Research On Key Technologies Of Hand Action Pattern Recognition Based On Surface Electromyography

Hand as an important organ of mankind is not only the main tool that is involved in production labor,but also an important communication tool.Aside from performing sophisticated gestures and movements,human hand can also accomplish all kinds of creative work and realize the transformation of the natural world.In addition,the hand can also help express thoughts,feelings and behavior intentions,which becomes important medium of auxiliary interaction.Surface electromyography(sEMG)as an electrophysiological signal associated with muscle activity contains abundant information of conscious during body movement,thus making it the ideal control source of the biological signal of bionic rehabilitation equipments and new type human-computer interaction devices.The in-depth study of pattern recognition of human hand action is based on sEMG signal,the goal is to achieve low-cost,fast and accurate hand action recognition ability,and then convert recognition results to control instructions with multi-degree freedom to drive the peripheral equipments.The study will finally meet the urgent need of physically handicapped patients for independent control over intelligent bionic hand,and then provide more choices of reliable and humane auxiliary rehabilitation equipment for the rehabilitation medical field.Meanwhile,it provides support for users to choose the appropriate action type according to the situation so as to express their own intentions and exchange information with the external environment.Therefore,this study is of great research significance and application value.There still exists many problems and challenges in order to recognize actions with high accuracy rate due to the complicated uncertainty of the sEMG and the diversity of human body movement.For instance,filtering of target muscle group,selection and extraction of sEMGfeatures,modeling method of consecutive action or movement are not good enough.Low recognition accuracy rate,limited kinds and quantity of recognition types,low timeliness of thealgorithm,are all the factors that prevent this technology from entering the application field.Aiming at solving these problems,this paper carried out the study in the following aspects:(1)Scheme selection of sEMG signal samplingsEMG signal has inherent characteristics like randomness,nonlinearity,and is susceptible to interference,therefore,the quality of the obtained EMG signal is closely related to the sampling scheme.In early studies,the need to choose appropriate muscle group to extract sufficient feature information is not urgent due to few action patterns,good result can often obtain by using experience to determine sampling system scheme.However,as increasing types of actions emerge to be identified,the drawbacks of empirical methods gradually appear.To minimize the randomness of the signal,as well as avoid the blindness of the electrode position selection,experimental research for determining optimal electrode position is carried out,and then an all-new sEMG sampling scheme is designed.(2)Action area division of the data and label sample preparationClassifier design based on machine learning has high demand on the type property and quantity of the training samples.Therefore,the extraction of large quantities of samples and the addition of action tags become essential.Fast and concise method for sEMG signal processing is investigated and detection algorithm for action area of sEMG signals is designed,the action type of the action samples can then be labelled automatically,thus facilitating the preparation of great amounts of training samples,improving the accuracy rate,reducing the overall time cost,and providing support for subsequent classifier training.(3)Extending the types and quantities of hand movements and hand gesturesIn order to meet the demand of action type diversity on bionic rehabilitation equipments and new type human-machine interaction applications,recognizable types of hand actions should be extended.A classification system that can recognize 20 types of hand actions is designed,different types of feature extracting algorithms and classification algorithms are analyzed,an appropriate configuration scheme that satisfies multi-pattern classification is finally determined through experimental comparison.(4)Comprehensive performance evaluation of multi-pattern classification systemIn order to avoid a biased judgment of the classification system performance just based on recognition accuracy rate,other factors that can influence the system are introduced to evaluate the performance of different configuration schemes,such as the complexity of feature extraction and pattern recognition algorithm,robustness and real-time performance,system response time,instantiation cost,and so on.