Research On Gesture Recognition Algorithm Based On SEMG Signals

With the popularity of electronic products,intelligent equipment has been closely related to human life,its role in human life is increasingly prominent.In the field of intelligent prosthetics,surface electromyography(sEMG)signals are easy to obtain and have a promising application prospect in the processing of prosthetic control.Gesture recognition based on sEMG signals are not affected by external conditions such as application region,temperature,humidity and light.In addition,sEMG signals can provide motion intention 50-100 ms before the actual movement,therefore,sEMG signals have become the most widely used control signal source.Aiming at gestures involved in daily activities and taking gestures recognition based on sEMG signals as the research background,this paper records and denoises sEMG signals during gestural movements,analyzes sEMG signals generated by such movements,and uses features extracted from various gestures under the premise of matching the algorithm framework proposed.Finally,the sEMG signals generated by the specified action is used to recognize the motion intention,and the recognition accuracy is higher.The main contents of this paper are as follows:(1)In view of the fact that the recorded sEMG signals generally contain more noise,such as baseline wandering,power line interference and Gaussian white noise.This paper mainly deals with Gaussian white noise,due to the deficiencys of spectrum aliasing in wavelet algorithm and mode aliasing in empirical mode decomposition algorithm,in order to avoid it affecting the experimental results,this paper proposes complementary ensemble empirical mode decomposition(CEEMD)and variational mode decomposition(VMD)combined sliding window soft threshold algorithm to denoise the sEMG signals.The algorithm exploits CEEMD for signal adaptive decomposition,to overcome the mode aliasing problem existing in EMD,and then objectively defines the modal component range of the noise reduction signal according to the autocorrelation function,and then exploits VMD for decomposition of the selected modal component,which avoids the mode aliasing problem because it is a finite frequency band and the center frequencies between components are mutually exclusive.Then the mode aliasing problem caused by wavelet packet is avoided.Finally,the sEMG signals is denoised by sliding window soft threshold method.(2)Traditional gesture recognition methods are poor in accuracy and memory utilization in gesture recognition based on sEMG signals,so a real-time gesture recognition method based on time domain feature extraction of sliding window is proposed.This paper focuses on the setting of relevant parameters in time domain feature extraction of sliding window,such as window function form,feature type and channel number and their combinations,window length of sliding window,overlap rate and decision value determination in post-processing.This algorithm uses the sliding window method to extract the time domain features to improve the utilization rate of sEMG signals.Useing the random forest algorithm to be classifier,After testing 40 gestures in the two datassets of Nina Pro database DB2 and DB4,the stability test of their performance show that the algorithm has a high recognition rate and stable performance.Aiming at the shortcomings of real-time gesture recognition algorithm,a gesture recognition method based on wavelet domain features and random forest is proposed.Through the discrete wavelet packet transform,each signal in the original sEMG signals set is decomposed into three layers,and the wavelet domain features of the signals are extracted: wavelet energy and wavelet coefficient variance,and then build the features set based on wavelet domain,and finally use the random forest algorithm for gesture recognition.And through the comparison test with the support vector machine,k-nearest neighbor,linear discriminant analysis of three kinds of classification algorithms,finally concluded that using the same wavelet domain features,random forest algorithm has better classification performance.