Research On Motion Recognition And Human-Computer Interaction Method Based On SEMG Signals

With the development of the computer technology,the way of human-computer interaction is becoming more and more convenient.As a novel way of human-computer interaction,motion recognition has been increasingly favored by researchers and the market.As a kind of physiological signal of human body,surface electromyography(s EMG)can detect muscle activity in the non-invasive way,which can be used for the human motion recognition.Compared with mainstream image-based motion recognition methods,s EMG-based methods have problems of unstable performance and high error rate.However,its superior portability and low energy consumption are the key to the realization of motion recognition on wearable devices in the future.Therefore,this paper mainly studies the motion recognition algorithm based on s EMG signals for improving the recognition performance and design a motion recognition system.Its main work is as follows:Firstly,this paper proposes a novel motion recognition algorithm named convolutional vision transformer(Cvi T),which mainly includes three parts: data preprocessing,feature extraction and action classification.Firstly,the original signal is filtered,sliced and normalized in the preprocessing part.Then the feature extraction and the classification are carried out by the proposed model to output the result.In Nina Pro DB2,the proposed method achieves 80.02% with the window length of 200 ms.In the subset of Nina Pro DB2(Exercise E1),the proposed method achieves 83.47% and 84.09% with the window length of 200 ms and 300 ms respectively.In the subsets of Nina Pro DB5(Exercise A,Exercise B),the proposed method achieves 76.83% and 73.23%respectively.The experimental results demonstrate that the proposed Cvi T has better performance than most current approaches.Then,this paper proposes a motion recognition method based on multi-task learning to solve the low recognition efficiency of single-task learning.Also,the proposed method enlarges the number of recognition tasks based on s EMG signals,which is very suitable for the model deployment in complex multi-task scenarios.Specifically,the algorithm requires preprocessing and data enhancement of multi-label data.Then,a shared-feature extraction module is used to extract the features of the multi-label data,and the feature maps are input to several independent classifiers to output the classification results of different tasks.Experiments have shown that reducing the sliding length can improve the classification accuracy of single-task models or multi-task models and the recognition accuracy increases by 1.64% at most.At the same time,this paper explores the influence of the multi-task mechanism in the basic model.The multi-task model can achieve the classification accuracy of 98.95%,which is higher than the accuracy achieved by single-task models.Compared with the size of the single-task model,the size of the multi-task model is shrunk by nearly half.Also,the multi-task model is more conducive to the deployment of the large-scale model,which has much broad application prospects.Finally,this paper proposes a structure of motion recognition system based on s EMG signals.It mainly includes signal transmission module and motion recognition module.The signal transmission module is used to transmit the s EMG signals to the intelligent terminal device by the appropriate signal transmission protocols.Also,it is used to store data locally and in the cloud for the establishment of relevant s EMG datasets.The motion recognition module is mainly used to process the s EMG signals received by the intelligent terminal device and finally output the current result of motion information.