Real-Time Soccer Analytics on the Edge : A Lightweight Framework and Data Standardization

초록

Soccer video analysis has advanced rapidly with recent progress in artificial intelligence and computer vision, enabling the automatic extraction of tactical and performance-related information from broadcast footage. Early systems, such as the classical pipeline of Liu et al. (2009), established the foundations of automatic player detection, labeling, and tracking using handcrafted visual features. More recent industry-oriented approaches demonstrated the practical feasibility of deep-learning-based soccer analytics, while modern generic multi-object tracking methods such as ByteTrack achieved strong benchmark performance. Despite these advances, differences in task formulation, evaluation protocols, and data representation hinder direct comparison and reveal the absence of a unified standard for soccer video analytics. Motivated by these observations, this study presents a modernized lightweight framework that performs automatic soccer player labeling and ball possession estimation and is specifically designed for real-time deployment on consumer-grade hardware. The framework integrates YOLOv8 to detect players and the ball, adaptive K-Means clustering to achieve unsupervised team-color labeling, and Norfair-based tracking enhanced with temporal inertia and a hold-rule mechanism to stabilize player identities and possession timelines. The ball detector is initialized from a COCO-pretrained YOLOv8 model and fine-tuned using approximately 1,000 manually annotated soccer-ball samples. Evaluation is conducted on four full-match broadcast videos, including a held-out match used exclusively for validation. Experimental results show that the proposed system achieves high detection performance (player precision 96.3%, recall 93.8%; ball precision 96.1%, recall 92.4%) while sustaining 25–30 FPS on a MacBook Air (M3), confirming its practicality for real-time soccer analytics. In addition, this work formally defines the Soccer Analytics Format for Spatial-Visual Annotation (SAF-SVA), a lightweight JSON-based schema that unifies player tracking, team labeling, and ball possession representation, enabling reproducible experimentation and fair benchmarking. Overall, the proposed framework provides a robust and scalable foundation for real-world soccer video analytics.

키워드