A New Algorithm Based on Cascaded Convolutional Neural Network for Driver Fatigue Detection

In order to improve the accuracy and speed of driver fatigue detection, in this paper we propose a new algorithm based on cascaded convolutional neural network for driver fatigue detection. First, we design a lightweight multi-task cascade convolutional neural network called LMTCNN for face detection and eyes key points positioning. Then, according to the key positions of the eyes, images of eyes can be captured. We convert the images to HSV color space and complete pupil positioning in H channel. Finally, ellipse fittings are used for obtaining pupil contour. The ratio of ellipse long axis and short axis can indicate the opened/closed state of eyes. We count the time of eyes closure during a period to calculate the PERCLOS and further judge the fatigue status of driver. Through comparative experiments, we demonstrate our algorithm outperforms mostly current driver fatigue detection approaches in both accuracy and speed.