Punch force estimation from videos during boxing training

Abstract This study presents an innovative method for estimating punch forces during training using a graph neural network (GNN). The model was trained on human pose data, extracted from video footage, with corresponding punch force values recorded by a load cell, serving as the target output. Experimental tests involving six participants showed that, under a subject-dependent approach, the predicted punch forces had an average mean absolute error of approximately 84.8 N compared to the actual measurements. By estimating force from training videos, this technique offers a novel means to evaluate performance based on an athlete’s posture, providing coaches with a valuable tool to monitor and enhance training processes.