Transfer Learning Brain-computer Interface For Intelligent Assisted Driving

As the number of new drivers in China has soared and there are loopholes in traffic supervision,the number of traffic accidents caused by irregular driving behaviors remains high.Therefore,it is of great social significance to strengthen the regulation of driving behaviors during driving.At present,the intelligent driving technology is one of the most effective solutions to the above problems,but prior art solution ignoring the core safety factor of the driver itself.If the driver 's driving intention and driving behavior can be effectively predicted and further used to assist driving,the probability of traffic accidents could be expected to be significantly reduced.To this end,this article introduces the motor imagery brain computer interface(MI-BCI)technology into the field of intelligent assisted driving to effectively obtain the driver's driving intention,In order to realize the rapid transition of driver-vehicle state,and enhance the real-time of information interaction between adjacent vehicles,and effectively avoid traffic accidents.On the other hand,transfer learning is introduced to solve the problem of long training sessions of the BCI system and effectively improve the efficiency of the BCI system.This article mainly contains the following work.Firstly,the generation mechanism of motion imaginary EEG was studied.To deal with the characteristics of nonlinearity and non-stationarity,we extracted EEG features from time-domain,frequency-domain,time-frequency-domain to comprehensively characterize the phenomenon of ERD/ERS.Furthermore,introduced two common pattern recognition methods,support vector machine(SVM)and BP neural network.Secondly,we introduced transfer learning into MI-BCI system.Then,considering that the existing transfer learning methods are not efficient in transfer and the limitations of the applicable field,a hybrid transfer learning model with integrating the advantages of instance transfer and feature transfer learning methods was built in this paper.Step one,we realized the transfer of the instance level by introducing the principle of sample weight polarization to improve the classical TrAdaBoost algorithm,which can optimize training samples in the source domain to some extent.Step two,to further narrow the distance between the source domain and the target domain,the large margin projectedtransductive support vector machine was applied to complete the transfer of the feature level,thus maximizing the transfer efficiency.Furthermore,the proposed method was applied to the BCI competition dataset(Dataset IIb data set)for offline test to verify the efficiency and universality of the algorithm.Finally,a transfer learning BCI system for intelligent driving assistant was designed,which included driving simulation platform and software and hardware system.And then,test the practicability,efficiency and universality of motion imagery classification algorithm based on hybrid transfer learning online.