Abstract:

Most existing studies on driver’s mental load evaluation give the driver’s mental load classification label with the presence or absence of sub-tasks in driving scenes， but drivers may sometime fall into self-thinking， leading to the increase in their mental load in normal driving scenes. In addition， even the same driving sub-tasks may not have the same effects on the mental load of different drivers due to the discrepancy of individuals. As a result， there may exist some noisy labels in the data set collected by traditional methods， hence affecting the training results of the mental load evaluation models of drivers. In view of these problems， the method of confidence learning is adopted to process （detect and filter） the mental load classification labels of drivers in this paper. By using the processed labels， with electroencephalogram， electrocardiogram and skin electricity signal features as model inputs， the driver’s mental load models based on algorithms of support vector machine， random forest， K-near neighbor， decision tree， logic regression， and multi-layer perceptron are constructed to comparatively analyze the effects of noisy label processing on the enhancement of the performance of different models. The results show that after the noise label processing by using confidence learning， the performances of various driver’s mental load models constructed remarkably improve， among which support vector machine model achieves the best results in performance enhancement.

Key words: traffic safety, mental load, confidence learning, machine learning, noisy labels