A computer vision-based approach for performance analysis of boxing sparring and competition
Quantitative performance analysis of boxing bouts is presently limited by the time it takes to manually process video footage, as well as the expertise that is required. We propose to develop an automatic performance analysis and action classification system for boxing sparring and competition bouts that is based on a computer vision action recognition pipeline. In recent years, advances in deep learning and the creation of large image datasets has produced numerous models for action recognition and pose estimation from monocular camera footage. However, many of these data sets focus on action recognition for day-to-day tasks, and thus accuracy in classification of pose and action diminish quickly when models trained on generic datasets are applied in specific domains. We propose to augment datasets of video footage with motion capture data by developing a synthetic dataset pipeline wherein variations of the motion, 3D environment, and camera angles are generated by a photorealistic game engine. Through domain randomization and adaptation techniques, our novel synthetic dataset will result in a model for robust classification of boxer actions and other metrics, such as ring position, and allow for unprecedented motion and performance analysis for boxing.
View Full Project DescriptionSheldon Andrews;David Labbé
Own the Podium
Computer science
Arts, entertainment and recreation; Other services (except public administration); Retail trade
École de technologie supérieure
Accelerate