ANTHROPOMETRIC BODY ZONE-BASED SOCCER PASS DETECTION AND TEAM CLASSIFICATION WITH INTEGRATED COMPUTER VISION MODELS

Read full text Summarise with VIRAT
Authors: Namjildorj Batbaatar, Stephen Karungaru, Kenji Terada, Munkhbat Gantumur, Altangerel Ayush
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Real-time detection and classification of player actions in soccer matches remains challenging due to complex player interactions, frequent occlusions, and the need for accurate team-specific possession analysis. This paper introduces a novel soccer event detection system that achieves 84.51% accuracy at 18.33 FPS through the integration of anthropometric body zone analysis with multiple computer vision models. Our approach combines YOLOv8 for multi-object detection, ByteTrack for tracking, and an innovative unsupervised team classification method using SigLIP embeddings with UMAP-accelerated dimensionality reduction and K-means clustering. To resolve ambiguities in ball-player association, we develop a principled anthropometric approach that segments player bounding boxes into upper and foot zones based on human biomechanical ratios, supplemented by ResNet-based CNN classifiers for complex occlusion scenarios. The system detects comprehensive soccer events including standard passes (84.36% accuracy), headed passes (86.05% accuracy), throw-ins (87.5% accuracy), and chest control interactions – capabilities not addressed in previous related works. Experimental validation on diverse soccer footage demonstrates superior performance compared to existing methods, with improvements of 2.41-10.81% over state-of-the-art approaches while maintaining real-time processing capabilities. The anthropometric body zone analysis represents a novel application of human biomechanics to sports video analysis, enabling robust ball-player association across varying camera angles and player poses. Our GPU-optimized pipeline, featuring caching mechanisms, achieves computational efficiency suitable for live match analysis, processing standard broadcast video with minimal real-time latency on consumer hardware.

Fields of Study & Subjects

Video Surveillance and Tracking Methods Human Pose and Action Recognition Video Analysis and Summarization

Publication details

Journal / source
Institutional Repositories DataBase (IRDB)
Volume
22
Issue
1
Pages
109–128
Publication date
1 February 2026
Language
EN