The Research Of Motor Imagery Recognition In Virtual Reality

Electroencephalogram (EEG) is a kind of bioelectricity signal raised on the brainof the scalp, which is caused by intracranial synapses and dendrites nerve potential.When human in active thinking or under different sensory stimuli, EEG comes to be adifference. The brain-computer interface (BCI) system based on EEG is usually usingthis kind of difference of EEG to control external devices or direct communicationwith the outside world. At present, the recognition of multiclass motor imagery EEGsignals are mainly concentrated in offline analysis, classification of4-class motorimagery has been realized. But for on-line BCI research, mainly is only focused onthe two dimensions of control (left and right, or action and stop, etc.),multi-dimensional control still exist a lot of problem: poor real-time performance, lowclassification accuracy, collected EEG signals with strong noise pollution, and so on.According to the present problems existing in the BCI technology and Combinedwith the requirements the subject, an asynchronous BCI control of roaming system ofvirtual reality based on multiclass moter imagery is developed.The system consists ofEEG signal acquisition, de-noising preprocessing, feature extraction, multiclass motorimagery pattern recognition, and realized to roaming in the virtual reality system usemotor imagery. The thesis’s specific work and innovation include:(1) In order to eliminate the noise mixed in Electroencephalogram (EEG), thispaper proposes an EEG de-noising method based on double-density discrete wavelettransform using neighbor-dependency thresholding. Firstly, high frequencycoefficients of multilayer signals are obtained by the double-density discrete waveletdecomposition. Then the detail coefficients are shrinked with neighbor-dependencythresholding algorithm, which takes into account the statistical dependencies of theneighboring coefficients. Finally, the original signal eliminated noise is resumed byreconstructing shrinked detail coefficients. The de-noising experiments of standardnoise-adding signal is carried on, the simulation results show that with approximateshift invariance method of double density wavelet de-noising effect is significantlybetter than the generation of discrete wavelet transform, under the strong noiseenvironment effect is more obvious; In addition, due to the neighbor-dependencythresholding algorithm takes into account the statistical dependencies of theneighboring coefficients, smoothing of the mutation coefficient, burr reducedobviously after de-noising. Real EEG de-noising experiment results show that the proposed algorithm can be smoothed out sudden changes in EEG disturbance signal;Spectrum analysis results also show that the algorithm effectively suppress the highfrequency noise, and the main frequency range corresponding to the power spectrumand the power spectrum of the original EEG, keep the details of the most useful in theEEG, for subsequent EEG feature extraction and pattern recognition has laid a goodfoundation.(2) Proposesed a method of motor imagery EEG feature extraction based on thenormalized AR power spectrum. Firstly estimates the AR power spectrum of EEG atelectrodes C3, Cz and C4, and then normalizes the power spectrum of movementrelated Mu rhythm (8~12Hz) and Beta rhythm (20~24Hz) to constitute the featurevector.The simulation results show that after the normalization of AR model spectrumestimation obviously improve the consistency and enhance the generalization abilityof the characteristic value. In addition, according to Mu rhythm wave has closerelationship with body movement and its nonlinear characteristics, this paper proposesa method of motor imagery EEG feature extraction based on Mu rhythm wave energyentropy. To select the algorithm which owns real-time and high identification,correctness, analyses the measured movement imagine eeg data, and use theMahalanobis distance classifier and support vector machine (SVM) analyzes andcompares the two methods of feature extraction, which adopts the normalized ARmodel power spectrum estimate methods of feature extraction and support vectormachine to recognition three kinds of motor imagery, the highest average recognitionaccuracy is80.83%.(3) An online based on motor imagery control of brain-computer interface_virtual reality system is built, and according to the requirements of online control，the method of the integration of the designed algorithm is described in detail, onlinedata processing strategy analysis (using a sliding time window), online controlstrategy (weighted voting last control command), virtual scene reconstruction andBCI communication between the Client and the building of the agreement. Involvedin the online control experiment was carried out and succeeded in identifying theimagination left hand, right hand and foot movement and eyesclose, performances ofthree subjects to participate in online control are80%、80.625%and80%respectively,and use exercise tomotor imagery to roaming in the virtual reality system.