Evaluation of a markerless motion capture system for measuring mechanical work during tennis strokes

This study evaluated the agreement between markerless motion capture and criterion methods for estimating mechanical work (external and internal) performed during tennis strokes. Sixteen tennis players performed 10 serve, forehand and backhand movements whilst motion data were captured concurrently with a custom 3D markerless system (utilising open-source pose estimation; HRNet and OpenPose) and two criterion methods: marker-based motion capture and ground reaction forces. Centre of mass kinetic and potential energy were calculated and used to compute external mechanical work from all methods, whilst segment kinetic energies were calculated and used to compute internal mechanical work from the kinematic approaches only. Small biases (≤4%) and moderate limits of agreement (LoA; ±6-10%) from Bland-Altman analyses indicated strong agreement with the criterion systems for external work across multiple strokes. Although small biases of 4-7% (HRNet) and 1-2% (OpenPose) were observed for internal work, high-frequency artefacts in markerless kinematics (particularly evident for OpenPose) led to high random errors (LoA of ± 8-13% [HRNet] and ± 12-21% [OpenPose]) and large inter-individual differences. Our markerless approach shows promise for measuring external mechanical work during tennis and has the potential to be implemented in the field as a non-invasive tool for on-court workload monitoring.