HistoTrack++: A Vision-Based System for Temporal Bout Segmentation, Multi-Target Tracking and Kinematic Analysis in Overhead Combat Sports Videos

Analyzing raw, untrimmed training videos in combat sports is challenging due to occlusion, motion clutter, and identity ambiguity. This paper presents HistoTrack++, an end-to-end framework for overhead boxing footage that addresses these challenges via automated bout segmentation, robust identity tracking, and interpretable movement analytics. The pipeline begins with a segmentation module combining a fine-tuned YOLOv8 detector with domain-driven spatial-temporal heuristics, achieving 97.6% segmentation accuracy on 40 h of sparring data from 75 athletes. For tracking, we introduce HistoTrack, a histogram-guided, identity-preserving tracker that maintains 0.95 MOTA at 23 FPS, even under heavy occlusion. Four coach-informed kinematic metrics—hotspot density, directional histograms, engage/disengage dynamics, and zone usage—provide actionable insights aligned with athlete development. Extensive evaluation against 21 existing models demonstrates that HistoTrack + + outperforms prior methods in accuracy, identity consistency, and interpretability in challenging overhead views. Although validated on boxing, the modular segmentation–tracking–analytics framework is domain-agnostic and transferable to other multi-object overhead video scenarios, including judo, wrestling, and team-sport training. With ~ 10 M parameters and real-time performance, the framework is deployable both locally and on scalable server infrastructures. HistoTrack + + thus offers a practical, scalable approach for automated video analytics in combat sports, with broader implications for tactical modeling, longitudinal performance assessment, and real-time sports video understanding.