This study experimentally evaluates whether shoulder joint kinematics can be accurately reconstructed using calibration free asynchronous RGB-D cameras under rehabilitation relevant conditions. We present a markerless framework that jointly learns temporal alignment, geometric consistency, and pose reconstruction using continuous-time modeling based on Neural Ordinary Differential Equations and implicit representations, eliminating the need for hardware synchronization or manual camera calibration. The system was validated in a controlled laboratory setting against a BTS Smart-DX optical motion capture reference during five clinically relevant shoulder movements. Performance was assessed for single- and dual-camera configurations. The dual-camera setup achieved a mean joint position error of 15.4 ± 2.8 mm with low temporal jitter (5.9 ± 0.7 mm), while the single-camera configuration showed reduced accuracy and higher sensitivity to occlusion. The results demonstrate that calibration-free asynchronous RGB-D systems can provide feasible shoulder kinematics, with a clear accuracy–complexity trade-off between single-and dual-camera deployments.
Abromavičius et al. (Thu,) studied this question.