Author Identifier

Wayne Poon

http://orcid.org/0000-0002-3098-2055

Date of Award

2024

Document Type

Thesis

Publisher

Edith Cowan University

Degree Name

Doctor of Philosophy (Integrated)

School

School of Medical and Health Sciences

First Supervisor

Oliver Barley

Second Supervisor

Syed Afaq Ali Shah

Third Supervisor

Chris Abbiss

Abstract

Evaluating the detection and classification of punches in boxing performance presents challenges such as high punch volume, rapid execution, and dynamic athlete movement, resulting in potential occlusion. This thesis aimed to i) examine the reliability of single- and multiple-camera system configurations in classifying punches, ii) assess the reliability of punch classification with and without headguards, and iii) create a custom dataset to develop an Artificial Intelligence (AI) model for automated punch detection and classification. In Study One, three observers assessed a simulated boxing bout using single- and multiple-camera systems in a test-retest manner. The multiple-camera system showed superior intra-observer reliability (Kappa 0.71 to 0.92) compared to the single-camera system (Kappa 0.46 to 0.68). Inter-observer reliability across observers was moderate for the single-camera system (Kappa 0.46 to 0.56) and fair for the multiple-camera system (Kappa 0.23 to 0.43). Disagreement rates were notably lower in the multiple-camera system (3.78% to 14.59%) compared to the single-camera system (15.13% to 34.05%). A strong correlation was observed between occlusion level and assessment difficulty in the single-camera system (d-value 0.974 to 0.992, p < 0.01). Study Two examined the reliability of assessing punches with and without headguards using a multiple-camera system. Intra[1]observer reliability ranged from slight to almost perfect (Kappa 0.20 to 0.95) with headguards and was more consistently high without headguards (Kappa 0.59 to 0.81). Inter-observer reliability was low to moderate with headguards (Kappa 0.27 to 0.54) and moderately consistent without headguards (Kappa 0.52 to 0.55). Disagreement rates were substantially higher with headguards (up to 41.46%) compared to without headguards (up to 19.49%). In Study Three, 1358 punches were annotated to create a dataset for AI model training and testing. YOLO version 5 large (YOLOv5l) and X-large (YOLOv5x) were the best-performing models, achieving an F1 score of 0.56 and mAP@0.5:0.95 of 0.19. A custom detection pipeline incorporating post-processing techniques showed 100% intra-model reliability. When compared with human observer consensus, YOLOv5l with a weighted threshold of 0.3 demonstrated moderate to substantial agreement (Kappa 0.57, percent agreement 81%). The AI models completed assessments in approximately 2 minutes, significantly faster than human observers (29 minutes to 2 hours). This thesis demonstrates that boxing performance analysis is affected by camera positioning and protective equipment. The developed automated punch detection system shows promise in streamlining analysis but highlights the ongoing need for human expertise. These findings have important implications for standardising boxing performance analysis and integrating automated systems in combat sports.

DOI

10.25958/sawg-j364

Access Note

Access to this thesis is embargoed until 21st December 2028

Available for download on Thursday, December 21, 2028

Share

 
COinS