Introduction: Marker-based kinematics is hardly compatible with sports movement analysis, especially because it hinders natural movement. Hence, deep-learning based markerless solutions are being developed, such as Pose2Sim [1]. Sports in competition may also be subject to setup constraints, preventing the use of heavy equipment and installation. The use of lightweight action cameras, with post-calibration and post-synchronization, can be pertinent in this context. This study aimed at investigating the accuracy of the Pose2Sim pipeline in such suboptimal conditions. We assessed the measures of key performance indicators in boxing, which involves swift and 3-dimensional full-body movements.
Methods: One male elite boxer performed 6 repetitions of a boxing sequence composed of 3 punches (jab, high hook, low hook). His 3D motion was captured both with a marker-based protocol, and with a light-weight markerless one. The latter involved 8 GoPro cameras, which were post-calibrated based on the dimensions of the boxing ring, and post-synchronized based on time-lagged correlations of 2D keypoints speeds between paired cameras. The markerless kinematic analysis was then performed with Pose2Sim. Both protocols used the same OpenSim [2] model to optimize inverse kinematics. Displacement of lead foot, pelvis, lead fist, and velocity of lead fist were analyzed for the jab. Rotation of rear foot, pelvis, displacement of rear fist, and velocity of rear fist were analyzed for the hooks. Waveform similarity was assessed with the inter-protocol coefficient of multiple correlation [3]. Time and magnitude of peaks were also compared.
Results: Results from the marker-based and markerless approaches demonstrated excellent waveform similarity (CMC>0.94), and times-to-peak exhibited intervals of under one frame (i.e., 17 ms). Magnitudes-at-peak were very close for displacements (below 6 cm), but not for fist velocity on jab (1.2m/s difference) nor for rotations on hooks (up to 20° difference, e.g. for pelvis rotation).
Conclusion: Despite the use of action cameras, and of suboptimal calibration and synchronization procedures, our lightweight markerless protocol gives satisfying results for the analysis of key performance indicators in boxing, especially for limb displacements. Nevertheless, results should be taken with caution for velocity and for rotation measurements. Such a light-weight markerless protocol could be useful in situations where more accurate marker-based approaches are not conceivable, such as sports competitions.