sys.stdout = open(logfile, "w")
stream = NetVideoStream(args.server_ip, args.server_port, args.image_width,
args.image_height, args.image_channels)
if not stream.start():
print("[ERROR] could not connect to Server...")
exit(2)
signal.signal(signal.SIGUSR1, signal_handler)
result_video_writer = None
result_video_queue = None
if args.create_result_video:
result_video_queue = Queue(maxsize=args.video_queue_size)
result_video_writer = VideoWriter(args.image_width, args.image_height,
args.video_folder,
result_video_queue)
result_q = Queue(maxsize=args.result_queue_size)
result_server = ResultServer(args.client_port, result_q)
result_server.start()
config = tools.load_config(args.config)
od = ObjectDetection(config, args.image_width, args.image_height,
args.gpu_usage, args.threshold).start_up()
dt = t.time()
fps_dt = t.time()
fps = FPS().start()
cv2.imshow("retail_out", image)
cv2.waitKey(1)
t = Thread(target=read)
t.start()
thread1 = Thread(target=infer_yolo, args=(
timestamp,
margin,
points,
))
thread1.start()
video_writer = VideoWriter(
out_path.get() +
"/t_mobile_demo_out_{}.avi".format(time.strftime('%Y_%m_%d_%H_%M')),
image.shape[1], image.shape[0], 25)
while True:
# print(detector_op.is_closed())
trk_op.wait()
if trk_ip.is_closed():
# print("Here")
video_writer.finish()
break
ret, inference = trk_op.pull()
if ret:
trackers = inference.get_result()
frame = inference.get_input().get_image()
patches = inference.get_data()
# trails = inference.get_meta_dict()['trails']
def start():
try:
# ssdmobilenet v2 labels
LABELS = config.LABELS
# face recognition labels
le = pickle.loads(open(config.LABEL_PATH, 'rb').read())
#classes to be detected
LIST_CLASSES = 'person'.split(',')
# recognizer label (svm model)
recognizer = pickle.loads(open(config.MODEL_PATH, 'rb').read())
# load face features exractor from disk
embedder = FeatureExtractor(config.EMBEDDING_MODEL_PATH)
# load face recognizer from disk
faceDetector = FaceDetector(config.FACE_DECTOR_PATH)
# load pretrained ssdmobilenet from disk
pedestrianDetector = PedestrianDetector(
config.SSD_MOBILENET_V2_PATH,
faceDetector.detector.device_path())
# trackers
#load deepsort model from disk
encoder = gdet.create_box_encoder(config.TRACKER_MODEL_PATH,
batch_size=1)
#instatiate metrics
metric = nn_matching.NearestNeighborDistanceMetric(
'cosine', args['cosine_distance'], None)
# instatiate tracker
tracker = Tracker(metric)
if not args.get('input', False):
cap = cv2.VideoCapture(1)
else:
cap = cv2.VideoCapture(args['input'])
# instatiate video writer
videoWriter = VideoWriter(args['output'],
cap.get(cv2.CAP_PROP_FPS))
# let's declare some variable
H, W = None, None
# objects dict holds different ids and their corresponding class name
objects = {}
# holds the tracked objects and the number of consecutive frames they have been marked as disappeared
disappearedIds = {}
# total frame counts will be used to compute average fps
frameCounts = 0
# frames determine number of frame allows for an object to be disappeared
#frames = 0
# track object id
trackedOjects = []
# video writer process
process = None
# inference start time
startTime = datetime.datetime.now()
while cap.isOpened():
grabbed, frame = cap.read()
# stop while loop if is there is no frame available
if not grabbed:
break
#stop while loop if q is pressed
key = cv2.waitKey(1)
if key == ord('q'):
break
if H is None or W is None:
H, W = frame.shape[:2]
# start video writer process
if args['output'] and process is None:
writeVideo = Value('i', 1)
frameQueue = Queue()
process = Process(target=videoWriter.start,
args=(writeVideo, frameQueue, H, W))
process.start()
if frameCounts % 3 == 0:
results = pedestrianDetector.detect(
frame, args['confidence'])
bboxes = []
scores = []
#frames += 1
#loop over detection results
for r in results:
# label index
label_id = int(r.label_id + 1)
# check if detected object is in the list of the object we are interrested in.
if label_id in LABELS.keys(
) and LABELS[label_id] in LIST_CLASSES:
# scale bounding box
bbox = r.bounding_box.flatten() * np.array(
[W, H, W, H])
bbox = bbox.astype('int')
# add object bounding box to bboxes
bboxes.append(bbox)
# add detection confidence to scores list
scores.append(r.score)
bboxes = np.array(bboxes)
scores = np.array(scores)
if frameCounts % 3 == 0:
#grab feartures from frame
features = encoder(frame, bboxes)
# perform deep sort detection
detections = [
Detection(bbox, score, feature)
for bbox, score, feature in zip(
bboxes, scores, features)
]
# grab boxes, scores, and classes_name from deep sort detections
pedestrianBoxes = np.array([d.tlwh for d in detections])
pedestrianScores = np.array(
[d.confidence for d in detections])
#perform non-maxima suppression on deep sort detections
indexes = preprocessing.non_max_suppression(
pedestrianBoxes, args['nms_threshold'],
pedestrianScores)
detections = [detections[i] for i in indexes]
# update tracker
tracker.predict()
tracker.update(detections)
# loop over tracked objects
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
pedestrianBox = track.to_tlbr()
startX, startY, endX, endY = int(pedestrianBox[0]), int(
pedestrianBox[1]), int(pedestrianBox[2]), int(
pedestrianBox[3])
trackedOjects.append(track.track_id)
# Perfrom face detection only if there is no object in objects dict or current track id is not objects keys
if len(objects) == 0 or track.track_id not in objects.keys(
):
faces = faceDetector.detect(frame, .3)
# loop over detected faces
for face in faces:
bbox = face.bounding_box.flatten() * np.array(
[W, H, W, H])
xmin, ymin, xmax, ymax = bbox.astype('int')
# check if detexcted face bboxe is inside current pedestrian bbox
isOverlap = helper.overlap(
(startX, startY, endX, endY),
(xmin, ymin, xmax, ymax))
#
if isOverlap:
# grab ROI (face in our case) and then pass it to face feature extractor
roi = frame[ymin:ymax, xmin:xmax]
vector = embedder.run(roi)
# preprocess the feature vector extracted
vector = np.expand_dims(vector, axis=0)
# pass the preprocessed vector to perform prediction us trained ML model (SVM in our case)
proba = recognizer.predict_proba(vector)
ind = np.argmax(proba[0])
# filter out weak face detection
if proba[0][ind] > .6:
label = le.classes_[ind]
# perform list comprehension to avoid two different ids using same name.
if label in objects.values():
#and track.track_id not in objects.keys():
#del objects[track.track_id]
objects = {
k: v
for k, v in objects.items()
if v != label
}
continue
if label in objects.values(
) and track.track_id in objects.keys():
helper.drawBox(
frame,
(startX, startY, endX, endY),
label, (0, 255, 0))
else:
cv2.rectangle(frame, (xmin, ymin),
(xmax, ymax),
(0, 255, 0), 2)
helper.drawBox(
frame,
(startX, startY, endX, endY),
label, (0, 255, 0))
# add ids to objects
objects[track.track_id] = label
# set number of disappear for current track id to 0
disappearedIds[track.track_id] = 0
else:
# tracked already recognized people
for key, name in objects.items():
if key == track.track_id:
disappearedIds[track.track_id] = 0
helper.drawBox(frame,
(startX, startY, endX, endY),
name, (0, 255, 0))
# copy disappearedIds dict and increase the number of disappeared for a specific tracked object
idsCopy = disappearedIds.copy()
for i in idsCopy:
if i not in trackedOjects:
disappearedIds[i] += 1
#delete a given object if its number of dissapeared reached the maximum.
# 300 is the number of maximum consecutive frame a given tracked id (object) is allowed to be marked as disappeared
#until it is deleted from tracking.
if disappearedIds[i] == 300:
del disappearedIds[i]
if i in objects.keys():
del objects[i]
print(objects)
trackedOjects = []
font = cv2.FONT_HERSHEY_SIMPLEX
elips = (datetime.datetime.now() - startTime).total_seconds()
frameCounts += 1
fps = frameCounts / elips
text = 'Average FPS: {:.2f}'.format(fps)
cv2.putText(frame, text, (10, 20), font, 0.5, (0, 255, 0), 2)
# put frame in queue
if process is not None:
frameQueue.put(frame)
cv2.imshow('Face Recognition and Pedestrian Tracking', frame)
# stop writting process
if process is not None:
writeVideo.value = 0
process.join()
cap.release()
cv2.destroyAllWindows()
except Exception as e:
raise e
continue
detector_ip.push(Inference(image.copy()))
# print('waiting')
detector_op.pull_wait()
# print('done')
ret, inference = detector_op.pull()
if ret:
i_dets = inference.get_result()
trk_ip.push(Inference(i_dets))
# sleep(0.1)
t = Thread(target=read)
t.start()
video_writer = VideoWriter(out_path.get() + "/t_mobile_demo_out_4.avi",
image.shape[1], image.shape[0], 25)
while True:
# print(detector_op.is_closed())
trk_op.pull_wait()
if trk_ip.is_closed():
# print("Here")
video_writer.finish()
break
ret, inference = trk_op.pull()
if ret:
trackers = inference.get_result()
frame = inference.get_input().get_image()
for trk in trackers:
if not trk.is_confident():