Detecção e rastreamento da bola em partidas de futebol a partir de fluxos MPEG-TS: um estudo de viabilidade com visão computacional

Abstract

This work presents a complete pipeline for detecting and tracking a soccer ball in broadcast footage captured directly from MPEG-TS using the DTU-238 receiver and StreamXpert. After extracting the video with FFmpeg, a YOLOv8n model is used for frame-by-frame detection, followed by specialized filters to reduce false positives. An optical flow module estimates camera motion to stabilize the scene, enabling continuous tracking even when detections are missing. The estimated trajectory is compared and visually overlaid with manually annotated ground truth, producing a final annotated video. Results indicate that reliable ball-tracking can be achieved from broadcast TV streams, demonstrating the potential of this approach for sports analysis applications.