Evaluation of Commercial Camera-Based Solutions for Tracking Hand Kinematics

Tracking hand kinematics is essential for numerous clinical and scientific applications. Markerless motion capture devices have advantages over other modalities in terms of calibration, set up, and overall ease of use; however, their accuracy during dynamic tasks has not been fully explored. This study examined the performance of two popular markerless systems, the Leap Motion Controller 2 (LM2) and MediaPipe (MP), in capturing joint motion of the digits. Data were compared to joint motion collected from a marker-based multi-camera system (Vicon). Eleven participants performed six tasks with their dominant right hand at three movement speeds while all three devices simultaneously captured the position of hand landmarks. Using these data, digit joint angles were calculated. The root mean squared error (RMSE) and correlation coefficient (<i>r</i>) relative to the Vicon angles were computed for LM2 and MP. LM2 achieved a lower error (<i>p</i> < 0.001, mean RMSE = 14.8°) and a higher correlation (<i>p</i> = 0.007, mean <i>r</i> = 0.58) than the MP system (mean RMSE = 22.5°, mean <i>r</i> = 0.45). Greater movement speed led to significantly higher RMSE (<i>p</i> < 0.001) and lower <i>r</i> (<i>p</i> < 0.001) for MP but not for LM2. Error was substantially greater for the proximal interphalangeal joint than for other finger joints, although <i>r</i> values were higher for this joint. Overall, the LM2 and MP systems were able to capture motion at the joint level across digits for a variety of tasks in real time, although the level of error may not be acceptable for certain applications.