Research On Vehicle Detection And Tracking Method In Expressway Environment

In recent years,intelligent driving technology has attracted extensive and warm attention worldwide,and perception technology plays an important role in it.Highway is a very typical driving scene with high speed and monotonous scene.It is of great practical significance to solve the problem of driverless driving in the highway environment.Unmanned driving in highway environment needs to detect more distant dynamic vehicle targets through perception module,and obtain more accurate position and speed information of surrounding vehicles.Therefore,in view of the above environmental awareness requirements,the fusion of lidar and millimeter-wave radar is used to get accurate dynamic vehicle position and attitude information near the point cloud data of lidar,and the vehicle detection and tracking method is realized in the highway environment.Firstly,in order to obtain all obstacle information in the environment,it is necessary to segment the point cloud of lidar into road surface.On the one hand,it obtains the information of static obstacles in the environment,on the other hand,it reduces the difficulty of subsequent vehicle detection.Highway highly structured features make stable pavement removal more feasible,but there are still some difficulties.Because of the existence of temporary inspection and construction section,the method of combining high-precision map is not reliable at any time;the commonly used obstacle detection method based on search has higher recall rate and lower accuracy rate for road surface detection,and the lack of points is disadvantageous to the follow-up vehicle detection module;the method based on fitting,such as RANSAC,is slow,but also has certain limitations.Therefore,this paper proposes a road surface fitting method based on radial gradient feature points to segment road surface by combining search and fitting methods.It can not only get a more robust effect in various scenarios of expressway,but also achieve a relatively fast computing speed.Secondly,this paper achieves fast and relatively accurate vehicle detection through a one-stage 3D detection network based on efficient CNN.Compared with other methods,object detection based on in-depth learning has greater advantages in accuracy,but the more accurate two-stage detection network often has slower detection speed and can not meet the needs of Expressway perception speed.In this paper,Focal Loss reduces the imbalance of difficult and easy samples in 3D detection,and improves the accuracy of one-stage network.Thus,a one-stage network with relatively fast detection and two-stage network accuracy can be realized.Finally,the post-fusion framework of detection results of lidar and millimeter-wave radar is realized by state estimation based on Unscented Kalman filter and data association based on greedy algorithm.Aiming at the difference of fusion between point model and box model,a method of matching and updating bounding boxes based on feature points is proposed.By selecting feature points based on conditional probability,the feature points between detection results and predicted bounding boxes are matched.Thus,the center of the detected target is corrected to improve the accuracy of tracking results.In summary,this paper proposes a vehicle detection and tracking method based on highway scene,which improves the speed and accuracy of vehicle detection.