Project Abstract

The quality of a Taekwondo pattern?��s performance is a widely used metric of the advancement of a practitioner of the martial art. With a particularly subjective method of evaluation, the project endeavoured to develop a system which provides a less qualitative method of assessment to work to prevent inconsistent and unfair decisions during examinations and competitions.Utilising the pose estimation library OpenPose, data collected during the project?��s development was used to generate 25-keypoint skeletal models of subject?��s poses. Two differing approaches to the challenge of classification were pursued: a machine learning-based approach, and a point-based conditional approach. With videos of pattern performances as inputs, the system worked to generate videos overlayed with a skeletal model, in addition to a predicted movement and accuracy score. These approaches were then compared, assessing their suitability to the project?��s context, their accuracy, and which techniques were more error-prone than others.The project has contributed to the Taekwondo community, providing a foundation for further research and work to develop a fully reliable application, allowing practitioners a means to self-evaluate the quality of their own patterns, and to provide alternatives to in-person pattern examinations and competitions.

Keywords: Computer Vision, Machine Learning, Human Activity Recognition

 Conference Details

Session: Poster Session B at Poster Stand 73

Location: Sir Stanley Clarke Auditorium at Wednesday 8th 09:00 – 12:30

Markers: Xianghua Xie, Mukesh Tiwary

Course: BSc Computer Science, 3rd Year

Future Plans: I have a job lined-up