Player Re-Identification in Basketball

Sammanfattning

Player Re-Identification (Re-ID) is a fundamental component for comprehensive team sports analysis. However, applying traditional 2D appearance-based trackers presents unique challenges, as confined courts, complex tactics, and highly similar team uniforms cause severe mutual occlusions and frequent identity switches. Furthermore, erratic player movements and image degradation severely compromise visual identification, making long-term Re-ID exceptionally difficult. To address these domain-specific challenges, this thesis proposes a robust, multi-modal Re-ID framework. Rather than relying solely on global appearance features, our system extracts three orthogonal modalities: omni-scale visual appearance, explicit semantic jersey numbers via a domain-adapted Scene Text Recognition pipeline (YOLO11 and PARSeq), and fine-grained shoe color attributes using zero-shot segmentation (SAM 3). To mitigate single-frame anomalies caused by dynamic sports movements, we implement a temporal feature aggregation module utilizing average pooling for continuous visual features and cumulative majority voting for discrete semantic identifiers. At the tracking level, these modalities are dynamically fused using an XGBoost-based classification engine to effectively handle varying feature reliability. Additionally, a Dynamic Global ID mechanism employing an Exponential Moving Average (EMA) update strategy smoothly adapts to continuous visual drift while preserving historical robustness. Extensive evaluations on the TrackID3x3 dataset demonstrate that our multi-modal approach significantly outperforms traditional baselines, establishing a highly resilient paradigm for continuous player tracking and complex sports analytics.