Existing hockey analytics do not incorporate important context when analyzing shots, such as defender positioning and traffic in front of the net. We present a computer vision pipeline for detecting shots in hockey broadcasts, classifying players and their teams, and converting this data into 2D coordinates with an additional a rink projection. Armed with this additional context, we hope to influence richer analytics such as Defender-Adjusted Expected Goals (DAxG).
Example of full pipeline - clip with role/team classification, clip with anchor detection (faceoff/goal), and output 2D projection image from shot detection frame:
clip_309_tracked_github.mp4 |
clip_309_rink_anchor_tracked_github.mp4 |
