An Gesture Recognition Technique Based On Multi-channel Surface EMG Signals

Surface Electromyography(s EMG)is a kind of physiological electrical signal measured on the skin,which contains a lot of movement information generated by muscle contraction.Because this collection process is non-invasive,it is very suitable for consumer,medical,industrial and other application scenarios requiring human-computer interaction,but there are few commercial applications that can really be implemented.The main limitations are low identification accuracy,weak transmission real-time,poor stability,high device cost,and these factors lead to the promotion of commercial products is more difficult.Therefore,in order to solve the accuracy problem of gesture intention recognition,more and more acquisition channels are used in the related research of surface EMG signals,and the recognition accuracy is also getting higher and higher.However,it also inevitably increases the cost of devices and the cost of space-time calculation,such as gesture recognition delay and excessive storage space occupation.In addition,although previous studies have high accuracy in public datasets,their measured performance for new individuals is often poor and cannot meet good universality.Due to the limitations of the above comprehensive performance,this thesis designed a set of EMG acquisition and recognition software and hardware system.Eight channels of surface EMG signals were collected from human upper arm,and a set of 21-class gesture database was constructed.Because deep learning has great potential for feature extraction of signals,this thesis designed a lightweight gesture recognition algorithm using deep learning method.In this thesis,the public gesture database and the private gesture database were verified and repeatedly tuned,so that the model indicators such as accuracy,lightweight stability and individuation reached a good performance balance,and finally realized the real-time multi-gesture classification task,so that it could meet the industrial application scenarios.The main content of this thesis includes the following aspects:1.The basic pruned neural network model is designed according to the characteristics of data distribution by verifying the data on three open datasets which is widely analyzed.Using this model,we proposed a spatial attention mechanism without increasing the computational overhead,which can improve the performance of high-density EMG database.In the static database of Zhejiang University,the accuracy of the optimal attention mechanism network reached 96.06%,which was4.44% higher than that of the baseline classification without attention mechanism.In the German dynamic database,the accuracy of SOTA is 96.07%.In order to realize the lightweight deep learning network,this thesis used 1DCNN and voting mechanism to further improve the accuracy.The near-saturation accuracy of 98.6%can be achieved on the Zhejiang University 8 classification gesture task,and this method is a lightweight algorithm model with obvious redundancy optimization of spatial-temporal overhead.2.After analyzing the typical EMG database,this thesis built a set of EMG acquisition and recognition hardware and software system.Among them,the self-designed EMG acquisition and recognition software was used as the software system basis,and the 8-channel EMG chip developed by the laboratory and the edge AI computing platform of NVIDIA were used as the hardware system basis to jointly complete the collection,processing and recognition of gesture data.After setting up this hardware and software system,this thesis designed a set of data collection specifications and built a set of SIAT-EMG with 10 people and 21 classes of general gesture database,so as to further design a lightweight and individualized network that conforms to the characteristics of data distribution.3.Based on the independent database,the hyper-parameter tuning of the deep learning network is carried out again,and a network model with such indexes as accuracy,stability,individuation and lightweight is trained.For 21-class gestures task,the optimized network model achieves 83% accuracy without using voting mechanism,and the model has a relatively balanced recognition accuracy for each subdivision gesture.Finally,this thesis deployed the network on the edge computing device(NVIDIA Jeston Nano platform)to evaluate the performance of the actual model and conduct online gesture classification.The computational amount and storage overhead required by the model in this thesis have been reduced exponentially.The size of the deep learning model is only 21 K Bytes,and the computational amount is only 0.84 M FLOPS.Online multi-gesture task recognition is accurate and smooth,with almost no delay of gesture recognition.In this thesis,the software and hardware of EMG acquisition and recognition were built,21-class static gesture databases were constructed,and a lightweight neural network was established by deep learning to realize real-time,accurate and low-power EMG intention recognition,which is of guiding significance for the landing of EMG wearable products.