Research On Driver Fatigue Detection Based On Deep Learning

Research shows that fatigue driving is one of the main causes of road traffic accidents.Therefore,the research of fatigue detection algorithm is of great significance for improving road traffic safety.At present,the high-performance embedded platform makes the use of deep learning for fatigue detection become reliable.This paper conducts an in-depth study on realtime fatigue detection technology on embedded platforms.The main research contents are as follows:(1)Driver face detection algorithm.Aiming at the real-time face detection problem of embedded platform,this paper designs a fast face detection algorithm for three-level cascade convolutional neural network.The network uses bounding box regression to reduce the number of cascaded networks,quickly generate face candidate frames by using a full convolutional network,and replace the traditional fully connected layer structure with a global average pooling layer,with little loss of precision.Reduce the volume of the network to improve the detection speed,and use the location reliability to improve the traditional non-maximum suppression algorithm.(2)Driver landmark detection and head pose estimation algorithm.First This paper designs a lightweight feature extraction unit to construct the skeleton network structure of landmark detection and head pose estimation.Based on the idea of multi-task learning,the landmark detection task and the head pose estimation task are jointly learned by sharing a network,which improves the accuracy of landmark detection under the condition of attitude change,and at the same it can obtain accurate head pose estimation without additional computational cost.(3)Driver fatigue recognition algorithm.Aiming at real-time driver fatigue recognition,this paper designs a fatigue recognition algorithm based on multi-feature spatiotemporal fatigue feature sequence.The algorithm extracts multiple facial fatigue features,including eye,mouth and head pose.Splice multiple facial fatigue features in time dimension to construct spatiotemporal fatigue feature sequences,then use cyclic neural network based on long and short time memory structure to recognition fatigue.The fatigue recognition method designed in this paper not only has good detection accuracy,but also can meet the needs of real-time fatigue recognition on embedded platforms.(4)Embedded platform fatigue detection software design and implementation.Based on the algorithm proposed in this paper,this paper developed a fatigue detection software on the Jetson TX2 embedded platform.The software interface is friendly and beautiful,and the function is complete.It can perform fatigue monitoring in real time and give alarm information in time.It can also record fatigue video for storage.