Evaluating the accuracy of table tennis 3D markerless motion capture based on two video cameras: a pilot study

This pilot study presents a 3D markerless motion capture system for table tennis using binocular cameras, based on the OpenPose framework and linear triangulation to obtain 3D joint motion data. The system's accuracy was compared to a marker-based system, and the impact of different camera.

A comparative experiment was conducted between the proposed markerless system and a marker-based reference system. Three camera configurations were tested: RR-100 (right-right placement, 100 cm baseline), RR-145 (right-right placement, 145 cm baseline), and RF-100 (right-front placement, 100 cm baseline). Dynamic Time Warping (DTW) similarity of elbow joint angle curves was used as the primary accuracy metric, and statistical comparisons (p-values) were performed to identify significant differences in system errors across configurations.

(1) At a camera position of RR-100, the DTW similarity of the elbow joint angle curves between the markerless and marker-based systems was 0.941, 0.719, and 0.854, respectively, similar to the marker-based system. The DTW similarity reached 0.948, 0.730, and 0.887 for RR-145 and 0.949, 0.729, and 0.860 for RF-100. (2) Significant differences in system errors were found at varying baseline distances with L = RR, showing better accuracy at 145 cm (p < 0.01). (3) Significant differences in errors were found with L = RF and a 100 cm baseline, with better accuracy at RF (p < 0.01).

(1) The markerless system, though slightly less accurate, can also capture joint motion data. (2) Increasing baseline distance improves accuracy when the camera position is fixed. (3) With a fixed baseline distance, the system with the RF camera position outperforms the RR position in accuracy.