Research On Gesture Recognition Based On Convolutional Neural Network

All movements of the human body are accomplished by coordinating multiple muscles under the control of the neural control system.Electromyography（EMG）,which is accompanied by muscle movement,is an important electrophysiological signal that reflects the movement consciousness and movement state of the limb in real time.In the field of intelligent prosthetics,surface EMG signals become the most widely used control signal source because they are easily acquired and can control prosthetics more directly,naturally and non-invasively.This thesis uses the current research deep learning as a classifier for gesture recognition.At the same time,the most dexterous gestures of the human body are taken as the research object,and the surface EMG signal pattern recognition is deeply explored.Moreover,for the surface electromyography signal pattern recognition,the types of gestures are rarely studied,and the defects of the fine motion patterns are lacking.The gestures designed in this study include similar actions and different actions.It is believed that the research results of this thesis will help to achieve a more natural and smooth electromyography control system.At the same time,it will have important guiding significance for the research and application of myoelectric prosthesis control,human-computer interaction and even stroke rehabilitation training.The specific research work of this paper mainly includes:（1）Due to the inherent randomness,non-linearity,weak signal and vulnerability of surface EMG signals,the quality of acquiring EMG signals is closely related to the specific acquisition form and scheme.In this paper,the mechanism and characteristics of surface electromyography are studied.Based on this,the factors affecting surface EMG signal are summarized,and the distribution of forearm muscles is analyzed to develop a new acquisition scheme for surface EMG signals.（2）In this paper,the time domain,frequency domain and time-frequency domain analysis methods are briefly introduced,and the feature extraction method suitable for classifiers is selected according to the experimental data comparison features.Finally,the FFT dimension reduction spectrum is selected in this paper.The wavelet coefficient at T₀ after wavelet decomposition is used as the input characteristic of the convolutional neural network.The original sEMG signal is used to conduct the control experiment to verify the feasibility of the experiment.（3）Three kinds of convolutional neural networks with different structures are designed according to the characteristics of EMG signals.At the same time,the signal labels are made with the help of double threshold method.The feature matrix of the above different feature extraction methods and the corresponding output signal tags are sent to three convolutional neural networks for training.Finally,it shows that the dimensionality reduction spectrum features have the best performance,and the recognition rate reaches 95.13%.