Research On Embedded RSVP System Based On Brain Computer Interface

Brain computer interface(BCI)is one of the important ways of human-computer cooperation.The BCI based on Rapid Serial Visual Presentation(RSVP)can detect the Event-Related Potentials(ERP)in human Electro-Encephalo Gram(EEG)by machine to realize the target detection in sequence information.RSVP-BCI technology has broad application prospects and development potential in military operations,medical analysis,leisure and entertainment and other fields.At present,the feature extraction process of EEG signals in RSVP system is cumbersome,the accuracy of traditional classification algorithms is difficult to improve,and the processing of EEG signals is mostly based on PC,which can not meet the needs of mobile portability.Therefore,this thesis studies the feature extraction and classification methods based on deep learning,and designs a human-computer interface based on RSVP in Raspberry Pi to meet the needs of mobile,intelligent and portable.The specific work is summarized as follows:Firstly,based on the specific application scenario of bad information detection in the Internet,the RSVP-BCI target detection system is built,which is called RSVP system for short.The influencing factors and precautions in the experimental process were discussed.The EEG signals of 12 subjects were collected.The main data preprocessing steps of EEG artifact removal,filtering and baseline correction were introduced.Secondly,in order to show the separability of EEG signals generated by target image stimulation and non target image stimulation,we analyze and compare their characteristics in time domain,space domain and frequency domain;The methods of coherent averaging,wavelet transform and x DAWN feature extraction are introduced;Build LDA,SVM,HDCA,x DAWN + LDA and Mallat + SVM models on offline data for comparison,and select the model with the best comprehensive performance for online experimental verification.The results show that Mallat + SVM has the best classification performance on off-line data,and the AUC value reaches 0.9;The online experimental results show that the recall rate is 69.2% and the false alarm rate is 18.6%.The classification performance of online experiment is poor,but it still ensures that most of the target information can be output on average,which verifies the feasibility of RSVP system.Then,the feature extraction and classification methods based on convolution neural network and short-term memory neural network are studied,A CNN-LSTM cascade model is proposed to automatically extract the features in time domain and space domain,and improve the classification accuracy.Compared with traditional classification algorithms and single CNN and LSTM networks,CNN-LSTM achieves the best classification performance on offline data,and the AUC reaches 0.914.The online experimental results show that compared with Mallat + SVM,the TPR is increased by 0.15 and the FPR is reduced by 0.032.The results show that the CNN-LSTM model proposed in this thesis can automatically extract time-domain and spatial features,greatly simplify the cumbersome process of feature extraction,and improve the classification performance to a certain extent.Finally,the RSVP system based on Raspberry Pi platform is realized,and the human-computer interaction page of RSVP system is designed with the help of Tkinter,a python built-in GUI development tool.The algorithms with excellent classification performance in Chapter 3 and Chapter 4 are migrated to Raspberry Pi platform.The results show that the Raspberry Pi platform has lower real-time performance than windows,but it can still meet the actual needs,which shows that the RSVP human-computer interaction interface designed in this thesis has good portability,At the same time,it can provide reference for embedded equipment to realize EEG signal processing.