A Lightweight Tracking-Based Motion Stream for Hockey Skating Action Recognition

Abstract

Fine-grained action recognition of individual players is a fundamental problem for automatic game understanding, yet existing approaches in team sports primarily rely on local visual appearance or human pose, largely ignoring how players move within the playing environment. We introduce a lightweight tracking-based motion stream that captures global player kinematics using only four features per frame: rink-coordinate displacement and body orientation. Although remarkably simple, the proposed motion stream achieves 79.34% action recognition accuracy on the challenging FineHockeySkate [20] benchmark, within 1% of a skeleton-based model and 4% of a much larger video-based model. More importantly, the motion stream provides complementary information unavailable to local appearance and pose representations. We therefore propose a late-stage fusion framework that combines the motion stream with independently trained skeleton-based and video-based models. Experiments demonstrate that the proposed motion stream complements both modalities, with all pairwise fusion combinations outperforming their strongest individual component. Selecting the best model by mean class accuracy (MCA), our video+motion fusion achieves a new state of the art on FineHockeySkate with 65.84% MCA and 85.16% overall accuracy.