def get_video_reader(args, video_name):
date, hr_slot, camera = parse_meva_clip_name(video_name)
video_path = os.path.join(args.video_path, date, hr_slot,
video_name + ".avi")
if args.use_lijun_video_loader:
vcap = VideoReader(video_path)
frame_count = int(vcap.length)
elif args.use_moviepy:
vcap = VideoFileClip(video_path, audio=False)
frame_count = int(vcap.fps * vcap.duration) # uh
vcap = vcap.iter_frames()
else:
try:
vcap = cv2.VideoCapture(video_path)
if not vcap.isOpened():
raise Exception("cannot open %s" % video_path)
except Exception as e:
raise Exception("warning, cannot open %s" % video_path)
# opencv 3/4
frame_count = vcap.get(cv2.CAP_PROP_FRAME_COUNT)
# start reading frames into queues now
video_queuer = VideoEnqueuer(
args,
vcap,
frame_count,
frame_gap=1, # no skipping frames
prefetch=100,
start=True,
is_moviepy=args.use_moviepy,
batch_size=1)
get_frame_batches = video_queuer.get()
return get_frame_batches, video_queuer.num_batches
frame_count = int(vcap.fps * vcap.duration) # uh
vcap = vcap.iter_frames()
else:
try:
vcap = cv2.VideoCapture(videofile)
if not vcap.isOpened():
raise Exception("cannot open %s" % videofile)
except Exception as e:
# raise e
# just move on to the next video
print("warning, cannot open %s" % videofile)
continue
# opencv 2
if cv2.__version__.split(".")[0] == "2":
frame_count = vcap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT)
else:
# opencv 3/4
frame_count = vcap.get(cv2.CAP_PROP_FRAME_COUNT)
# initialize tracking module
if args.get_tracking:
tracking_objs = args.tracking_objs.split(",")
tracker_dict = {}
tracking_results_dict = {}
tmp_tracking_results_dict = {}
for tracking_obj in tracking_objs:
metric = metric = nn_matching.NearestNeighborDistanceMetric(
"cosine", args.max_cosine_distance, args.nn_budget)
tracker_dict[tracking_obj] = Tracker(
metric, max_iou_distance=args.max_iou_distance)
def load_track_and_features(args, video_name, p_file, v_file, p_extractor,
v_extractor, hs):
date, hr_slot, camera = parse_meva_clip_name(video_name)
# start loading video_frames first
video_path = os.path.join(args.video_path, date, hr_slot, video_name + ".avi")
if args.use_lijun_video_loader:
vcap = VideoReader(video_path)
frame_count = int(vcap.length)
elif args.use_moviepy:
vcap = VideoFileClip(video_path, audio=False)
frame_count = int(vcap.fps * vcap.duration) # uh
vcap = vcap.iter_frames()
else:
try:
vcap = cv2.VideoCapture(video_path)
if not vcap.isOpened():
raise Exception("cannot open %s" % video_path)
except Exception as e:
raise Exception("warning, cannot open %s" % video_path)
# opencv 3/4
frame_count = vcap.get(cv2.CAP_PROP_FRAME_COUNT)
# start reading frames into queues now
video_queuer = VideoEnqueuer(
args, vcap, frame_count, frame_gap=1, # no skipping frames
prefetch=100,
start=True, is_moviepy=args.use_moviepy,
batch_size=1)
get_frame_batches = video_queuer.get()
def load_track_file(file_path, homography):
"""load a tracking file into dict of numpy arrays."""
# assuming sorted by frameid
data = []
with open(file_path, "r") as f:
for line in f:
frame_idx, track_id, left, top, width, height, conf, _, _, _ = line.strip().split(",")
data.append([frame_idx, track_id, left, top, width, height, conf])
if not data:
return {}
data = np.array(data, dtype="float32") # [N, 7]
# compute topdown points
foot_points_x = data[:, 2] + data[:, 4] / 2. # [N]
foot_points_y = data[:, 3] + data[:, 5]
foot_points = np.stack([foot_points_x, foot_points_y], axis=0) # [2, N]
# [2, N]
top_down_points = warp_points(foot_points, homography)
top_down_points = np.transpose(top_down_points, [1, 0]) # [N, 2]
# [N, 9]
data = np.concatenate([data, top_down_points], axis=1)
track_ids = np.unique(data[:, 1]).tolist()
track_data = {} # [num_track, K, 9]
for track_id in track_ids:
track_data[track_id] = data[data[:, 1] == track_id, :]
return track_data
# track_id -> data
p_tracks = load_track_file(p_file, hs[camera])
v_tracks = load_track_file(v_file, hs[camera])
# get each frame's boxes to extract
frame_data = {} # frame_idx -> a list of boxes,
def get_track_boxes(tracks, cat_name):
for track_id in tracks:
idxs = list(range(0, len(tracks[track_id]), args.feature_box_gap))
idxs = idxs[:args.feature_box_num]
boxes = tracks[track_id][idxs, :] # [k, 7]
for box_idx, box in enumerate(boxes):
frame_idx = box[0]
tlwh = box[2:6]
if not frame_idx in frame_data:
frame_data[frame_idx] = []
frame_data[frame_idx].append((tlwh, track_id, box_idx, cat_name))
get_track_boxes(p_tracks, "Person")
get_track_boxes(v_tracks, "Vehicle")
# 2. go through the video once and crop all the images to extract features
# assuming not conflict between person/vehicle track_id
p_track_to_feat = {} # "track_id" => features
v_track_to_feat = {} # "track_id" => features
for batch in tqdm(get_frame_batches, total=video_queuer.num_batches):
image, scale, frame_idx = batch[0]
image = image.astype("uint8") # need uint8 type
if frame_idx in frame_data:
for tlwh, track_id, box_idx, cat_name in frame_data[frame_idx]:
# check box valid
if valid_box(tlwh, image):
x, y, w, h = tlwh
x, y, w, h = int(x), int(y), int(w), int(h)
#print(x, y, w, h)
#print(image[y:y+h, x:x+w])
box_img = cv2.cvtColor(
image[y:y+h, x:x+w], cv2.COLOR_BGR2RGB)
if cat_name == "Person":
if track_id not in p_track_to_feat:
p_track_to_feat[track_id] = []
p_track_to_feat[track_id].append(box_img)
elif cat_name == "Vehicle":
if track_id not in v_track_to_feat:
v_track_to_feat[track_id] = []
v_track_to_feat[track_id].append(box_img)
# extract features
def get_features(track_to_imgs, extractor):
for track_id in track_to_imgs:
box_imgs = track_to_imgs[track_id]
track_to_imgs[track_id] = extractor(box_imgs).cpu().numpy() # [K, 512]
if args.use_avg_pool:
# [1, 512]
track_to_imgs[track_id] = np.mean(
track_to_imgs[track_id], axis=0, keepdims=True)
get_features(p_track_to_feat, p_extractor)
get_features(v_track_to_feat, v_extractor)
data = {}
def gather_data(track_data, track_features, cat_name):
data[cat_name] = {}
for track_id in track_data:
# ignore track with no valid boxes
if track_id in track_features:
data[cat_name][track_id] = (
track_data[track_id], track_features[track_id])
gather_data(p_tracks, p_track_to_feat, "Person")
gather_data(v_tracks, v_track_to_feat, "Vehicle")
return data