A Real-Time Basketball Player Tracking Method Based on Occlusion-Aware Re-identification with a Dynamic Feature Gallery

In modern basketball, tactical strategy analysis is vital for improving team per formance. However, challenges arise due to player occlusion and uniformity of appearances due to similar jerseys, which complicate player tracking in broadcast footage. Accurate tracking is essential for simplifying analysis and optimizing the control PTZ camera orientation, optimizing the capture of game footage. Yet, existing algorithms struggle with the high occlusion rates and similar appearances of players, making tracking difficult. Therefore, this paper proposes a real-time player tracking algorithm that achieves high accuracy. This paper proposes a real-time single-player tracking algorithm that combines the advantages of position-based and feature-based approaches, introducing a Rule based Occlusion Detection mechanism and a Dynamic Feature Gallery (DFG) to address the high occlusion scenarios among players. Our design ensures that appear ance features are considered for re-identification only during occlusion, minimizing misidentification risks and reducing computational costs. Additionally, appearance features are conditionally sampled to reduce the computational cost of frame-by- frame calculations. Furthermore, a Siamese Network-based player feature extractor is employed for stable re-identification.