Hand Gestures Recognition Based On Optimization Of SEMG Training Set Weight

The hand is an important function and sports organ of human body.For upper limb amputation patients,accidents caused by upper limb losing and limb destroy not only severely affects the normal life of the patients,but also makes great harm to their psychological health.Therefore,the installation of artificial limbs has become an effective way to help them restore the hand function.Among them,myoelectrical prosthesis is widely used for its intuitive control and flexible feature.However,in the process of myoelectrical prosthesis,the gesture execution can not always reach the ideal effect of recognition.How to improve the phase recognition of myoelectric prosthetis becomes research focus in biomedical engineering field.Prosthetic hand control includes the acquisition of sEMG signal,feature extraction,pattern recognition and action execution.Among them,the hand gesture recognition effect largely depend on the way of building the training sample set in pattern recognition.Therefore,in this paper,we proposed a strategy based on the optimization of the training sample weights of surface eletromygram signals,which is used to improve the recognition effect of hand gesture,so as to adapt to the development of dexterous hand control.The whole process of hand motion can be decomposed into several independent gesture with different phases,and gestures between adjacent phases have similar pose and trajectory.Therefore,we firstly study the main factor that affect the similar hand gesture recognition.In this paper,we designed an experiment including 2 groups similar hand gestures(grasping and pinch group),to analyze the influence on recognition of similar hand gestures when we change the way of training set feature space construction.Firstly,we acquired the sEMG of 7 subjects performing similar hand gestures,after feature extraction,the training sample proportion among the similar gestures is adjusted in feature space.The result showed that in both grasping and pinch group,the non-averaged training sample proportion of sEMG feature vector among the similar gestures significantly obtain the higher recogniton rate than traditional averaged proportion construction.In addition,the whole process of hand motion involve the change of speed in different phases.Therefore,we study the main factor that affect the recognition of the hand gesture at different speed.We recruited 5 subjects to perform fist-grasping and finger-pinch hand motions at three different speed,and simultaneously acquired the sEMG signals and 3D coordinate data of five finger.Then four time domain features—root mean square(RMS)?absolute amplitude(WAV)?slope symbols change(SSC)and zero crossing(ZC)were extracted from sEMG signals,then the moving finger angular speed according to the 3D coordinate data was calculated.Following,the regression analysis was used to determine which sEMG characteristic values were linearly related to the speed,then the feature space of training set is constructed by using the characteristic values.At last,we adjusted the training sample proportion of three kinds of speed,then put them into classifier.The result showed that hand motion at different speed can make better classification rate when construct non-averaged training set sample in feature space,than that of averaged proportion.In order to further verify the feasibility of non-averaged sEMG training set sample distribution used in gesture recognition of prosthetic hand.A virtual hand recognition and control system based on surface electromyography is designed.Here,we applied the training sample proportion calculated off-line of three phase – start(S),middle(M)and end(E)of a whole hand motion to test online.Here,in fist-grasping gesture – start:middle:end =25%:25%:50%;In fist-closed gesture – start:middle:end=30%:40%:30%;In finger-pinch gesture – start:middle:end =35%:30%:35%,and the off-line peak recognition rate were 84.2%,81.2% and 83.1%.Later,a classifier is constructed based on the weighted sample set in off-line process,then put them to test on-line.The recognition rate were 86.8% in middle stage,60% and 57.4% in start and end stages.The result showed that the non-averaged EMG training set can effectively improve the recognition of different phases of hand motions,and can meet the requirements of real-time control.This paper carried out a preliminary study on training sample weight optimization of surface EMG based on myoelectric prosthetic hand.We proposed and validated a nonaveraged training sample distribution strategy can make improved recognition effect on similar hand gestures,hand motion at different speed and different phases of a whole gesture,which could provide new idea for subsequent sEMG dexterous prosthetic hand control.