Deep learning-based player tracking in sports videos

Abstract

Multi-camera player tracking is a fundamental prerequisite for advanced sports analytics, yet it remains a computationally challenging task due to frequent inter-player occlusions, rapid motion, and the visual homogeneity of team uniforms. This thesis presents a robust end-to-end pipeline for the detection and tracking of football players using a calibrated four-camera setup. The proposed system integrates state-of-the-art deep learning techniques with geometric computer vision. We employ a fine-tuned object detector paired with ByteTrack for local perception. To resolve the Multi-Dimensional Assignment (MDA) problem across views, we introduce a Hierarchical Divide- and-Conquer fusion strategy. Unlike naive greedy clustering approaches, this method utilises recursive bipartite matching with a multi-cue cost function incorporating position, velocity, shape, and colour histograms. Furthermore, a Temporal Hinting mechanism is implemented to recover player identities following extended occlusions or spatial discontinuities. Comparative evaluation against a greedy geometric baseline demonstrates substantial improvements in tracking accuracy, with the hierarchical approach achieving 0.844 GS-HOTA compared to 0.416 for the baseline-a 103% relative improvement. Comprehensive evaluation on held-out test sequences across temporal horizons from 5 to 45 minutes confirms exceptional detection stability, with detection accuracy (DetA) maintaining 93.6% and MOTA sustaining 97.0% regardless of sequence length. The system exhibits a consistent identity switch rate of approximately 110 switches per minute, demonstrating temporal stability without compounding drift. These results establish a strong foundation for automated game state reconstruction and tactical analysis in professional sports.