陳容正
In recent years, information technology has become an indispensable part of var ious sports events. By capturing on-field information through cameras and analyzing the footage using computer vision and artificial intelligence techniques, richer sports data can be provided. Player trajectories, in particular, can offer valuable tac tical insights, such as player interactions, team chemistry, and commonly employed strategies. However, existing multi-player tracking methods often face challenges like severe player occlusion, pose deformation, and rapid camera movement, leading to suboptimal tracking results. To address these issues, this paper explores the design of feature galleries for re-identification in basketball player tracking and pro poses a multi-player tracking and matching framework tailored to basketball broad cast videos. This framework improves tracking by designing a feature gallery classification method that considers the players’ activity range and the number of players on the court. First, the system uses a positional-based tracking model to obtain the initial trajectories, followed by analyzing these trajectories to identify the moments when trajectories disappear and reappear, and further performs trajectory matching. Based on player positions and occlusion scenarios, the cases requiring further matching are classified into three categories, and a re-identification mechanism is used to rematch the missing and reappearing trajectories in each category. Additionally, we designed an interface that allows users to manually check and adjust the tracking results to ensure the accuracy of the trajectories and the feasibility of subsequent trajectory applications.