Ice Hockey Puck Detection and Tracking from Broadcast Video

Abstract

Accurate localization and tracking of fast-moving sports objects in broadcast video remain a fundamental problem in computer vision, enabling downstream applications such as tactical analysis, player analytics, automated highlight generation, officiating assistance, and augmented reality broadcasting. In ice hockey, this problem is particularly difficult because the puck occupies only a few pixels, rapidly accelerates, exhibits motion blur, becomes occluded by players and boards, and is observed under continuously changing camera viewpoints. These characteristics substantially degrade the performance of conventional appearance-based detection and tracking methods. This thesis presents an end-to-end framework for puck detection and tracking that combines contextual reasoning, geometric scene understanding, and temporal motion modelling. First, I propose a context-aware puck detector, termed Player Localization Using Contextual Cues (PLUCC). By using player segmentation as an explicit contextual prior, the proposed method exploits the strong spatial relationship between players and the puck to improve localization accuracy over appearance-only detection approaches. The thesis then assesses the suitability of modern visual object trackers for broadcast hockey footage through a comprehensive evaluation of modern tracking algorithms. Experimental results show that existing appearance-based trackers remain fundamentally limited by the puck’s small image footprint, rapid motion, frequent occlusions, and visually similar distractors, motivating an alternative tracking paradigm. To address these limitations, we introduce a geometry-aware tracking framework in which puck motion is represented in a temporally stable planar coordinate system rather than directly in image space. A monocular 3D reconstruction pipeline estimates the broadcast camera pose and reconstructs the rink surface, while semantic segmentation removes dynamic objects to enable robust plane estimation. Projecting detections into this canonical coordinate system substantially reduces perspective distortion and camera-induced motion, producing geometrically consistent trajectories throughout time. Temporal motion modelling is subsequently performed within the planar representation using sequence-based learning, allowing the model to focus on the puck’s underlying physical motion rather than compensating for camera motion. Collectively, the proposed framework demonstrates that combining contextual detection with geometry-aware motion modelling provides a more robust foundation for puck tracking than conventional appearance-based approaches alone. Although developed for broadcast ice hockey, the suggested approach is largely sport-agnostic and is readily applicable to other sports involving small objects moving on approximately planar playing surfaces, including soccer, tennis, basketball, lacrosse, and baseball.