Vehicle Detection And Tracking For Behavior Recognition Based On Video Images

The continuous improvement of the scale and layout of the expressway network in China has greatly facilitated people’s travel.However,the severity of traffic accidents on expressways is much higher than that on ordinary roads due to the large flow,fast speed,high proportion of large vehicles and abnormal driving behaviors on expressways.Therefore,vehicle detection and tracking,timely monitoring of vehicle state,to ensure the safety of highway operation and operation efficiency is of great significance.The layout of monitoring equipment and the rapid development of computer vision technology in the whole process of the main highway sections provide data basis and technical support for the processing of surveillance video images.This paper uses highway surveillance video images to carry out research on vehicle detection,tracking and behavior recognition based on deep learning algorithm.The main research contents are as follows:(1)A vehicle image sample library based on highway surveillance video is constructed.Firstly,vehicles are divided into three categories according to their appearance characteristics: Car,Bus,Truck,proposed a semi-automatic labeling method to complete the rectangular frame labeling of 58,358 vehicle targets,and reasonably divided the data set according to the number of various vehicles to ensure the consistency of the number distribution of training set,verification set and test set,and complete the production of vehicle detection data set.After that,the Veri776 vehicle reidentification data set was selected to improve the recognition ability of the tracking algorithm for vehicle appearance features.Then,a total of 13,080 frames of the multi-angle tracking evaluation data set of the expressway was produced to evaluate the actual tracking effect.(2)An optimization model of vehicle detection based on YOLOV5 network is established.Firstly,a more efficient CG3 bottleneck structure was designed to optimize the network structure,and the P2 scale layer of the detection head was added.Then,ECA attention mechanism was introduced to improve the ability of detection model to extract key features.Then,a class-weighted Mosaic data enhancement method is proposed to solve the problem that the number of vehicles in the data set is not balanced.After the above optimization means,the m AP50 index and m AP50:95 index of the vehicle detection model in the test set were increased from 92.8% and 70% to 96.3%and 76%,respectively.(3)A multi-vehicle tracking optimization model based on Deep Sort algorithm is established.Firstly,in view of the problem that the appearance feature extraction network in Deep Sort has poor ability to identify vehicle features,Resnet18 is proposed to replace the original network,and triad loss constraint is introduced to the network output features.The m AP and m INP of the appearance feature extraction network in the Veri776 data set of vehicle reidentification reached 72.86% and 32.71%,respectively.Then,the parameters of the matching algorithm are adjusted to improve the running speed and tracking stability of the multi-vehicle tracking model according to the highway vehicle tracking scenarios.(4)A vehicle trajectory tracking and vehicle behavior recognition system based on highway surveillance video is designed and developed.Firstly,the method of obtaining road parameters and identifying the behaviors of vehicle stopping,reversing and illegal lane change is proposed.Then,the interface and background processing module of the system are designed and developed.The system can realize real-time automatic detection and tracking of vehicles at an average speed of 25 fps.At the same time,the number and average speed of vehicles in the current screen are counted,vehicle behavior is detected and the number of frames and vehicle ID are recorded.Finally,the tracking effect and real-time performance of the system are verified by using the high-speed monitoring video of Xinyuan.