def detect_image_folder(image_folder, output_path, model_config_path,
model_checkpoint_path):
"""
Runs object detection on a folder of images. Saves the results to a csv
"""
img_reader = ImageReader(image_folder)
model = Detector(model_config_path, model_checkpoint_path)
results = model.detect_images(img_reader.load_images(),
img_reader.filenames)
results.to_csv(output_path, index=None)
def full_run_single(video_id,
video_dir,
static_dir,
frame_by_frame_results_dir,
static_results_dir,
crop_boxes_dir,
ignore_mask_dir,
detector_config_path,
detector_model_path,
reid_model_path,
reid_model_backbone,
crop_results_dir,
anomaly_results_dir,
bg_interval=4,
bg_alpha=0.05,
bg_start_frame=1,
bg_threshold=5,
raw_detect_interval=30,
crop_min_obj_size=8,
crop_row_capacity=3,
crop_box_aspect_ratio=2,
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):
"""
Runs the full anomaly detection pipeline on a video
video_id: video id/name
video_dir: folder the video is in
static_dir: folder to put the background images in
frame_by_frame_results_dir: folder to put the raw video detection results in
static_results_dir: folder to put the background image detection results in
crop_boxes_dir: folder to put the crop boxes in
ignore_mask_dir: folder to put the ignore region mask in
detector_config_path: path to detector configuration file
detector_model_path: path to detector model checkpoint
reid_model_path: path to re-ID model checkpoint
reid_model_backbone: re-ID model backbone. eg. "resnet50"
bg_interval, bg_alpha, bg_start_frame, bg_threshold: see calc_bg_full_video function
raw_detect_interval: number of frames between detection on raw video
crop_min_obj_size, crop_row_capacity, crop_box_aspect_ratio: see create_crop_boxes function
ignore_count_thresh, ignore_area_thresh, ignore_score_thresh, ignore_gau_sigma: see create_ignore_mask function
abnormal_duration_thresh, detect_duration_thresh, undetect_duration_thresh, bbox_score_thresh,
light_thresh, anomaly_thresh, similarity_thresh, suspicious_duration_thresh:
See get_anomalies function
detector_verbose_interval: detector progress printing interval
verbose: verbose printing
"""
# Set up file paths
video_path = os.path.join(video_dir, f"{video_id}.mp4")
static_images_folder = os.path.join(static_dir, f"{video_id}")
fbf_results_path = os.path.join(frame_by_frame_results_dir,
f"{video_id}.csv")
static_results_path = os.path.join(static_results_dir, f"{video_id}.csv")
crop_boxes_path = os.path.join(crop_boxes_dir, f"{video_id}.csv")
crop_results_path = os.path.join(crop_results_dir, f"{video_id}.csv")
ignore_mask_path = os.path.join(ignore_mask_dir, f"{video_id}.npy")
anomaly_results_path = os.path.join(anomaly_results_dir, f"{video_id}.csv")
# Create folders
os.makedirs(static_images_folder, exist_ok=True)
os.makedirs(frame_by_frame_results_dir, exist_ok=True)
os.makedirs(static_results_dir, exist_ok=True)
os.makedirs(crop_boxes_dir, exist_ok=True)
os.makedirs(crop_results_dir, exist_ok=True)
os.makedirs(ignore_mask_dir, exist_ok=True)
os.makedirs(anomaly_results_dir, exist_ok=True)
# Read Video
raw_video = VideoReader(video_path)
# bg modeling
print("Creating background...")
calc_bg_full_video(video_path, static_images_folder, bg_interval, bg_alpha,
bg_start_frame, bg_threshold, verbose)
# Detection
detector = Detector(detector_config_path,
detector_model_path,
detector_verbose_interval,
class_restrictions=None)
# class_names = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car',
# 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
# 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',
# 'tvmonitor')
# detector.model.CLASSES = class_names
# detector.class_labels = class_names
## Raw Video
print("Detecting raw video...")
raw_images, raw_frame_nums = raw_video.load_video(raw_detect_interval)
fbf_results = detector.detect_images(raw_images, raw_frame_nums)
fbf_results.to_csv(fbf_results_path, index=False)
## Static Images
static_reader = ImageReader(static_images_folder)
static_frame_names = list(
map(lambda f: int(f[:-4]),
static_reader.filenames)) # "123.jpg" -> 123
print("Detecting background...")
static_results = detector.detect_images(static_reader.load_images(),
static_frame_names)
static_results.to_csv(static_results_path, index=False)
# Perspective Cropping
print("Creating crop boxes...")
create_crop_boxes(
fbf_results_path, crop_boxes_path, raw_video.img_shape,
crop_min_obj_size, crop_row_capacity,
crop_box_aspect_ratio) # either static/fbf results should work
# Should be able to use this in place of normal static images. Doesnt look feasable atm, way too long detection time
crop_boxes = pd.read_csv(crop_boxes_path).values
print("Detecting cropped background...")
crop_detect_results = detector.detect_images(static_reader.load_images(),
static_frame_names,
crop_boxes=crop_boxes)
crop_detect_results.to_csv(crop_results_path)
# # Ignore Region
print("Creating ingore mask...")
create_ignore_mask(fbf_results_path, ignore_mask_path, raw_video.img_shape,
ignore_count_thresh, ignore_area_thresh,
ignore_score_thresh, ignore_gau_sigma)
# Detect anomalies
print("Detecting anomalies...")
anomalies = get_anomalies_preprocessed(
video_path, reid_model_path, fbf_results_path, static_results_path,
ignore_mask_path, 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 [], []