The Research Of Control Methods Of Virtual Reality Based On Multi-class Motor Imagery Recognition

Brain Computer Interface（BCI）is an artificial and independent communicationchannel which does not depend on nerve cells and muscles. The brain could use thischannel to communicate with external devices. Electroencephalogram (EEG) is a kindof physiological signal, caused by postsynaptic potentials of numerous nerve cells inthe cerebral cortex. EEG can effectively reflect the state of mind and neuronal activity.The BCI system based on EEG has noninvasive and intuitive features, high researchvalue and good prospects, becoming an important branch of BCI research.Supported by the National Natural Science Foundation of China (61172134)and according to the requirement of issue of “The Research of Control Methods ofVirtual Reality Based on Multi-class Motor Imagery Recognition”, this paper startsfrom the mechanism and characteristics of EEG, systematically studies the signal’sacquisition, preprocessing, feature extraction, pattern recognition, then this papercombines with virtual reality technology to design and implement an on time BCIsystem based on four class motor imagery. Finally paper completes the experiments ofcontrolling the direction and movement of virtual scene by EEG on this system. Themain content and the innovation of this study are as follows:In the aspect of EEG de-nosing, Considering the defect of losing some importantdetails resulted from traditional soft threshold function, the paper proposes an EEGde-nosing method based on bivariate shrinkage functions and high density discretewavelet transform. Fristly, HDDWT is used to decompose the EEG signals, then theBivariate Shrinkage Functions is applied to shrink the coefficients. The simulatedsignals are analysed to demonstate that this method has an effective effect on noiseresistance and retain the edge character of EEG. Compared with threshold shrinkagefunction in common use, the algorithm has a better capability on details descriptionand signal reduction.In the aspect of EEG feature extraction, in order to eliminate the severe frequencyaliasing resulted from Discrete Wavelet Transform, an algorithm of motor imageryEEG feature extraction based on normalized energy of double tree complex wavelet isproposed. The simulated signals is analysed to show that DTCWT has a nice propertyof reducing aliasing effects. The sub-band energy extracted by DTCWT can reflect the motor image features better.Then combined with the nonlinear characteristics andERS/ERD phenomenon of EEG, DTCWT is used to decompose the EEG and thencalculate the permutation entropy of rhythm α and β. The experimental resultsshow that the discrimination of permutation entropy based on rhythm signals is moreobvious than feature extraction of single permutation entropy.In the stage of EEG recognition, this paper proposes an EEG recognition methodof Fuzzy Support Vector Machines based on genetic algorithm. FSVM fuzzifies thetraining samples by choosing an appropriate membership function, reduces themembership of noises and outliers, weakens the impact of outliers on theclassification model, and improves the classification performance of system. Then GAalgorithm is used to solve the problem of selecting optimal classification performanceparameters (punishment factor C, kernel function parameter γ).The paper selectsC3,C4,Cz,Cp4four channels which are related to left hand, right hand, feet andtongue motor imagery, extracts normalized DTCWT energy and permutation entropyof rhythm signals as the features of these four motor imagery. Offline experimentalresults show that GA-FSVM obtains a higher recognition rate compared with theconventional SVM. The average recognition rate is68.75%.The paper successfully builds a virtual reality platform controlled by motorimagery EEG signals, creates the overall proposal of this platform, redesigns BCIClient, virtual reality system by Matlab and C++programming, establishes a BCIsystem including the signal acquisition module, processing and recognition EEGmodule, virtual reality module. The subjects finish the experiment of controlling theonline synchronous BCI system by left hand, right hand, feet and tongue motorimagery. Experimental results show that the classification accuracy rate is close tooff-line analysis. The average recognition rate is65%. The paper attempts to applyfour kinds of motor imagery EEG signals to roam in the system of virtual reality andthe results are analyzed.