Computer Vision-Driven Ultimate Frisbee Tracking and Analytics

This paper presents work on the detection, tracking, and localization of the frisbee and players in footage of the sport Ultimate Frisbee. In the frisbee detection procedure, features are extracted from video frames using color, shape, and size cues obtained from digital image processing techniques. However, the problem of finding the frisbee position in real game footage is challenging because the appearance of the frisbee varies in size, shape, and brightness. In addition, many other moving objects may also appear similar to the frisbee, such as specific bright regions on players’ hats, jersey numbers, cleats, etc. This project aims to find the best method to localize both the players and the frisbee, with the ultimate goal of automatic pass counting. Methods tested to track the frisbee include: mophological-based tracking, background subtraction-based tracking, and template matching. While morphological-based processing was found to be most effective in tightly controlled footage, it was largely ineffective in full-scale game footage. Color-based segmentation and region labeling was used to successfully track player locations in both test footage and real game footage.