YOLO-Based Framework with Temporal Context and Network Analysis for Real-Time Basketball Video Understanding

Abstract

This paper proposes a real-time system for the automated recognition of player interactions in basketball videos using a YOLO-based deep learning framework integrated with temporal context and network analysis. The framework enhances YOLOv3 by incorporating information from adjacent frames to improve jersey number recognition and player tracking under occlusion. A novel algorithm, Joy2019, is introduced to assign consistent player identities over time using retrospective association and temporal smoothing. Passing interactions are represented as directed graphs, and a new Player Centrality (PC) score is developed to evaluate player influence using weighted actions. Experimental evaluation on annotated NBA footage demonstrates that Joy2019 improves jersey number recognition accuracy from 36.1% to 73.8%, and increases player detection rates to 90.2%. Compared to baseline YOLO outputs, the proposed method reduces mean absolute percentage error (MAPE) in pass graph reconstruction to 24.3– 33.1%. These findings demonstrate the potential of temporal deep learning and network science to improve the robustness, accuracy, and tactical insight of automated sports analytics systems.