Search

This study investigates the reliability of marker-less vision systems for contact detection between hand and hand-rim during wheelchair propulsion. The measurement system uses a camera collecting RGB and depth images. The hand is detected through Mediapipe, a software able to recognize the key points of the hand on the RGB image. Hand position is expressed with respect to the wheel. A classifier is used to determine if there is contact between hand and hand-rim. To validate this procedure, 17 non-disabled participants pushed the wheelchair on an ergometer during six tests, given by the combinations of holding a tennis racket in their right hand while propelling at three different speeds: 4, 5.4 km/h, and maximal sprint. Since validated contact detection methods are lacking in the literature, experts provided a reference by manually evaluating video recordings and force signals frame-by-frame. The results showed that the hand identification by Mediapipe is not influenced by the presence of the racket but by the speed. Contact events were detected in 99.5% of the cases. The mean error in contact time detection was −5 ms for the starts and 18 ms for the ends, the standard deviation was 48 ms for both, and the combined root mean square error (RMSE) was 48 ms for the starts and 50 ms for the ends. These values, once corrected for the systematic effects, lead to a standard uncertainty of approximately 0.05 s, corresponding to 15% of the average contact duration. The study highlights the potential use of marker-less vision systems for contact detection in wheelchair propulsion.