| With the increase of cars,traffic accidents are easy to happen on the road,which will have a significant impact on people’s life safety,and will inevitably cause a lot of economic losses.According to statistics,a large part of traffic accidents are caused by improper operation of drivers,such as playing with mobile phones,making phone calls,drinking water,leaving the steering wheel with both hands while driving,which are all the reasons for frequent traffic accidents.In order to ensure the safety of people’s lives and property,the research of dangerous driving detection algorithm becomes more and more important.Based on the analysis of the research objectives,this paper uses the improved YOLOv5 s algorithm to study the detection of six unsafe driving behaviors: left hand playing with mobile phone,right hand playing with mobile phone,drinking water,left hand playing with phone,right hand playing with phone,and both hands leaving the steering wheel.The main work and achievements of this paper are as follows:In view of the problem that too many YOLOv5 s parameters lead to poor performance on mobile devices with small computing power,the model lightweight design idea is adopted to replace the backbone network CSP-Darknet53 of YOLOv5 s with Mobile Net-v3 structure.In addition,too many parameters in some layers of Mobile Net-v3 may lead to longer detection time and dimension reduction in SENet may bring side effects to channel attention mechanism.Conduct ablation experiments on Mobile Net-v3,and use the ultra-lightweight attention mechanism ECA net to add to the back of the ablated network structure in the ablation experiment to offset the accuracy loss caused by the ablation of the original network structure.The YOLO-Mobile Net+ECA-10 network model formed by the above operations greatly reduces the amount of network parameters and has good detection effect.Explore the impact of replacing C3 module with C3 Ghost on the head part of the network model YOLO-Mobile Net+ECA-10 obtained above,and solve the problem of gradient divergence in C3 Ghost by using Shortcut for jump connection.It is concluded that using C3 Ghost to improve the C3 module in the original network head can improve the accuracy,increase the detection speed by nearly 2 FPS,and reduce the number of parameters by one third.When dangerous driving occurs on the side far away from the camera,the driver’s hand movements on the phone are blocked by his head,which makes it difficult to identify effectively,Add the SPPF pyramid pooling structure at the back of the Mobile Net＿Block to convert the input feature maps of different sizes into fixed-size features.Improve the receptive field of the network,and optimize the non-maximum inhibition method of the obtained network.It is concluded that using the DIOU non-maximum inhibition strategy combined with the model of YO-MO+ECA10+C3Ghost+Shortcut network can improve the detection effect of the driver’s phone call under occlusion.Through the experiment on the data set of driving behaviors marked in this paper,the experimental results show that the final detection accuracy of dangerous driving behaviors of drivers is 97.7%,the detection speed is 66.2 FPS,the recall rate is 98.1%,and the model size is 5.4M.The model meets the requirements of fast and accurate detection of six dangerous driving behaviors of drivers. |
