Research On Traffic Flow Detection And Light Control At Tunnel Entrance And Exit Based On Machine Vision ABSTRACT

The lighting system in highway tunnels is an important part of highway traffic engineering.The quality of tunnel lighting environment plays a very important role in driving safety.Traffic flow and speed are the main factors that affect the design of lighting system.On the premise that the speed is controllable in the actual operation of the tunnel,the requirements for lighting brightness under different traffic flow conditions vary greatly.If the lighting system always takes the value of brightness based on the value of the maximum traffic flow at the tunnel entrance and exit,it will lead to excessive energy consumption and a significant increase in construction and operation and maintenance costs.With the rapid development of computer vision technology,real-time target detection and tracking algorithm based on deep learning can solve this problem very well.This paper focuses on traffic flow detection and tunnel lighting control in practical application scenarios.The main research contents are as follows:(1)Firstly,the related knowledge of target detection algorithm and multi-target tracking algorithm is introduced,and an improved vehicle target detection model based on YOLOv5 s network is proposed.The C3 module in the original backbone network is replaced by the C2 f module.The C3 module in the original network performs the Add calculation after two convolution operations on the input to strengthen the transmission of feature information.Compared with the C3 module,the C2 f module has more branches and reduces one conv module,which can realize the lightweight of the model and make the model have faster detection speed.(2)The SPP module in the original backbone network is replaced by the SPPCSPC module.The SPP module achieves the fusion of local features and global features through convolution operation and pooling operation.SPPCSPC module has different pool core design compared with SPP module.It can divide the input into two segments and enter different branches,and finally concat the information flow output by all branches,which can further improve the detection accuracy of the model.(3)In order to solve the problems of mis-inspection and missed inspection of small and medium-sized target vehicles in practical application scenarios.Before introducing the attention mechanism SE module into the head output detection module of YOLOv5 s,the irrelevant features will be weakened,the recognition ability of the detection network for small vehicle targets will be enhanced,and training and testing will be conducted on the public data set Vis Drone2019.Taking the precision,recall and average precision of the model as the evaluation index,the experiment shows that the improved YOLOv5 algorithm has better detection performance for small target vehicles compared with the original network in the case of very dense traffic.(4)For the problem of vehicle number statistics,the multi-target tracking algorithm Deep SORT is adopted.The vehicle counting is realized by setting a virtual detection line in the video.Through video comparison,it is found that the improved YOLOv5-Deep SORT multi-target tracking algorithm has significantly improved the detection effect of small target vehicles under the same FPS,and there is no false detection phenomenon.(5)Finally,the basic structure of the tunnel lighting system is introduced,and a PID controller based on BP neural network is designed.Compared with the original PID control,the control strategy has better stability and energy saving,and the feasibility of the lighting control system is proved through matlab simulation experiments.