#time_0 = time.time()
# Load image and generate detections.
detections = create_det_from_model(det_model, img, 0.5, 0.4, min_detection_height, use_cuda)
#detections = create_detections(
# seq_info["detections"], frame_idx, min_detection_height)
#if seq_info['groundtruth'] is not None:
# gts = create_gts(seq_info['groundtruth'], frame_idx)
#detections = create_detections(
# seq_info["detections"], frame_idx, seq_info["image_filenames"][frame_idx], encoder, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# Update tracker.
tracker.predict()
tracker.update(detections)
#time_1 = time.time()
#total_time += time_1 - time_0
#total_frames += 1
# Update visualization.
if display:
#image = cv2.imread(
# seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
#vis.set_image(image.copy())
vis.set_image(img.copy())
#vis.draw_detections(detections)
#vis.draw_detections(gts)
vis.draw_trackers(tracker.tracks)
# Store results.
# NOTE: store from n_init frame(1-based index)
# for track in tracker.tracks:
# # NOTE: the condition is different from that in drawing tracks
# if not track.is_confirmed() or track.time_since_update > 1:
# continue
# # NOTE: store estimated state instead of observation
# 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=1)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
def run_multiple(sequence_dir, detection_dir, output_dir, min_confidence,
nms_max_overlap, min_detection_height, max_cosine_distance,
max_age, nn_budget, display, save_images_dir):
"""Run multi-target tracker on a particular sequence.
Parameters
----------
sequence_dir : str
Path to the MOTChallenge sequence directory.
detection_dir : str
Path to the detections file.
output_dir : 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.
save_images_dir : string
If not None, save the tracking result to indicated directories
"""
all_sequences = sorted(glob.glob(os.path.join(sequence_dir, '*')))
if len(all_sequences) == 0:
raise ValueError("There is no folder in " + sequence_dir)
for sequence_dir in all_sequences:
video_name = sequence_dir.split('/')[-1]
output_file = os.path.join(output_dir, video_name + '.npy') #'.txt')
print(video_name)
detection_file = os.path.join(detection_dir, video_name + '.npy')
try:
os.stat(detection_file)
os.stat(sequence_dir)
except:
raise NameError(detection_file + ' or ' + sequence_dir +
" doesn't exist!")
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric, max_age=max_age)
results = []
def frame_callback(vis, frame_idx):
# print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
detections = [
d for d in detections if d.confidence >= min_confidence
]
# 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]
# Update tracker.
tracker.predict()
tracker.update(detections)
# Update visualization.
if display:
image = cv2.imread(seq_info["image_filenames"][frame_idx],
cv2.IMREAD_COLOR)
image_name = seq_info["image_filenames"][frame_idx].split(
'/')[-1]
vis.set_image(image.copy(), image_name)
# 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()
# print("track._class:",track._class)
# print("track.confidence:",track.confidence)
# print("track.feature_to_save:",track.feature_to_save.shape)
results.append(
np.hstack([
frame_idx, track.track_id, bbox[0], bbox[1], bbox[2],
bbox[3], track._class, track.confidence,
track.feature_to_save
]))
# Run tracker.
if display:
visualizer = visualization.Visualization(
seq_info, update_ms=5, save_images_dir=save_images_dir)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
# Store results.
np.save(output_file, np.array(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)
# f.close()
# shutdown the window
if display:
cv2.destroyWindow(visualizer.viewer._caption)
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.
"""
print("init model....")
use_gpu = True
exact_list = ['7']
model = models.init_model(name='resnet50', num_classes=751, loss={'softmax', 'metric'}, use_gpu=use_gpu,
aligned=True)
checkpoint = torch.load("./checkpoint_ep300.pth.tar")
#global id
global_next_id = []
global_next_id.append(1)
# tmp2 = {}
# for item in checkpoint.items():
# key = item[0]
# values = item[1]
# newkey = key[7:]
# tmp2[newkey] = values
# model.load_state_dict(tmp2)
model.load_state_dict(checkpoint['state_dict'])
##sex and age
transform_test = transforms.Compose([
# transforms.Scale(224,224),
transforms.ToTensor(),
transforms.Normalize((0.429, 0.410, 0.382), (0.287, 0.285, 0.435))
])
model_sex_age = sex_age.ResNet50(sex_classes=2, age_classes=5)
tmp = torch.load('./sex_age/epoch5.pkl')
tmp2 = {}
for item in tmp.items():
key = item[0]
values = item[1]
newkey = key[7:]
tmp2[newkey] = values
model_sex_age.load_state_dict(tmp2)
model_sex_age.cuda()
model_sex_age.eval()
myexactor = FeatureExtractor(model, exact_list)
model.eval()
model.cuda()
seq_info_list = []
metric_list = []
tracker = []
angle_length = len(sequence_dir)
for index in range(0, angle_length):
seq_info_list.append(gather_sequence_info(sequence_dir[index], detection_file[index]))
metric_list.append(nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget))
tracker.append(Tracker(metric_list[index]))
results1,results2,results3 = [],[],[]
#define feature gallery
dic_feature = {}
def frame_callback(vis, frame_idx, seq_info, viewer,angle):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(
seq_info["detections"], frame_idx, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
image = cv2.imread(
seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
# Update tracker.
tracker[angle-1].predict()
tracker[angle-1].update(detections, image, myexactor, dic_feature, frame_idx, global_next_id, angle)
id_dic = {}
for index, track in enumerate(tracker[angle-1].tracks):
str_id = str(track.track_id)
if str_id in id_dic.keys():
track.track_id = global_next_id[0]
global_next_id[0] += 1
else:
id_dic[str_id] = (index, track.state)
# Update visualization.
if display:
vis.set_image(image.copy(),viewer,angle)
#print("deep_sort angle: "+str(angle))
vis.draw_detections(detections,viewer,angle)
vis.draw_trackers(tracker[angle-1].tracks,viewer,angle)
# Store results.
for track in tracker[angle-1].tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlwh()
print(angle)
if angle == 1:
print("angle1")
results1.append([
frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3], track.sex, track.person_age])
if angle == 2:
print("angle2")
results2.append([
frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3], track.sex, track.person_age])
if angle == 3:
print("angle3")
results3.append([
frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3], track.sex, track.person_age])
### save gallery
if (frame_idx) % 4 == 0:
for i in range(4):
if bbox[i] < 0 :
bbox[i] = 0
img = image[int(bbox[1]):int(bbox[1] + bbox[3]), int(bbox[0]):int(bbox[0] + bbox[2])]
img = cv2.resize(img, (128, 256), interpolation=cv2.INTER_CUBIC)
temp = img.copy()
transform_test = T.Compose([
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
img = transform_test(img)
img = torch.unsqueeze(img, 0)
##sex age forward
track_id = track.track_id
img = img.cuda()
sex_output, age_output = model_sex_age(img)
pred1 = sex_output.data.max(1)[1]
pred2 = age_output.data.max(1)[1]
age = age_dict[str(int(pred2))]
sex = sex_dict[str(int(pred1))]
track.person_age = age
track.sex = sex
##end
f1 = myexactor(img)
a1 = normalize(pool2d(f1[0], type='max'))
if str(track_id) not in dic_feature.keys():
dic_feature[str(track_id)] = []
dic_feature[str(track_id)].append((a1, angle, (bbox[0]+bbox[2]*0.5, bbox[1]+bbox[3])))
else:
if len(dic_feature[str(track_id)]) > 100:
del(dic_feature[str(track_id)][0])
dic_feature[str(track_id)].append((a1, angle, (bbox[0]+bbox[2]*0.5, bbox[1]+bbox[3])))
if display:
visualizer = visualization.Visualization(seq_info_list, update_ms=5,angle_length=angle_length)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback,angle_length)
# Store results.
f = open("./hypothese1.txt", 'w')
for row in results1:
print('%d,%d,%.2f,%.2f,%.2f,%.2f,%s,%s,1,-1,-1,-1' % (
row[0], row[1], row[2], row[3], row[4], row[5], row[6], row[7]),file=f)
f2 = open("./hypothese2.txt", 'w')
for row in results2:
print('%d,%d,%.2f,%.2f,%.2f,%.2f,%s,%s,1,-1,-1,-1' % (
row[0], row[1], row[2], row[3], row[4], row[5], row[6], row[7]), file=f2)
f3 = open("./hypothese3.txt", 'w')
for row in results3:
print('%d,%d,%.2f,%.2f,%.2f,%.2f,%s, %s, 1,-1,-1,-1' % (
row[0], row[1], row[2], row[3], row[4], row[5], row[6], row[7]), file=f3)
def run(image_dir, groundtruth, detections, min_confidence, nms_max_overlap,
min_detection_height, max_cosine_distance, nn_budget, display):
"""Run multi-target tracker on a particular sequence.
Parameters
----------
image_dir : str
Path to the MOTChallenge sequence directory.
detection_dir : str
Path to the detections file.
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.
"""
seq_info = gather_sequence_info(image_dir, groundtruth, detections)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# 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)
out = []
curr_idx = -1
img = None
for row in results:
if curr_idx != row[0]:
if img != None:
out += [img]
img = {
'videoName': video_name,
'name': '{}-{}.jpg'.format(video_name,
str(row[0]).zfill(6)),
'index': row[0],
'labels': []
}
curr_idx = row[0]
img['labels'] += [{
'category': None,
'box2d': {
'x1': row[2],
'y1': row[3],
'x2': row[2] + row[4],
'y2': row[3] + row[5]
},
'id': row[1]
}]
out += [img]
return out
def run(video_path,
track_path,
feat_path,
save_output,
out_dir,
cap_dir,
concat,
smoothing,
save_fig,
is_plot,
lag=30):
video_name = os.path.basename(video_path)
seq_info = gather_sequence_info(video_name, video_path, feat_path)
if (concat): track_path = track_path[:-4] + '_join.csv'
df = pd.read_csv(track_path, header=None)
if (smoothing): results = smooth(df, smooth_method='golay')
else: results = df.values
if (save_fig):
capture(video_path, cap_dir, results, seq_info, is_plot=is_plot)
return
cap = cv2.VideoCapture(video_path)
print('Video Path:', video_path, '\tFeatures:', feat_path)
print(lag)
points = {}
def frame_callback(vis, frame_idx):
# print("Frame idx", frame_idx)
image_np = np.array(cap.read()[1])
vis.set_image(image_np)
mask = results[:, 0].astype(np.int) == frame_idx
track_ids = results[mask, 1].astype(np.int)
boxes = results[mask, 2:6]
points[frame_idx] = []
for track_id, box in zip(track_ids, boxes):
l, t, w, h = np.array(box).astype(int)
x, y = int(l + w / 2), int(t + h)
points[frame_idx].append([track_id, x, y])
if (frame_idx > lag):
remove_idx = frame_idx - lag
if remove_idx in points:
del points[remove_idx]
vis.draw_groundtruth(track_ids, boxes, points)
visualizer = visualization.Visualization(seq_info, update_ms=50)
if save_output:
if (concat):
visualizer.viewer.enable_videowriter(
os.path.join(out_dir, video_name[:-4] + '_opt.mp4'))
else:
visualizer.viewer.enable_videowriter(
os.path.join(out_dir, video_name[:-4] + '_reg.mp4'))
visualizer.run(frame_callback)
cap.release()
cv2.destroyAllWindows()
def run(model, 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
----------
model: protobuffr
Path to tensorflow feature model
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.
"""
#encoder = create_box_encoder(model, batch_size=64)
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
#metric = nn_matching.NearestNeighborDistanceMetric(
# "euclidean", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
results = []
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
#if seq_info['groundtruth'] is not None:
# gts = create_gts(seq_info['groundtruth'], frame_idx)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# Update tracker.
tracker.predict()
"""
if display:
image = cv2.imread(
seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
vis.set_image(image.copy())
#vis.draw_detections(detections)
#vis.viewer.color = 0, 255, 0
#vis.draw_detections(gts)
#vis.draw_trackers(tracker.tracks)
"""
tracker.update(detections)
# if display:
# vis.viewer.color = 255, 0, 0
# vis.draw_trackers(tracker.tracks)
# 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_detections(gts)
vis.draw_trackers(tracker.tracks)
#"""
# Store results.
# NOTE: store from n_init frame(1-based index)
"""
for track in tracker.tracks:
# NOTE: the condition is different from that in drawing tracks
if not track.is_confirmed() or track.time_since_update > 1:
continue
# NOTE: store estimated state instead of observation
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=10)
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 run(sequence_dir, 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.
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.
"""
global total_frames, total_time
cfg_file = "yolo3/cfg/yolov3.cfg"
#weight_file = "yolo3/yolov3.weights"
weight_file = 'yolo3/backup/MOT17Det/yolov3-mot17det_10000.weights'
use_cuda = 1
det_model = create_model(cfg_file, weight_file, use_cuda)
seq_info = gather_sequence_info(sequence_dir, None)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
#metric = nn_matching.NearestNeighborDistanceMetric(
# "euclidean", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
results = []
img = cv2.imread(seq_info["image_filenames"][1])
sized = cv2.resize(img, (det_model.width, det_model.height))
sized = cv2.cvtColor(sized, cv2.COLOR_BGR2RGB)
boxes = do_detect(det_model, sized, 0.5, 0.4, use_cuda)
print("processing: %s" % os.path.basename(sequence_dir))
def frame_callback(vis, frame_idx):
#print("Processing frame %05d" % frame_idx)
global total_frames, total_time
img = cv2.imread(seq_info["image_filenames"][frame_idx])
time_0 = time.time()
# Load image and generate detections.
detections = create_det_from_model(det_model, img, 0.5, 0.4,
min_detection_height, use_cuda)
#detections = create_detections(
# seq_info["detections"], frame_idx, min_detection_height)
#if seq_info['groundtruth'] is not None:
# gts = create_gts(seq_info['groundtruth'], frame_idx)
#detections = create_detections(
# seq_info["detections"], frame_idx, seq_info["image_filenames"][frame_idx], encoder, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# Update tracker.
#tracker.predict()
#tracker.update(detections)
time_1 = time.time()
total_time += time_1 - time_0
total_frames += 1
# Update visualization.
if display:
#image = cv2.imread(
# seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
#vis.set_image(image.copy())
vis.set_image(img.copy())
vis.draw_detections(detections)
#vis.draw_detections(gts)
#vis.draw_trackers(tracker.tracks)
# Store results.
# NOTE: store from n_init frame(1-based index)
for track in tracker.tracks:
# NOTE: the condition is different from that in drawing tracks
if not track.is_confirmed() or track.time_since_update > 1:
continue
# NOTE: store estimated state instead of observation
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=1)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
print("average time for one frame: %.5f, FPS: %.5f" %
(total_time / total_frames, total_frames / total_time))
# 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 run_ghma(sequence_dir, detection_file, output_file,
min_confidence, nms_max_overlap, min_detection_height, max_cosine_distance, nn_budget,
display, run_type):
seq_info = gather_sequence_info(sequence_dir, detection_file)
image_shape = seq_info["image_size"][::-1]
aspect_ratio = float(image_shape[1]) / image_shape[0]
image_shape = 1024, int(aspect_ratio * 1024)
first_idx = seq_info["min_frame_idx"]
last_idx = seq_info["max_frame_idx"]
#if pre_computed:
if run_type == "pre_computed":
print(" *********** Pre_computed mode ***********")
if os.path.isfile('0130/01/det/det.txt'):
if not os.path.getsize('0130/01/det/det.txt'):
detFlag = False
else:
detFlag = True
else:
detFlag = False
if not detFlag:
net = load_net(cfg_path, weight_path, 0)
meta = load_meta(meta_path)
det_file = open('0130/01/det/det.txt', 'w')
for idx in range(first_idx, last_idx+1):
print(idx)
result = detect_(net, meta,
seq_info["image_filenames"][idx].encode('utf-8'),
11,
target=range(8),
thresh=min_confidence)
for j in range(len(result)):
det_file.write("%d,%d,%f,%f,%f,%f,%f,-1,-1,-1\n" %
(idx, result[j][0], result[j][2], result[j][3], result[j][4], result[j][5], result[j][1]))
det_file.close()
else:
print(">> Detections already exsits, skip yolo detection step")
if os.path.isfile('./temp/01.npy'):
if not os.path.getsize('./temp/01.npy'):
extFlag = False
else:
extFlag = True
else:
extFlag = False
if not extFlag:
f = create_box_encoder("resources/networks/mars-small128.ckpt-68577", batch_size=32, loss_mode="cosine")
generate_detections(f, "./0130/", "./temp/", None)
else:
print(">> Features already exists, skip extraction step")
seq_info = gather_sequence_info(sequence_dir, "./temp/01.npy")
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
results = []
elif run_type == "instant":
print(" *********** Instant Mode ***********")
encoder = create_box_encoder("resources/networks/mars-small128.ckpt-68577",
batch_size=32, loss_mode="cosine")
net = load_net(cfg_path, weight_path, 0)
meta = load_meta(meta_path)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
results = []
else:
raise Exception(" Unknown run type ")
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
if run_type == "pre_computed":
detections = create_detections(seq_info["detections"], frame_idx, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
elif run_type == "instant":
detections = calc_cur_det(net, meta, encoder, seq_info, frame_idx, min_confidence)
# 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]
# 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)
vis.save_image("./frame/{}.jpg".format(frame_idx))
# 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=100)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
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.
"""
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
results = []
"""
seq_info['detections'] --> array of shape (detections, 10 + feat vect size)
Every row corresponds to a single detection (every frame can have various detections and therefore various rows)
For each detection, store:
- The 10 first columns correspond to the MOT challenge notations:
[<frame>, <id>, <bb_left>, <bb_top>, <bb_width>, <bb_height>, <conf>, <x>, <y>, <z>]
OBS. The <id>, <x>, <y>, <z> params are set to -1.
- The rest of columns correspond to the feature vectors.
"""
det = seq_info["detections"]
images = sorted(seq_info['image_filenames'].keys())
def frame_callback(vis, frame_idx):
image_id = images.index(frame_idx)
det = seq_info["detections"][image_id:]
if image_id==366:
print()
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(
seq_info["detections"], image_id, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# 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([
image_id, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3], 1])
# Run tracker.
if display:
visualizer = visualization.Visualization(seq_info, update_ms=5) # 5
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
# Store results.
#def results2json(results, output_file):
data = list()
for i, row in enumerate(results):
print(i)
if i == 412:
print()
d = dict()
d['image_id'] = row[0]
d['bbox'] = [row[2], row[3], row[4], row[5]]
d['score'] = 1
d['category_id'] = 1
data.append(d)
with open(output_file, 'w') as outfile:
json.dump(data, outfile)
"""
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.
"""
# 初始化tracker
seq_info = gather_sequence_info(
sequence_dir, detection_file) # 加载检测文件detection_file：.npy
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
# 调用回调函数
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections. 加载图像并生成检测。
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height) #生成检测
detections = [d for d in detections if d.confidence >= min_confidence]
# Run non-maxima suppression. 运行非极大值抑制。
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. 进行跟踪
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 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.
"""
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
results = []
# Create label
# classs = []
#
# for i in range(len(tri)):
# classs.append(label[tri[i]])
# print(classs)
#
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
#input()
# Load image and generate detections.
detections = create_detections(
seq_info["detections"], frame_idx, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# print("Detections: ", detections)
# input()
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
label = np.array([d.label for d in detections])
#print('boxes',boxes)
#print('scores',scores)
indices = preprocessing.non_max_suppression(
boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# build label
label_name = ['Di bo','Xe dap','Xe may',
'Xe hang rong','Xe ba gac','Xe tac xi',
'Xe hoi','Xe ban tai','Xe cuu thuong',
'Xe khach','Xe buyt',
'Xe tai','Container','Xe cuu hoa']
label_actual = []
for i in label:
i = int(i)
#print('i',i)
label_actual.append(label_name[i])
# 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())
#print(boxes)
# vis.draw_detections(detections,label_actual)
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()
#print('label',track.label)
results.append([
frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3],track.label])
#print('Doneee')
# Run tracker.
if display:
print('seq_info',seq_info)
visualizer = visualization.Visualization(seq_info, update_ms=5)
else:
visualizer = visualization.NoVisualization(seq_info)
cam = sequence_dir.split("/")[-1]
#output_path_frames = os.path.join("saved_frames", cam)
visualizer.run(frame_callback)
#print("Sequence dir: ", sequence_dir)
print("Cam: ", cam)
#print(output_path_frames)
# Store results.
# for k,row in enumerate(results):
# print('%d,%d,%.2f,%.2f,%.2f,%.2f,1,-1,-1,-1'.format(
# row[0], row[1], row[2], row[3], row[4], row[5]))
#print("Ahihi")
f = open(output_file, 'w')
#print("The abs path: ", os.path.join(os.path.abspath(output_file), cam + ".txt"))
for k,row in enumerate(results):
print('%d,%d,%.2f,%.2f,%.2f,%.2f,1,%d,-1,-1' % (
row[0], row[1], row[2], row[3], row[4], row[5],row[6]),file=f)
def run(sequence_dir,
detection_file,
output_file,
min_confidence,
nms_max_overlap,
min_detection_height,
max_cosine_distance,
nn_budget,
display,
video_dir=None,
max_age=None,
world_viewer_threshold=None):
"""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.
"""
sequence_dir_list = []
video_dir_list = []
if video_dir:
video_dir_list = video_dir.split(",")
else:
sequence_dir_list = sequence_dir.split(",")
detection_file_list = detection_file.split(",")
camera_num = len(detection_file_list)
seq_info_dic = {}
tracker_dic = {}
global_track = []
global_id = [0]
for i in range(camera_num):
if video_dir:
seq_info_dic[i] = gather_video_info(video_dir_list[i],
detection_file_list[i])
else:
seq_info_dic[i] = gather_sequence_info(sequence_dir_list[i],
detection_file_list[i])
# print ("initial seq_info_dic is {}".format(seq_info_dic))
for i in range(camera_num):
tracker_dic[i] = Tracker(nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget),
i,
camera_num,
max_age=max_age)
results = []
tracking_file = open('tracking.txt', 'w+')
def video_frame_processing(vis, frame_idx, index, image):
# print("Processing frame %05d" % frame_idx)
frame_indices = seq_info_dic[index]["detections"][:, 0].astype(np.int)
mask = frame_indices == frame_idx
detection_list = []
for row in seq_info_dic[index]["detections"][mask]:
bbox, confidence, feature = row[1:5], row[5], row[6:]
# if bbox[3] < min_height:
# continue
detection_list.append(Detection(bbox, confidence, feature, index))
print('input nums: ', len(detection_list))
# filter detections by confidence
detections = [
d for d in detection_list if d.confidence >= min_confidence
]
#input()
# extract detection info
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
# filter detection by NMS
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
print('filtered nums: ', len(detections))
tracker_dic[index].predict()
matching_results = tracker_dic[index].update(detections, tracker_dic,
global_id, global_track,
world_viewer_threshold)
print('matching results: ', matching_results)
for res in matching_results:
det = detections[res[1]]
tracking_file.write('{} {} {} {} {} {} {}\n'.format(
frame_idx, det.tlwh[0], det.tlwh[1], det.tlwh[2], det.tlwh[3],
det.confidence, res[0]))
tracking_file.flush()
if display:
if index == 0:
vis.reset_image()
vis.append_image(image.copy())
vis.draw_trackers(tracker_dic[index].tracks, index)
# Store results.
# for i in range(camera_num):
# for track in tracker_dic[i].tracks:
# if not track.is_confirmed() or track.time_since_update > 1:
# continue
# bbox = track.to_tlwh()
# results.append([
# frame_idx, track.global_id, bbox[0], bbox[1], bbox[2], bbox[3]])
def frame_callback(vis, frame_idx, index):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info_dic[index]["detections"],
frame_idx, index, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# print ("the camera_index is {}, the coordiantes of detections is {}".format(index,[detection.tlwh for detection in detections]))
# Update tracker.
tracker_dic[index].predict()
tracker_dic[index].update(detections, tracker_dic, global_id,
global_track, world_viewer_threshold)
# Update visualization.
if display:
if index == 0:
vis.reset_image()
image = cv2.imread(
seq_info_dic[index]["image_filenames"][frame_idx],
cv2.IMREAD_COLOR)
vis.append_image(image.copy())
vis.draw_trackers(tracker_dic[index].tracks, index)
# Store results.
# for i in range(camera_num):
# for track in tracker_dic[i].tracks:
# if not track.is_confirmed() or track.time_since_update > 1:
# continue
# bbox = track.to_tlwh()
# results.append([
# frame_idx, track.global_id, bbox[0], bbox[1], bbox[2], bbox[3]])
# find the maximun frame id
last_frame = {}
max_frame_idx = 0
for i in range(camera_num):
last_frame[i] = seq_info_dic[i]["max_frame_idx"]
if last_frame[i] > max_frame_idx:
max_frame_idx = last_frame[i]
if video_dir == None:
visualizer = visualization.Visualization(seq_info_dic, global_id,
camera_num, 5, last_frame)
visualizer.run(frame_callback, tracker_dic, camera_num, max_frame_idx)
else:
# for video sequence
# print ("seq_info_dic is {}".format(seq_info_dic))
visualizer = visualization.Visualization(seq_info_dic, global_id,
camera_num, 5, last_frame,
video_dir_list)
visualizer.run(video_frame_processing, tracker_dic, camera_num,
max_frame_idx)
# 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 cam_detection(encoder, cam_id, min_confidence, max_cosine_distance,
nms_max_overlap, nn_budget):
cap = cv2.VideoCapture(0)
if not cap.isOpened():
print("Camera {} open failed.".format(cam_id))
return
ret, frame = cap.read()
seq_info = gather_sequence_info(cam_id, frame)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
jump_step = 3
print('Start frame_callback')
def frame_callback(vis, frame_idx):
print('Processing frame: {:d}'.format(frame_idx))
if not cap.isOpened(): return
st_time = time.time()
# Load image and generate detections.
ret, frame = cap.read()
# Object detection -- SSD
rclasses, rscores, rbboxes = process_image(
frame,
select_threshold=min_confidence,
nms_threshold=nms_max_overlap) #boxes in (x1, y1, x2, y2) format
# Select only person
cls_ind = 15
inds = np.where(rclasses == cls_ind)[0]
if len(inds) == 0:
dets = []
else:
cls_boxes = rbboxes[inds, :]
cls_boxes_v = cls_boxes.copy()
#convert normalized (ny1, nx1, ny2, nx2) to pixel coordinate (x1, y1, x2, y2)
cls_boxes_v[:, 1] = cls_boxes[:, 0] * frame.shape[0]
cls_boxes_v[:, 0] = cls_boxes[:, 1] * frame.shape[1]
cls_boxes_v[:, 3] = cls_boxes[:, 2] * frame.shape[0]
cls_boxes_v[:, 2] = cls_boxes[:, 3] * frame.shape[1]
cls_boxes_v[:, 2:] -= cls_boxes_v[:, :
2] #transfer to (x,y,w,h) format
cls_scores = rscores[inds]
dets = np.hstack(
(cls_boxes_v, cls_scores[:, np.newaxis])).astype(np.float32)
print('Filtered detection size: {}'.format(np.asarray(dets).shape))
# Feature extraction
detections = feature_extraction(encoder, frame, dets)
# Update tracker.
tracker.predict()
tracker.update(detections)
# Update visualization.
vis.set_image(frame.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]
])
end_time = time.time()
print('Processing time: {:.3f}s'.format(end_time - st_time))
# Run tracker.
visualizer = visualization.Visualization(seq_info, update_ms=5)
visualizer.run(frame_callback)
cap.release()
def run(sequence_dir, result_file, show_false_alarms=False, detection_file=None,
update_ms=None, video_filename=None, IDnum=0):
"""Run tracking result visualization.
Parameters
----------
sequence_dir : str
Path to the MOTChallenge sequence directory.
result_file : str
Path to the tracking output file in MOTChallenge ground truth format.
show_false_alarms : Optional[bool]
If True, false alarms are highlighted as red boxes.
detection_file : Optional[str]
Path to the detection file.
update_ms : Optional[int]
Number of milliseconds between cosecutive frames. Defaults to (a) the
frame rate specifid in the seqinfo.ini file or DEFAULT_UDPATE_MS ms if
seqinfo.ini is not available.
video_filename : Optional[Str]
If not None, a video of the tracking results is written to this file.
track_id : currently track id
IDnum : not 0 is foot tracking display
"""
seq_info = deep_sort_app.gather_sequence_info(sequence_dir, detection_file)
results = np.loadtxt(result_file, delimiter=',')
if show_false_alarms and seq_info["groundtruth"] is None:
raise ValueError("No groundtruth available. Cannot show false alarms.")
def frame_callback(vis, frame_idx):
#프레임별로 처리
print("Frame idx", frame_idx)
image = cv2.imread(
seq_info["image_filenames"][frame_idx], cv2.IMREAD_COLOR)
vis.set_image(image.copy())
if seq_info["detections"] is not None:
detections = deep_sort_app.create_detections(
seq_info["detections"], frame_idx)
vis.draw_detections(detections)
mask = results[:, 0].astype(np.int) == frame_idx
track_ids = results[mask, 1].astype(np.int) #해당 frame_id인 mask값들 중 [1]들인 id값 추출
boxes = results[mask, 2:6]
vis.draw_groundtruth(track_ids, boxes)
# 발위치 10개 중 y값만 빼기
h_file = os.path.dirname(result_file) # result/text/
with open(h_file + '/ID_h.txt', 'r') as f_hi:
line_splits = [int(l.split(',')[1]) for l in f_hi.read().splitlines()[1:]]
i = 0
#print(line_splits)
if show_false_alarms:
groundtruth = seq_info["groundtruth"]
mask = groundtruth[:, 0].astype(np.int) == frame_idx
gt_boxes = groundtruth[mask, 2:6]
for box in boxes:
# NOTE(nwojke): This is not strictly correct, because we don't
# solve the assignment problem here.
min_iou_overlap = 0.5
if iou(box, gt_boxes).max() < min_iou_overlap:
vis.viewer.color = 0, 0, 255
vis.viewer.thickness = 4
vis.viewer.rectangle(*box.astype(np.int))
if IDnum != 0: # Tracking 하는 ID만 보여주고 발 표시 !!!!!!!
vis.viewer.circle(
box[0] + box[2] / 2, box[1] + box[3], 3)
#if int(box[1] + box[3]) > line_splits[i] - 3 and int(bbox[1] + bbox[3]) < line_splits[i] + 3:
# vis.viewer.circle(
# box[0] + box[2] / 2, box[1] + box[3], 3)
# i += 1
if update_ms is None:
update_ms = seq_info["update_ms"]
if update_ms is None:
update_ms = DEFAULT_UPDATE_MS
visualizer = visualization.Visualization(seq_info, update_ms)
if video_filename is not None:
visualizer.viewer.enable_videowriter(video_filename)
visualizer.run(frame_callback)
def run(video_name, video_path, feat_path, track_path, min_confidence,
nms_max_overlap, min_detection_height, max_cosine_distance, nn_budget,
display):
seq_info = gather_sequence_info(video_name, video_path, feat_path)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric, max_age=50, n_init=5)
results = []
cap = cv2.VideoCapture(video_path)
print('Video Path:', video_path, '\tFeatures:', feat_path)
def frame_callback(vis, frame_idx):
# print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# Update tracker.
tracker.predict()
tracker.update(detections)
# Update visualization.
if display:
vis.set_image(cap.read()[1])
# vis.draw_detections(detections)
count_human = vis.draw_trackers(tracker.tracks)
else:
count_human = 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()
ID = count_human.index(track.track_id) + 1
results.append([frame_idx, ID, bbox[0], bbox[1], bbox[2], bbox[3]])
# Run tracker.
if display:
visualizer = visualization.Visualization(seq_info, update_ms=50)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
cap.release()
# Store results.
f = open(track_path, '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 run(sequence_dir,
detection_file,
output_file,
min_confidence,
nms_max_overlap,
min_detection_height,
max_cosine_distance,
nn_budget,
display,
stock=False,
track_class=None,
**kwargs):
"""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.
"""
if track_class is not None:
seq_info = gather_sequence_info(sequence_dir, detection_file,
track_class)
else:
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
if stock:
print('initializing a stock tracker')
tracker = Tracker(metric,
max_age=kwargs['max_age'],
max_iou_distance=1.0 - kwargs['min_iou_overlap'])
else:
print('initializing a modified tracker')
# the tracker now has the class as an optional argument
tracker = MyTracker(
metric,
max_age=kwargs['max_age'],
max_iou_distance=1.0 - kwargs['min_iou_overlap'],
tracker_type=kwargs["tracker_type"],
flow_dir=kwargs["flow_dir"],
update_kf=kwargs["update_kf"],
update_hit=kwargs["update_hit"]
) # the IOU is inverted as 1 - IOU in the cost matrix
results = []
if kwargs["track_subset_file"] is not None:
good_frames = np.loadtxt(kwargs["track_subset_file"])
else:
good_frames = None
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# this is is what should be called detections
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
# this is the high confidence detections
high_confidence_detections = [
d for d in detections if d.confidence >= min_confidence
]
# These are the low confidences ones and we don't need to run NMS on them because the goal is to retain as much information as possible
# these should be cmpletely disjoint
low_confidence_detections = [
d for d in detections if d.confidence < min_confidence
]
#HACK temporarily set the value of detections to None to make sure it's not used
detections = None
# Run non-maxima suppression.
# these should only be from high conf detections
# hc = high_conf
hc_boxes = np.array([d.tlwh for d in high_confidence_detections])
hc_scores = np.array(
[d.confidence for d in high_confidence_detections])
indices = preprocessing.non_max_suppression(hc_boxes, nms_max_overlap,
hc_scores)
hc_nms_positive_detections = [
high_confidence_detections[i] for i in indices
] # I think you can just do this by indexing
# this should negate the value from the line above
# there might be a cleaner way to do this with sets
hc_nms_negative_detections = [
high_confidence_detections[i]
for i in range(len(high_confidence_detections)) if i not in indices
]
assert len(hc_nms_positive_detections) + len(
hc_nms_negative_detections
) == len(
high_confidence_detections
), "This operation should just be partitioning detections into two subsets"
# Update tracker.
tracker.predict()
# read the next image because we will actually be using it now
image = cv2.imread(seq_info["image_filenames"][frame_idx],
cv2.IMREAD_COLOR)
assert image is not None, "the image now needs to be properly read"
if stock:
tracker.update(hc_nms_positive_detections)
else:
tracker.update(hc_nms_positive_detections,
bad_detections=hc_nms_negative_detections +
low_confidence_detections,
image=image,
use_unmatched=kwargs["use_unmatched"],
frame_idx=frame_idx)
# Update visualization.
if display:
vis.set_image(image.copy())
if seq_info['groundtruth'] is not None:
vis.draw_groundtruth(
*create_groundtruth(seq_info['groundtruth'], frame_idx))
#HACK for showing detections
#vis.draw_detections(hc_nms_positive_detections)
vis.draw_trackers(
tracker.tracks,
create_groundtruth(seq_info['groundtruth'], frame_idx)
) # clean up the double use of create_groundtruht
# Store results.
for track in tracker.tracks:
#MOD write all the tracks, even if not detected recently
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,
video_output_file=kwargs["video_output_file"],
vis_method=kwargs["vis_method"])
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback, good_frames)
# 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)
# score results
if seq_info['groundtruth'] is not None:
scorer = Scorer()
# I think the groundtruths might be off by a factor of 2
scores = scorer.score_lists(results, seq_info['groundtruth'],
good_frames)
assert type(
scores
) == pd.DataFrame, "The results are supposed to be a dataframe"
last_slash_idx = output_file.rfind("/")
scores_output_file = "{}scores{}".format(
output_file[:last_slash_idx + 1], output_file[last_slash_idx + 1:])
argv_output_file = "{}argv{}".format(output_file[:last_slash_idx + 1],
output_file[last_slash_idx + 1:])
# write the scores to a (very short) file
current_name = scores.index[0]
argv = re.sub("[',]", "", str(kwargs["argv"]))
scores = scores.rename({current_name: argv})
scores.to_csv(scores_output_file)
with open(argv_output_file, 'w') as outfile:
outfile.write(str(kwargs["argv"]))
else:
print("There are no groundtruths so no scoring can be done")
last_slash_idx = output_file.rfind("/")
argv_output_file = "{}argv{}".format(output_file[:last_slash_idx + 1],
output_file[last_slash_idx + 1:])
with open(argv_output_file, 'w') as outfile:
outfile.write(str(kwargs["argv"]))
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.
IDnum : int
Tracking ID_num
"""
# 프레임 사진 있는 위치, 영상번호 (seq_info["image_filenames"])
# 프레임번호 (seq_info["image_filenames"][frame_idx])
# 다 gather_sequence_info에 있음
if 'y' == input("is there ID [y/n] : "):
IDnum = int(input("ID_num for tracking : ")) #########
#range_down = input("ID tracking range_down : ") #########
#range_up = input("ID tracking range_up : ") #########
else:
IDnum = 0
if 'y' == input("foot display [y/n] : "): ############# y값 함수 구하고나면 발 위치보기
foot_dis = True
else:
foot_dis = False
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
target_h = []
update_ms = seq_info["update_ms"]
max_frame_idx = seq_info["max_frame_idx"]
i = 0
def frame_callback(vis, frame_idx):
#print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections]) # (x, y, w, h)
scores = np.array([d.confidence
for d in detections]) # Detector confidence score
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 or foot_dis:
image = cv2.imread(seq_info["image_filenames"][frame_idx],
cv2.IMREAD_COLOR)
vis.set_image(image.copy())
if IDnum == 0:
vis.draw_detections(detections)
vis.draw_trackers(tracker.tracks)
else: #찾는 ID 있을 때
vis.draw_target_trackers(tracker.tracks, IDnum)
if foot_dis: # Tracking 하는 ID만 보여주고 발 표시 !!!!!!!
vis.draw_foot(tracker.tracks, IDnum)
# h 저장
h_file = os.path.dirname(output_file) # result/text/
with open(h_file + '/ID_h.txt', 'r') as f_hi:
line_splits = [
int(l.split(',')[1]) for l in f_hi.read().splitlines()[1:]
]
# Store results.
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
############################################################### tracking 대신 bbox만 넣어주고 계산하기
bbox = track.to_tlwh()
results.append([
frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3]
])
if int(track.track_id) == int(IDnum):
#print("find ID-01")
if bbox[1] + bbox[3] < seq_info["image_size"][0]:
target_h.append(
[track.track_id, bbox[1] + bbox[3],
bbox[3]]) # id의 y값 별 h
if (
frame_idx >= 40
): # start frame 이걸로 설정 ################################
# MOT16-02에서 할머니 멈추기 전까지가 260frame
endT = 117 * update_ms / 1000
# endT = 5
vel_py.foot_world(frame_idx, track.track_id, bbox, IDnum,
update_ms, max_frame_idx, endT)
# velocity 추정 끝나면 count = 0으로 계산 그만시킴
# foot 10 보기
for i in range(10):
if int(bbox[1] + bbox[3]) > line_splits[i] - 1 and int(
bbox[1] + bbox[3]) < line_splits[i] + 1:
print(int(bbox[1] + bbox[3]))
vis.draw_foot(tracker.tracks, IDnum)
# Run tracker.
if display:
visualizer = visualization.Visualization(seq_info, update_ms=5)
else:
visualizer = visualization.NoVisualization(seq_info)
##### 영상 저장 ###########
#video_output_dir = os.path.curdir + '/result/video'
#video_filename = os.path.join(video_output_dir, "%s_all_tracking.avi" % (os.path.basename(sequence_dir))) # video name은 seq_info["sequence_name"]
#video_filename = os.path.join(video_output_dir, "%s_ID%s_tracking.avi" % (os.path.basename(sequence_dir), IDnum))
#video_filename = os.path.join(video_output_dir, "%s_ID%s_foot 10.avi" % (os.path.basename(sequence_dir), IDnum))
#if video_filename is not None:
# visualizer.viewer.enable_videowriter(video_filename)
#########################
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)
# [frame_idx, track.track_id, bbox[0], bbox[1], bbox[2], bbox[3]]
f.close()
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.
MOT
MOTChanllenge 序列文件所在的目录
detection_file : str
Path to the detections file.
检测的文件目录
output_file : str
Path to the tracking output file. This file will contain the tracking
results on completion.
追踪结果输出的文件目录
min_confidence : float
Detection confidence threshold. Disregard all detections that have
a confidence lower than this value.
最小置信度阈值
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.
最小检测框高度
max_cosine_distance : float
Gating threshold for cosine distance metric (object appearance).
余弦距离门控阈值
nn_budget : Optional[int]
Maximum size of the appearance descriptor gallery. If None, no budget
is enforced.
外观描述库的最大大小，如果大于该值，后边的外观描述信息会将旧的外观描述信息覆盖掉
display : bool
If True, show visualization of intermediate tracking results.
是否将追踪结果进行可视化的展示
"""
# 收集流信息，包括图片名称，检测结果以及置信度等
seq_info = gather_sequence_info(sequence_dir, detection_file)
# metic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)
# 将度量方式对象传入到追踪器对象中，创建一个追踪器对象
tracker = Tracker(metric)
results = []
# 循环帧：frame_idx < last_idx则运行frame_callback，且每次循环frame_idx += 1
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
# 筛选小于min_confidence的目标，并构造一个Detection对象构成的列表
# Detection是一个存储图中一个bbox结果
# 需要：1. bbox(tlwh形式) 2. 对应置信度 3. 对应embedding
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
# 筛选检测框
detections = [d for d in detections if d.confidence >= min_confidence]
# Run non-maxima suppression.
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
# 使用非极大抑制,找出局部范围内得分最大的框
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
# 更新检测框列表
detections = [detections[i] for i in indices]
# Update tracker.
# tracker给出一个预测结果，然后将detection传入，进行卡尔曼滤波操作
tracker.predict()
tracker.update(detections)
# Update visualization，
# 进行可是化操作，如果选择了跟踪结果可视化，图片框中展示的图片
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.
# 根据display选择不同的可视化对象
if display:
visualizer = visualization.Visualization(seq_info, update_ms=5)
else:
visualizer = visualization.NoVisualization(seq_info)
# 调用对应的run方法开始进行追踪。
visualizer.run(frame_callback)
# 将追踪的结果存储到output_file中
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 main():
start = time.time()
counter = []
writeVideo_flag = False
fps = 0.0
filename_path = os.path.join(result_path, 'submission.txt')
list_video, list_ids = load_list_video(list_video_path, id_path)
result_file = open(filename_path, 'w')
max_cosine_distance=0.8
nn_budget = 100
nms_max_overlap = 1.0
display = True
for video in list_video:
path = os.path.join(video_path, video)
ROI = load_roi(zones_path, video)
vis=visualization.Visualization(img_shape=(960,1280,3), update_ms=2000)
metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric)
results = []
print("Processing video: ", video )
video_capture = cv2.VideoCapture(path)
pause_display = False
frame_num = 0
while True:
start = time.time()
# print(count)
video_capture.set(cv2.CAP_PROP_POS_FRAMES, frame_num)
ret, frame = video_capture.read() # frame shape 640*480*3
# gray = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
# frame = np.zeros_like(frame1)
# frame[:,:,0] = gray
# frame[:,:,1] = gray
# frame[:,:,2] = gray
if ret != True:
break
# print(count1)
#print(frame.shape)
w = int(video_capture.get(3))
h = int(video_capture.get(4))
result = []
t1 = time.time()
img = letterbox(frame, new_shape=img_size)[0]
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416
img = np.ascontiguousarray(img)
dets = run_detect(model,img,device,frame)
detectionss=[]
for det in dets:
feature = gdet.HOG_feature(frame, det[:4])
detectionss.append(Detection(det[:4], det[4], feature, det[-1]))
#detectionss.append(Detection(det[:4], det[4], det[-1]) for det in dets)
img = np.zeros((h, w, 3), np.uint8)
img = frame.copy()
min_confidence = 0.4
detections = [d for d in detectionss if d.confidence >= min_confidence]
# 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]
# Update tracker.
tracker.predict()
tracker.update(detections)
if display:
vis.set_image(frame.copy())
vis.draw_detections(detections)
vis.draw_trackers(tracker.tracks)
res = vis.return_img()
draw_roi(ROI, res)
cv2.imshow('frame', res)
print('frame_num', frame_num)
if not pause_display:
key = cv2.waitKey(2)
if key == ord('q'):
break
if key == ord(' '):
pause_display = not pause_display
frame_num += 1
else:
key = cv2.waitKey(0)
if key == ord('q'):
break
if key == ord(' '):
pause_display = not pause_display
print(" ")
print("[Finish]")
video_capture.release()
if writeVideo_flag:
out.release()
#list_file.close()
result_file.close()
cv2.destroyAllWindows()
def run(min_confidence, nms_max_overlap, min_detection_height,
max_cosine_distance, nn_budget, max_time, feature_model):
encoder = create_box_encoder(feature_model, batch_size=32)
detector = Detector()
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
cap = cv2.VideoCapture(0)
seq_info = {
"sequence_name":
"real-time",
# "image_filenames": image_filenames,
# "detections": detections,
# "groundtruth": groundtruth,
"image_size": (int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)),
int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))),
"min_frame_idx":
0,
"max_frame_idx":
max_time * 20,
"feature_dim":
127,
"update_ms":
500
}
def frame_callback(vis, frame_idx):
# read image from camera
ret, frame = cap.read()
frame = np.array(frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
cap.release()
cv2.destroyAllWindows()
exit(0)
# generate bounding box
detections = detector.detect(frame.copy(), frame_idx)
# only display image if no detection
if detections.shape[0] <= 0:
vis.set_image(frame)
return
# generate deep feature
features = encoder(frame.copy(), detections[:, 2:6].copy())
detections = np.array([
np.r_[(row, feature)]
for row, feature in zip(detections, features)
])
detections = create_detections(detections, min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# Update tracker.
tracker.predict()
tracker.update(detections)
# Update visualization.
vis.set_image(frame)
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.
visualizer = visualization.Visualization(seq_info, update_ms=5)
visualizer.run(frame_callback)
cap.release()
cv2.destroyAllWindows()
def run(sequence_dir, detection_file, output_file, min_confidence,
nms_max_overlap, min_detection_height, max_cosine_distance, nn_budget,
display):
"""
multi-target tracker 실행
Parameters
----------
sequence_dir : str
시퀀스 데이터 경로
detection_file : str
검출 파일
output_file : str
output file 경로 / 추적 결과를 포함
min_confidence : float
confidence 임계값 / 이보다 낮은 confidence는 무시하게 된다.
nms_max_overlap: float
Maximum detection overlap (NMS 임계값)
min_detection_height : int
height 임계값 / 이보다 낮은 height는 무시하게 된다.
max_cosine_distance : float
cosine distance metric 임계값
nn_budget : Optional[int]
Maximum size of the appearance descriptor gallery. If None, no budget is enforced.
display : bool
True : 시각화
"""
# 시퀀스 정보수
seq_info = gather_sequence_info(sequence_dir, detection_file)
# cosine distance metric 사용
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
# 프레임 마다 호출
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# detection 생성
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# NMS 실행
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]
# tracker 업데이트 이 부분에서 detection을 업데이트 해서 객체를 추적한다.
# 내가 내 코드에 결합하고 싶을 때 사용하는 방법을 이해하기 위해 포스트를 작성했기 떄문에
# 자세하게 보고싶으면 SORT 포스트를 보자 비슷한거 같다.
tracker.predict()
tracker.update(detections)
# 시각화
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)
# 결과를 저장한다.
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]
])
# 객체를 추적하는 시작부분 frame_callback을 매 프레임 호출한다.
if display:
visualizer = visualization.Visualization(seq_info, update_ms=5)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
# 결과를 저장한다.
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 run(sequence_dir, detection_file, output_file, min_confidence,
nms_max_overlap, min_detection_height, max_cosine_distance, nn_budget,
display, clss):
"""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.
"""
capture = cv2.VideoCapture(sequence_dir)
seq_info = gather_sequence_info(sequence_dir, detection_file, capture)
capture.set(1, seq_info['min_frame_idx'])
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
seq_info['image_size'], clss,
min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# Update tracker.
tracker.predict()
tracker.update(detections)
# Update visualization.
if display:
#capture.set(1, frame_idx)
_, image = capture.read()
#pdb.set_trace()
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 > 15:
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)
# Post-process results.
results_array = np.array(results)
track_idxs = results_array[:, 1]
unique, counts = np.unique(track_idxs, return_counts=True)
#pdb.set_trace()
# for u, c in zip(unique, counts):
# if c < 100:
# results_array = results_array[track_idxs!=u]
# track_idxs = track_idxs[track_idxs!=u]
foo, unique_idxs = np.unique(track_idxs, return_inverse=True)
results_array[:, 1] = unique_idxs
results_processed = results_array.tolist()
# Store results.
f = open(output_file + '_' + clss + '.txt', 'w')
for row in results_processed:
print('%.2f,%.2f,%.2f,%.2f,%.2f,%.2f' %
(row[0], row[1], row[2], row[3], row[4], row[5]),
file=f)
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.
"""
seq_info = gather_sequence_info(sequence_dir, detection_file)
metric = nn_matching.NearestNeighborDistanceMetric("cosine",
max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
results = []
def frame_callback(vis, frame_idx):
print("Processing frame %05d" % frame_idx)
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
# 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 run(sequence_dir, detection_file, output_file, min_confidence,
nms_max_overlap, min_detection_height, max_cosine_distance, nn_budget,
display, offset, n_frames, max_iou_distance, max_age, n_init,
max_tracks, metric_param, alpha_ds):
"""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.
"""
seq_info = gather_sequence_info(sequence_dir, detection_file, offset,
n_frames)
metric = nn_matching.NearestNeighborDistanceMetric(alpha_ds,
max_cosine_distance,
nn_budget)
tracker = Tracker(metric,
max_iou_distance=max_iou_distance,
max_age=max_age,
n_init=n_init,
max_tracks=max_tracks,
metric_param=metric_param)
results = []
def frame_callback(vis, frame_idx):
aaa = n_frames / 10
if frame_idx % aaa == 0:
print("Processing frame {} / {}".format(frame_idx, n_frames))
# Load image and generate detections.
detections = create_detections(seq_info["detections"], frame_idx,
min_detection_height)
detections = [d for d in detections if d.confidence >= min_confidence]
# 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]
#print(len(detections))
# 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],
1, -1, -1, -1
])
# Run tracker.
if display:
visualizer = visualization.Visualization(seq_info, update_ms=5)
else:
visualizer = visualization.NoVisualization(seq_info)
visualizer.run(frame_callback)
#Reorder from 1 to n_tracks:
ids = list(np.unique(np.array(results)[:, 1]))
results = np.array(results)
results[:, 1] = [ids.index(x) for x in results[:, 1]]
# Store results.
np.save(output_file, results)
