Ultimate Frisbee Field Boundary Computer Vision Project

Ultimate Frisbee Drone

Use consistent source video for ultimate frisbee games to create computer vision models

Goals

1. Use endzone cones to try field boundary detection, homography
2. Light / Dark player object detection
3. Frisbee object detection

Steps

• download source video footage (See image names for yt IDs)
• dataset 1
  • capture frames every 60 seconds
  • label visible endzone cones
  • label info boxes when present (score updates, injuries, timeouts, etc . . .)
• dataset 2
  • add additional frames to dataset 1, every 15 seconds of video
  • to be annotated

Cone Label Schematic

Frame Capture Script

# paths can be strings or Path objects, yt_id should be string
# update progress bar with each image saved or each frame processed

def get_frames(video_path, image_path, yt_id, interval=5):
    if not video_path.exists():
        print('video not found')
        return
    if not image_path.exists():
        image_path.mkdir()

    # open video, get frames per second
    cap = cv2.VideoCapture(str(video_path))
    fps = int(cap.get(cv2.CAP_PROP_FPS))
    counter = 0

    # pbar = tqdm(total=int(cap.get(cv2.CAP_PROP_FRAME_COUNT)/(fps*interval)),
    #             position=0, leave=True)
    pbar = tqdm(total=int(cap.get(cv2.CAP_PROP_FRAME_COUNT)),
                position=0, leave=True)
    while cap.isOpened():
        success = cap.grab()
        pbar.update(1)
        if success:
            counter += 1
            if counter % (fps * interval) == 0:
                save_path = Path(image_path, f'{yt_id}_{counter}.png')
                # pbar.update(1)
                if not save_path.exists():
                    _, image = cap.retrieve()
                    cv2.imwrite(str(save_path), image)
        else:
            break
    
    pbar.close()
    cap.release()
    cv2.destroyAllWindows()
    return