Driver Fatigue And Distracted Behavior Detection Based On Deep Learning

In recent years,the number of traffic accidents in China has increased year by year.Through relevant data statistics and research,it is found that driver distraction behavior and fatigue driving are the main causes of traffic accidents.Therefore,timely monitoring and early warning of the above two major incentives is particularly important.Current algorithms are difficult to achieve accurate and rapid detection.This thesis will focus on these two issues,with the main research content and innovation points as follows:1)To solve the problem of fatigue driving detection,we first detect the opening and closing states of the human eyes and mouth,and propose a lightweight algorithm based on a single shot multi box detector(SSD).This algorithm transforms the problem of facial feature point localization in traditional methods into a deep learning target detection problem,thereby more effectively detecting fatigue states.Firstly,Ghostnet is used to replace the VGG16 backbone in the original SSD algorithm to extract the network,making the overall network parameters smaller;Subsequently,the integration of the Focus Loss function and the Soft Non Maximum Suppression(Soft NMS)algorithm further improved the detection effect.Experimental results show that compared with the original SSD algorithm,the Mean Average Precision(m AP)on the Yawning Detection Dataset(Yaw DD)improves 2.35%,reaching 96.73%;2)Combining the improved algorithm with fatigue state judgment indicators,a fatigue detection system based on deep learning is designed and implemented,which consists of three parts: data preprocessing,detection,and judgment.Experimental verification of the system on the Yaw DD dataset can effectively detect the fatigue status of drivers.3)An M-YOLO detection method is proposed for driver distraction behavior.Firstly,the backbone of the YOLOv4 network is replaced by a Mobilenetv1 lightweight network,and some standard convolutions in the feature extraction network are replaced by a depthwise separable convolution(DSC),which greatly reduces network parameters;Subsequently,the fusion of Squeeze and Excitation(SE)attention mechanisms can better extract feature information;Finally,an Exponential Linear Unit(ELU)is introduced to solve the problem of neuronal inactivation caused by the use of Rectified Linear Unit(Re LU).Experimental verification on a State Farm dataset shows that the proposed algorithm m AP achieves 94.11%,with a speed of 57 frames per second(FPS),which is2.39% higher than the original YOLOv4 algorithm,significantly improving the detection speed and accuracy of the algorithm.