Intense Punch Exchange Detection in Boxing Videos Using Pose-Based Machine Learning Pipeline

View at publisher Summarise with VIRAT
Authors: P.K.D.G.D. Pitigala, L. Ranathunga
Cited by: 0
References: 10

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

This paper presents a novel automated system for detecting and extracting intense exchange segments from full-length boxing match videos using a comprehensive pose-based machine learning pipeline. The system integrates multiple deep learning models including a custom-trained YOLOv8 boxer detection model, pose estimation using YOLOv8-Pose, and a Random Forest classifier for punch classification based on kinematic features. The approach addresses the significant challenge of automatically identifying high-action segments in sports videos, which traditionally require manual annotation by experts. Our system processes videos frame-by-frame, extracting 17 body key points per boxer, calculating eight-dimensional kinematic feature vectors from sliding windows of arm movements, and classifying these motions as punches or non-punches. The detected punches are then temporally binned and ranked to identify the most intense 10-second segments. Experimental evaluation demonstrates promising results with 66.1% recall, and 59.5% precision in punch detection. The system successfully demonstrates the capability of this approach to generate highlight reels that capture the most action-packed moments of boxing matches, providing significant value for sports broadcasting, analysis, and fan engagement applications. This research demonstrates the potential of the proposed modular approach to be further fine-tuned for improved accuracy.

Fields of Study & Subjects

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

Publication details

Journal / source
IEEE International Conference on Industrial and Information Systems (ICIIS)
Pages
419–424
Publication date
16 January 2026
Language
EN