Real-Time Soccer Analytics: Event Detection, Player Tracking, and Spatial Mapping with Computer Vision

This paper presents a real-time soccer analytics system that leverages computer vision and AI to automate player tracking, spatial mapping, and event detection from a single broadcast video feed. The system integrates a fine-tuned YOLOv8s entity detector (90.7% mAP50) with two auxiliary YOLO26n-pose models for pitch (32 keypoints) and goalpost (8 keypoints per side) localization. A multi-threaded, asynchronous pipeline combines Slicing Aided Hyper Inference (SAHI) for small-object recall, ByteTrack for multi-object tracking, a Constant-Acceleration Kalman Filter for ball trajectory smoothing, and a robust HSV+LAB team assigner with temporal voting. Strict homography validation maps 2D broadcast views to standardized world coordinates, enabling precise spatial event logic (goal-line crossings, out-of-play boundaries). TensorRT FP16 optimization, GPU-native NMS, and CUDA stream execution recover inference overhead, delivering a stable ~20.5 FPS end-to-end on consumer-grade hardware. The pipeline achieves 89.3% direct visual ball detection (compared to ~3–8% in standard baselines), 100% valid homography frames, and reliable team possession tracking. Structured metadata and automatic event clip extraction are exported alongside live analytics, providing a scalable, low-latency soccer analysis solution without human intervention. The system is explicitly designed for semi-professional scenarios using a single scouting camera, clear field markings, and distinct team kits, operating efficiently on limited hardware (Intel i5-12500H, 24GB RAM, RTX 3050 Mobile 4GB).