Reshearch On Gesture Recognition Algorithm Based On Surface Electromyography And Parallel CNN

As a physiological signal of the human body,surface electromyography(s EMG)signal contains the related information of muscle state and human motion intention.And gesture recognition based on s EMG has always been a frontier technology and research hotspot in the field of human-computer interaction(HCI).Traditional methods mostly use machine-learning algorithms to recognize gesture based on s EMG,but their accuracies are not very high.In recent years,with the improvement of hardware performance,many researchers began to introduce deep-learning alogrithms into s EMG-based gesture recognition and ahieved certain results.However,there are still some limitations in their studies: the time sequence of s EMG and the synergistic effect between muscle groups are not taken into account,which lead to limited accuracy of the model.This thesis studies the above problems and the major contributions are summarized as follows.(1)Aiming at the problem that the time sequence of s EMG is seldom considered into s EMG-based gesture recognition,a parallel CNN-RNN based gesture recognition algorithm is proposed by introducing recurrent neural network(RNN)into convolutional neural network(CNN)architecture.The model adopts two parallel neural network branches,and CNN of each branch applies long and narrow convolutional kernels.In addition,long short-term memory is added in each branch to extract the time sequence features of the s EMG,and finally the feature fusion of the branches and classification are carried out.The experimental results on multiple Nina Pro databases show that,the combination of CNN and RNN can extract more spatiotemporal characteristics of s EMG than using them separately in the model,thus improving the accuracy of gesture recognition.(2)In order to futer improve the recognition rate,based on the hypothesis that gesture depends on the coordination of multiple muscle tissues,a deep learning algorithm of multi-stream fusion for s EMG-based gesture recognition is proposed in this thesis.Dividing the multiple streams based on the position of the electrodes on the arm,this algorithm models the obtained multiple steams with the parallel CNN,and fuses the feature maps of all streams for final classification.The experimental results on multiple Nina Pro databases show that in the framework of multi-stream fusion,associative modeling of s EMG generated by adjacent muscle groups on the arm can effectively improve the accuracy of gesture recognition.