Enhancing Soccer Player Tracking and Re-Identification with Dual Visual Embeddings and Tracklet Association

We present a modular player-tracking pipeline for fixed-viewpoint soccer videos that combines a transformer-based detector (RF-DETR), BoT-SORT for local association, Global Tracklet Association (GTA) for global reconnection, temporal interpolation, and appearance embeddings from CLIP-ReID and PRT-ReID. On the SoccerTrack Challenge ''challenge'' split, the system attains HOTA 0.55, LocA 0.82, and AssA 0.43. We hypothesized that fusing multiple appearance embeddings would consistently improve accuracy. Ablations on two in-domain training sequences show that CLIP-Only performs best, while Proposed (CLIP+PRT) improves over PRT-Only yet does not surpass CLIP-Only. GTA improves ID consistency when paired with interpolation, and reducing detector input resolution from 1456 to 560 substantially degrades accuracy due to weakened small-object detection. These findings indicate that CLIP embeddings are highly discriminative, calibrated fusion is required for PRT to help, and high resolution is a decisive factor. Future work will focus on calibrated fusion and lightweight detection for real-time use.