Document Type



Sports science practitioners often measure ground reaction forces (GRFs) to assess performance, rehabilitation and injury risk. However, recording of GRFs during dynamic tasks has historically been limited to lab settings. This work aims to use neural networks (NN) to predict three-dimensional (3D) GRF via pose estimation keypoints as inputs, determined from 2D video data. Two different NN were trained on a dataset containing 1474 samples from 14 participants and their prediction accuracy compared with ground truth force data. Results for both NN showed correlation coefficients ranging from 0.936 to 0.954 and normalised root mean square errors from 11.05% to 13.11% for anterior-posterior and vertical GRFs, with poorer results found in the medio-lateral direction. This study demonstrates the feasibility and utility of predicting GRFs from 2D video footage.