def full_run_sequential(video_id,
video_dir,
detector_config_path,
detector_model_path,
reid_model_path,
reid_model_backbone,
anomaly_results_dir,
bg_interval=4,
bg_alpha=0.05,
bg_start_frame=0,
bg_threshold=5,
raw_detect_interval=30,
ignore_count_thresh=0.08,
ignore_area_thresh=2000,
ignore_score_thresh=0.1,
ignore_gau_sigma=3,
abnormal_duration_thresh=60,
detect_duration_thresh=6,
undetect_duration_thresh=8,
bbox_score_thresh=0.3,
light_thresh=0.8,
anomaly_thresh=0.8,
similarity_thresh=0.95,
suspicious_duration_thresh=18,
detector_verbose_interval=20,
verbose=True,
crop_min_obj_size=8,
crop_row_capacity=3,
crop_box_aspect_ratio=2):
"""
Full run but runs one frame at a time. This should be more suitable for live processing.
Doesnt save any intermediate calculations.
See full_run_single for parameter info
"""
# Set up file paths
video_path = os.path.join(video_dir, f"{video_id}.mp4")
anomaly_results_path = os.path.join(anomaly_results_dir, f"{video_id}.csv")
# Create folders
os.makedirs(anomaly_results_dir, exist_ok=True)
# Read Video
raw_video = VideoReaderQueue(video_path, queue_size=8)
# bg modeling
print("Creating background...")
bg_images = calc_background(raw_video.load_video()[0], bg_interval,
bg_alpha, bg_start_frame, bg_threshold) # cpu
# bg_images = calc_bg_tensor(raw_video.load_video()[0], bg_interval, bg_alpha, bg_start_frame, bg_threshold) # gpu, doesnt seem to speed up much
bg_images = (img for img, _ in bg_images) # throw out frame
# Detection
detector = Detector(
detector_config_path,
detector_model_path,
detector_verbose_interval,
# Change this parameter depending what classes you need
# class_restrictions={5, 6, 13, 18}, # voc vehicles
class_restrictions={1, 2, 3, 5, 6, 7}, # coco vehicles
# class_restrictions={0}, # binary (0=vehicle, 1=not vehicle)
# class_restrictions=range(6), # only vehicles (0-5=vehicles, 6=not_vehicle)
# class_restrictions=None # use all classes
)
print(detector.model)
## Raw Video
print("Detecting raw video...")
raw_images, raw_frame_nums = raw_video.load_video(raw_detect_interval)
fbf_results_gen = detector.detect_images_generator(raw_images,
raw_frame_nums)
fbf_results_getter = detector.detect_images_getter(raw_video.get_frame)
fbf_results = ResultsDict(fbf_results_gen,
results_getter=fbf_results_getter,
name="fbf")
print("Detecting background...")
static_results_gen = detector.detect_images_generator(
bg_images, range(bg_start_frame, raw_video.nframes, bg_interval))
static_results = ResultsDict(static_results_gen, name="static")
# Comment out this whole block if you dont want to use crop boxes
print("Creating crop boxes")
# crop_box_gen = crop_box_generator(fbf_results, raw_video.img_shape)
# cropped_results_gen = detector.detect_images_generator(bg_images,
# range(bg_start_frame, raw_video.nframes, bg_interval),
# crop_box_gen)
# static_results.results_gen = cropped_results_gen
# Ignore Region
print("Creating ingore mask...")
# todo:
# I cant find a good way to create the ignore mask for live processing.
# It may need to not be used until a sufficient number of frames have been processed.
# It also isn't very quick at taking new regions into account. eg a car that stops on the grass wont get added for a while.
# ignore_alpha = 0.1
# ignore_alpha_2 = 1 - (1 - ignore_alpha) ** bg_interval # adjusted for different intervals
# ignore_mask_gen = create_ignore_mask_generator(static_results.iterator(), raw_video.img_shape, ignore_count_thresh,
# ignore_area_thresh, ignore_score_thresh, ignore_gau_sigma,
# alpha=ignore_alpha_2)
ignore_mask_gen = None # dont ignore anything
anomalies = get_anomalies_sequential(
raw_video, reid_model_path, fbf_results, static_results,
ignore_mask_gen, reid_model_backbone, bg_start_frame, bg_interval,
abnormal_duration_thresh, detect_duration_thresh,
undetect_duration_thresh, bbox_score_thresh, light_thresh,
anomaly_thresh, similarity_thresh, suspicious_duration_thresh, verbose)
if anomalies is not None:
anomaly_event_times = get_overlapping_time(anomalies)
# Save results
print("Saving Results...")
anomalies.to_csv(anomaly_results_path, index=False)
return anomalies, anomaly_event_times
else:
return [], []
