Dynamic sports environments: automatic feature players tracking and re-identification with similar appearances

Abstract

Tracking multiple players in sports videos remains challenging due to frequent occlusions, similar appearances among players, and rapid motion. This paper presents a comprehensive player tracking framework that integrates deep player re-identification (DPRI), an individual probabilistic overlapping bounding box occupancy mapping (IPOM) for robust spatial localization under severe overlap, and an identity-aware k-shortest path optimization algorithm (KSP-ID) for improved trajectory association. The DPRI module combines pose-based embeddings, jersey appearance cues, and facial features to generate discriminative identity representations. Experimental evaluation on the autonomous production of images based on distributed and intelligent sensing (APIDIS) and Shantou University (STU) datasets demonstrates strong performance, achieving 85.5% multiple object tracking accuracy (MOTA), 69.0% generalized multiple object tracking accuracy (GMOTA), and 90.2% rank-1 re-identification accuracy. Furthermore, the proposed framework significantly reduces identity switches (IDS) compared with conventional k-shortest path (KSP) approaches, probabilistic occupancy models, and recent deep learning-based tracking methods. These results demonstrate that integrating semantic identity information with spatial reasoning enables robust and efficient player tracking in highly dynamic sports environments.