Research On Driver Behavior Detection Method Based On Convolutional Neural Network

With the improvement of people’s material living standards,the fundamentals of transportation and the establishment of public service facilities in China is increasingly sound,and the number of domestic cars has increased rapidly,which greatly increases the probability of traffic accidents.According to statistics,80% of traffic accidents are mainly caused by drivers who are not attentive during driving.In order to ensure the safe driving behavior of drivers,it is essential to implement safe and intelligent monitoring of moving vehicles.In this context,after analyzing and comparing the existing methods for dangerous driving behavior monitoring in China and abroad,this paper studies the driving behavior detection method from the driver’s head and hand area based on the convolutional neural network.The specific work is as follows:(1)Research on the driver behavior detection method of YOLOV3 based on CBAM : YOLOv3 uses the YOLO layer to merge the feature maps extracted by the Darknet53 framework to detect,and the YOLO layer affects the detection results greatly.The CBAM module is a simple and efficient forward convolutional neural network attention module,which can improve the feature expression of key regions.Therefore,the fusion attention module CBAM was introduced into the YOLO layer to get the improved C-YOLOV3 structure,so that the algorithm focused on five common driving behaviors in the detection process: "correct driving,talking with right hand,talking with left hand,drinking water,and driving without concentration".(2)Research on the handheld call behavior detection method of YOLOv3 based on feature fusion: Study the driver’s handheld calling behavior further,and propose an improved YOLOv3 network for the small size like mobile phone target detection,low resolution,and inconspicuous features.model.In this paper,K-Means++ clustering is used to recalculate the anchor frame to obtain an anchor frame of the appropriate scale to achieve the effect of accelerating the convergence speed of the algorithm.By improving the FPN structure in the original YOLO network,the feature map after the feature fusion is selected.The method of sampling in the opposite direction re-fuses the features to achieve the purpose of improving detection accuracy.(3)In view of the problem that the driver’s data set provided by State Farm does not have the labeling information,this paper uses the labeling tool Label-Image to select numbers of images from the data set to manually label the target area,and the target in the training data set The area is marked according to the format applicable to YOLO,and a new marking file is generated.This paper conducted corresponding experiments on the labeled data set.The results show that the two kind of improved YOLOv3 network models studied in this paper can improve the accuracy of detection more effectively.Among them,the driver behavior detection method of YOLOv3 based on CBAM reduced the interference of irrelevant background factors and the detection accuracy has been significantly improved.In addition,the hand-held call behavior detection method of YOLOv3 based on feature fusion completes the detection of the driver’s hand-held phone,and improves the accuracy of the detection without affecting the detection speed,which proves the effectiveness of the method.