A Gesture Recognition Method Based on Yolov4 Network

Abstract

With the diverse development of human-computer interaction. The gesture recognition-based interaction has a large-scale application prospect in collaborative robotics and smart home control. However, due to the similarity of gestures and occlusion, the previous methods have problems with the poor accuracy and shift of detection box. Aiming at the above issues, a gesture recognition method based on the Yolov4 deep learning algorithm is proposed. Firstly, gesture images were collected and annotated, and the data was processed by the GridMask and scale adjustment of data augmentation in order to improve the generalization performance of the network. Then K-means clustering algorithm was used to cluster the annotation boxes in the annotation dataset, by this way, the anchor box of YOLOV4 was optimized to improve the IOU accuracy. Finally, during the training process, focal loss and Consine warmup were adopted to improve the unbalanced sample number of classes and overfitting of the network. The experimental results shows that the proposed algorithm outperforms the main target detection models which include Yolov4, Yolov3 and Faster RCNN, the average recognition accuracy of this method reaches 99.4% and the FPS is 33fps. The proposed algorithm has good real-time performance.