| Fatigue driving is one of the behaviors that seriously endanger traffic safety.Drivers who are in a state of fatigue reduce or lose their ability to control the vehicle,and cannot take necessary measures for sudden traffic incidents,resulting in unavoidable traffic accidents.Therefore,finding an effective automatic fatigue monitoring method is a work of great significance in recent years.Among various fatigue indicators,physiological signal indicators can objectively and timely respond to changes in human fatigue levels.The method of combining it with machine learning has gradually become a research hotspot in the field of fatigue driving monitoring.This paper focuses on the application of sparse representation(SR)theory to fatigue driving monitoring based on physiological signals,and proposes a generalized minimax-concave based kernel sparse representation regression(GMC-KSRR)algorithm,which is extended from different perspectives to provide a novel and effective solution for automatic monitoring of fatigue driving.(1)First,this study proposes a fatigue recognition algorithm GMC-KSRR with excellent performance,strong generalization ability and simple model.Specifically,the algorithm maps training samples to a high-dimensional reproducing kernel hilbert space(RKHS)to seek linear separability of complex features.Then,in RKHS,the sparse coefficients of test samples are calculated through the SR process based on GMC penalty.Unlike the traditional l1-norm,the GMC penalty is unbiased and does not underestimate the high-value components of sparse coefficients.Finally,predict the regression result in the label subspace according to the obtained sparse coefficients.On the SEED-VIG sub-dataset of the SJTU Emotion EEG Dataset(SEED),the algorithm’s advantages in recognizing fatigue-related physiological signals and its practical significance in fatigue driving monitoring were verified.(2)Secondly,this paper performs a multimodal extension based on GMC-KSRR,and proposes a multimodal algorithm method based on the sparse center index(SCI),GMC-MKSRR.SCI estimates the data quality of each modalities and determines the weights of each modalities at the decision level to achieve robust multimodal fatigue driving monitoring by penalizing disturbed modalities.The robustness of the proposed multimodal algorithm has been verified on the SEED-VIG dataset,indicating the strong application potential of the algorithm in actual driving environments.(3)Finally,this paper investigates the adaptation of the truncated l1distance(TL1)kernel to fatigue-related physiological features and the performance improvement in combination with sparse representation algorithms.Compared with the traditional radial basis function(RBF),the TL1 kernel has good nonlinear adaptive capability for the problem of differentiating complex features.In this study,a fatigue recognition framework based on TL1 and SR theory is proposed.First,the linearly indistinguishable physiological features are mapped to the reproducing kernel Krein space(RKKS)by TL1 kernel.Then the obtained kernel matrix is transformed into a symmetric positive definite(SPD)matrix according to the eigenspectrum correction method.Finally,the final prediction results are obtained by a sparse representation based regression process.The monitoring performance of the proposed framework is experimentally validated on the SEED-VIG dataset.The experimental results show that the TL1 kernel outperforms the RBF kernel in terms of performance and stability of the kernel parameters,which validates the effectiveness of the TL1 kernel in identifying physiological signals and the validity of the proposed framework for fatigue level estimation. |
PDF Full Text Request