Ice Hockey Action Recognition via Contextual Priors

Abstract

Skeleton-based action recognition models, which are developed for generic human-pose data, struggle with ice-hockey broadcasts player action recognition, where the players appear smaller, move abruptly, and wield sticks that are invisible to standard skeleton models. To address these issues, we propose CP-Hockey, a context-aware pipeline that incorporates two domain-specific priors. First, a temporal player’s boundingbox normalization stabilizes player scale across the player tracklet, raising top-1 accuracy from 31 % to 57 % on a six-class NHL dataset. Second, we design hockey-specific skeletons that include stick end-points and optional detailed head landmarks. A 15-keypoint body-plus-stick model improves the accuracy to 64 %, while our full 20-keypoint configuration reaches 65 %. Experimental results with STGCN++ and 2s-AGCN show that both contextual priors are necessary: scale normalization reduces spatial jitter, and stick keypoints disambiguate visually similar movements such as stickwork versus striking a puck with a stick. CP-Hockey establishes a strong baseline for fine-grained ice-hockey analytics and provides a blueprint for adapting skeleton pipelines to other equipment-centric sports.