Driver Drowsiness Detection based on 3D Convolution Neural Network with Optimized Window Size

Since driver drowsiness is one of the most significant causes of traffic accidents, it has been a key issue to detect driver's drowsiness. This paper introduces a driver drowsiness detection based on an optimized 3D convolutional network with only facial features. Our main contributions are that (1) we use only partial information from face images to detect driver drowsiness and (2) propose the best window size for drowsiness detection over video frames that allows us to take appropriate and enough contextual information into consideration to predict driver's drowsiness. In order to figure out what duration of the input video frames gives the best results for a 3D convolutional network, we conduct extensive experiments on the window sizes ranging from 5 to 128 frames together with several window overlapping options. Our approach has achieved an accuracy of 94.74% on the National Tsinghua University Driver Drowsiness Detection (NTHU-DDD) dataset, outperforming other 3D convolutional network-based state-of-art approaches.