Développement d’une méthode d’extraction automatique de métriques de performance en boxe basée sur la vision par ordinateur

Abstract

In boxing, studies have shown that punching efficiency, i.e. ratio between punches landed and the total number of punches thrown, is key for winning a bout. The movements of the boxers and their control of the ring space are also important as they are part of ring generalship, which is a component that is considered in the boxing scoring system. Therefore, monitoring such metrics is crucial to assess an athlete’s performance or to gain knowledge about a boxer’s fighting style. However, subjectively observing and manually annotating these events during a bout is a tedious task that must be done by a sport performance analyst. Previous attempts to automate the process required specialized equipment such as time-of-flight cameras, motion capture systems or inertial sensors. The markers and sensors required by these methods are not practical to use in real bouts since they can disrupt the natural motion of the boxers, may be damaged upon contact, or worst, may injure the athletes. In this work, we propose an innovative computer vision-based method to automatically extract performance metrics such as the number and type of punches thrown by a boxer from a shadowboxing video, and ring control through trajectories and heatmaps from a bout video. Our cost-effective approach requires only monocular images recorded by a single video camera, which eliminates the need of any specialized equipment. On average, it achieves 77 % weighted classification accuracy of the following stances and punches: unguarded, guarded, jab, cross, lead hook, rear hook, lead uppercut and rear uppercut. Moreover, it can track a boxer’s position in the ring within a 12,2 cm margin of error on average.

Date	14 Sept 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	David Labbé (Supervisor), Sheldon Andrews (Co-supervisor) & Jocelyn Faubert (Co-supervisor)