Fatigue State Detection of Locomotive Driver Based on Human Posture Features and Double-Stream Long Short-Term Memory Neural Network

Abstract

For the fatigue information conveyed by the upper part posture of the body, 13 feature points of the driver’s upper body in different states in this paper, such as sober fatigue, are collected on the high-speed railway simulator with a monocular camera, and then the sample data are obtained through correlation processing. Each sample in the feature set is continuous data extracted from continuous frames, including angle feature and relative position proportion feature. The training set is used to train the Double-Stream Long Short-Term Memory (LSTM) neural network, and the corresponding and Long Short-Term Memory neural network classifier is obtained. The trained Double-Stream Long Short-Term memory neural network classifier is used to classify the soberness, mild fatigue and severe fatigue of locomotive drivers. The model can achieve a good effect that the average classification accuracy of this model is close to 92.67%, and the F1 score is close to 92.71%, which verify the effectiveness and robustness of the method.