Rehabilitation gesture recognition based on improved Yolov5
Shiwei Cai, Wenlu Yang · 2024
In order to improve the rehabilitation of patients with hand motor dysfunction, an improved yolov5 network model was proposed. After the CA attention mechanism is introduced into C3 module, the model pays more attention to the position features of gestures, so as to improve the ability of capturing important features in gestures. At the same time, BiFPN can better integrate information at different scales and enhance the global understanding of rehabilitation gestures. Finally, replacing the loss function to WIoU reduces the classification loss of the pack class, making it more stable and converging faster. In order to verify the improved method, this paper made a data set containing eight gestures, and conducted training on the public data set and the homemade data set. Compared with the original yolov5 model, the accuracy of the public data set was increased by 3.3 percentage points, and the accuracy of the homemade data set was increased by 3 percentage points, and the recall rate and mAP were also improved to varying degrees. The detection accuracy of gesture recognition network is improved effectively.