Computer Vision-Based Classification of Manual Wheelchair Propulsion Patterns

Manual wheelchair users (MWC) are at high risk of upper limb overuse injuries due to inefficient propulsion techniques. While clinical guidelines recommend the semicircular (SC) propulsion pattern to reduce joint stress, accessible automated feedback classification tools are lacking in both clinical and remote rehabilitation contexts. This study had two goals: 1) to develop a vision-based classifier using 2D wrist trajectory images and a convolutional neural network (CNN) to distinguish SC from non-SC patterns at the cycle level; and 2) to validate its performance on simulator-based application data against expert consensus. Ten participants (including wheelchair users and non-disabled individuals) performed four propulsion patterns at three speeds on a wheelchair simulator. Wrist trajectories were recorded via RGB video and processed using MediaPipe markerless tracking, producing ${150}\times {150}$ -pixel images to train the CNN. For validation, a dataset of simulator-based application propulsion cycles was collected, and a sample of 180 images (127 SC, 53 non-SC per model prediction) was labeled by three experts, with consensus reached through majority vote. The CNN achieved 94% accuracy in controlled conditions, and 93.3% on simulator-based application data, with precision, recall, and F1 scores all at 95.3%. This low-cost, video-based system demonstrates the feasibility of automated propulsion pattern classification and shows potential for future integration into cycle-by-cycle, near real-time feedback tools, such as VR-based wheelchair simulators, to support skill acquisition and prevent injury.