Driver fatigue detection based on comprehensive facial features and gated recurrent unit

In recent years, deep learning-based driver fatigue detection algorithms have been increasingly used. However, traditional fatigue detection algorithms cannot effectively correlate contextual information of image frames. They perform better in individual image frames. Also, the accuracy and robustness of these algorithms are limited because they only consider particular frames. Therefore, a fatigue detection method based on integrated facial features and Gate Recurrent Unit (GRU) judgment neural network is proposed in this paper. We use a neural network including a GRU layer to efficiently distinguish the contextual information present in multiple image frames arranged in chronological order. Besides, we designed a multi-task convolutional neural network (MTCNN) model to extract comprehensive facial features. After obtaining the facial feature points' positions, we can calculate the aspect ratio between the upper and lower eyelids, the upper and lower lips, and the eyebrows to the chin. In addition to the above three features, we can also obtain the subject's three head pose angles by comparing the facial features with the typical 3D face model. Finally, we input the change curves of 6 features in 20 consecutive frames into the judgment network to learn the change rule and create a judgment network. After the learning is completed, the judgment network model will judge the six feature curves in the newly input 20 frames in real-time and output the driver's fatigue status. This fatigue detection method can take a real-time detection at 55 FPS on the workstation platform (TensorFlow 2.3.0, RTX2070s). On the Nvidia Jetson Xavier AGX embedded platform (TensorFlow lite, ARM 8-cores CPU), the method can take a real-time detection at 26 FPS. The accuracy of this fatigue detection method can reach 97.47%.