Fusion-Net50-Based Soccer Player Re-Identification and Team Clustering for Speed and Movement Analysis

Abstrak

Comprehensive analysis of soccer matches requires substantial human resources as well as specialized equipment with relatively high costs. The development of Computer Vision (CV) technology, particularly the Multi-Object Tracking (MOT) approach, enables contactless and efficient analysis of player performance. However, the main challenges in MOT for soccer, namely player re-identification, tracking ID switches, as well as inter-frame identity consistency, especially under occlusion conditions and low-confidence detections, can cause the analysis data to become invalid. This research aims to improve object identity consistency (Tracking ID) in video-based soccer match analysis by integrating detection, tracking, re-identification, and team identification methods. This research proposes the integration of YOLOv8n- and ByteTrack-based tracking methods with a Re-ID embedding mechanism, low-confidence detection handling, as well as ResNet-50-based Feature–Position Fusion to maintain identity consistency (Track ID) throughout the video sequence. Furthermore, team identification is consistently performed using a Color Embedding approach with a ResNet-50 backbone and the K-Means clustering algorithm. The experimental results show that the integration of all these methods successfully maintains consistent Track ID with a total of 25 Track IDs (Players 21 Track IDs, Goalkeeper 1 Track ID, Referee 2 Track IDs, Ball 1 Track ID). This performance is supported by evaluation metrics indicating strong performance, namely HOTA of 91.9%, MOTA of 89.8%, and MOTP of 81.0%. The effectiveness of the Re-ID module in maintaining object identity consistency is reflected in the IDF1 score of 94.9%. Supported by association metrics (AssA 93.5%, AssRe 94.8%, AssPr 96.3%), these results indicate that trajectories are formed with a relatively low level of association errors, thereby supporting subsequent analyses such as speed estimation and player movement visualization in a more stable manner.