Puck localization and multi-task event recognition in broadcast hockey videos

Puck localization is an important problem in ice hockey video analytics\nuseful for analyzing the game, determining play location, and assessing puck\npossession. The problem is challenging due to the small size of the puck,\nexcessive motion blur due to high puck velocity and occlusions due to players\nand boards. In this paper, we introduce and implement a network for puck\nlocalization in broadcast hockey video. The network leverages expert NHL\nplay-by-play annotations and uses temporal context to locate the puck. Player\nlocations are incorporated into the network through an attention mechanism by\nencoding player positions with a Gaussian-based spatial heatmap drawn at player\npositions. Since event occurrence on the rink and puck location are related, we\nalso perform event recognition by augmenting the puck localization network with\nan event recognition head and training the network through multi-task learning.\nExperimental results demonstrate that the network is able to localize the puck\nwith an AUC of $73.1 \\%$ on the test set. The puck location can be inferred in\n720p broadcast videos at $5$ frames per second. It is also demonstrated that\nmulti-task learning with puck location improves event recognition accuracy.\n