Multiple nonlinear features fusion based driving fatigue detection

Various studies showed driving fatigue is one import factor that caused traffic accidents, so it's of great significance to seek an effective detection method for the safety of life and property. Electroencephalogram (EEG) is regarded as the “gold standard” in fatigue detection. However, due to its non-linear, non-stationary and vulnerable to the environmental noise, it's still difficult to achieve an accurate and reliable recognition result. In this paper, we propose a driving fatigue detection method based on multiple nonlinear features fusion strategy. Firstly, six widely used nonlinear features for EEG signals are included for feature extraction. Second, those extracted features are further fused and classified with the multiple kernel learning (MKL) based SVM. Finally, we take the full use of automatic feature extraction and classification ability of deep neural network to analyze the critical EEG channels based on the optimal single nonlinear feature of spectral entropy. Experimental results show that the single nonlinear feature based model achieved the best recognition accuracy of 81.33% with spectral entropy. The proposed multiple nonlinear features fusion method of MKL obtained the best accuracy of 84.37% with four types of entropy features. Two typical feature extraction methods of autoregressive and power spectrum density are used as comparative work to illustrate the effectiveness of the established dataset and the proposed method. The selected two groups of key electrodes are verified through experiments.