A Hand Gesture Detection Algorithm and Quantization Implementation

Using deep learning to recognize the hand gestures of deaf-mutes quickly and to translate coherent sign language efficiently is of great social significance to solve the communication barriers between deaf-mutes and normal people. However, the detection network based on deep learning cannot be deployed to mobile terminals with limited memory and computing power due to its huge model and computation. To solve this problem, an efficient gesture detection algorithm based on pose estimation is proposed, that is, after judging whether people raise their hands to make gestures by pose estimation, detection network quickly recognizes hand gestures. Firstly, a small dataset of high frequency isolated hand gestures is constructed to train and test the network.Secondly, the structure of the original pose detection network OpenPose is optimized from backbone, initial stage and refinement stage to make the network more lightweight. Then, a pruning algorithm combining channel and layer is carried out to compress the network. After pruning, the size of the model was only 7MB, and the inference speed was significantly increased to 106fps, and the accuracy sacrifice was only 1%. After the fine-pruning, the accuracy rises and slightly more than the original. Finally, the network is deployed to the chip AR9301, which is accelerated by hardware accelerator quantization to further improve the performance.