Marker-less Vision System Based on RGB Camera for Wheelchair Tennis Contact Detection

This study evaluates the performance of a marker-less vision system to measure the start and the end of the hand contact phase during wheelchair propulsion. The system is composed by Mediapipe, a software that recognizes the hand on the RGB images and a classifier that determines whether the hand is in contact with the hand-rim for each frame. The system was validated with 17 subjects that propelled a wheelchair placed on a wheelchair ergometer at three different speeds (4 km/h, equal to 1.11 m/s, 5.4 km/h, equal to 1.5 m/s and maximal sprint) while holding a tennis racket or with free hands. The system, using data coming from RGB images only or combined with depth information, measured the start and end of the contact with a standard uncertainty of 55 ms. This value was comparable to 50 ms, obtained in a previous study using depth images 1]. This study also explains the implementation of three different reliability scores, aiming to evaluate the number of frames in which Mediapipe did not recognize the hand and the likelihood of each data point being classified as “contact” or “no-contact”. These scores showed a weak correlation with the error in contact detection.