Summarizing Footage from Sporting Events with Computer Vision

Scott Powell, Christian Newton, and Dr. Jason Grant, Computer Science Department, Middlebury College, 14 Old Chapel Road, Middlebury VT 05753

Computer vision is an important topic of research in computer science with applications ranging from self-driving cars to facial recognition in sports analysis. Object detection algorithms are a centerpiece of computer vision, and serve as a basis for much of the field. We used object detection algorithms to analyze basketball footage, and create a cut of “highlights” of a basketball game. We also used facial recognition to distinguish players, allowing us to create highlights specific to an individual player. 

We found that the YOLOv5 object detection algorithm performed better than other algorithms. While object detection worked well on large and static objects such as the people or backboards, it was unable to consistently detect or track a basketball in motion due to blurring in video footage with reduced framerates. At 30 FPS we could track the ball accurately while it was held, but not when it was passed or thrown. We were able to increase tracking success by using information about detections in the surrounding 100 frames to approximate the basketball’s location. Facial recognition was effective for people close to the camera, but at increased distances became an ineffective method of distinguishing players. We found recognition of players by jersey numbers worked better than face recognition and could be implemented as a supplement or a replacement of facial recognition in this context. Additionally, we found camera quality had a significant impact on the speed and efficiency of the algorithm, as high-definition cameras create larger images that take more time to process, but low-definition cameras might cause detection algorithms to fail if the object is blurry. This is important for other computer vision applications such as self-driving cars, which need to find a balance between accurately detecting objects and detecting them in real-time.

Additional Abstract Information

Presenter: Scott Powell

Institution: Middlebury College

Type: Poster

Subject: Computer Science

Status: Approved

Time and Location

Session: Poster 5
Date/Time: Tue 12:30pm-1:30pm
Session Number: 4059