Research On Multi Target Detection And Tracking Of Driverless Vehicle

Autonomous driving is an important part of the current technological development.At present,completely unmanned driving has been realized in specific environments such as parks,ports,and mining areas.Environmental perception is the cornerstone of unmanned driving technology.It provides necessary information for lower-level decision-making and planning,and is a necessary prerequisite for safe driving of unmanned vehicles.Lidar,as an important sensor in vehicle perception,can obtain three-dimensional point clouds in real time and efficiently.Obstacle detection and tracking research based on point clouds is one of the key issues in smart car dynamic environment perception.This article uses VLP-16 radar as the main sensor to perceive the surrounding environment of unmanned vehicles.The content mainly includes Lidar external parameter calibration,background point cloud removal,driving area extraction,obstacle detection and tracking research,the specific content is as follows:(1)Lidar external parameter calibration.In this paper,the carton calibration algorithm is used to obtain the coordinates of the "corner points" of the carton in the lidar and car body coordinate system,and to solve the rotation and translation matrix.Finally,the radar coordinate point cloud is converted to the car body coordinate system.(2)Remove the background point cloud.First use the RANSAC and GPF algorithms,and use a plane model to represent the ground,and the obtained ground cannot be used for slope recognition.Therefore,this paper proposes an algorithm to effectively identify the slope: firstly divide the front of the vehicle into two areas,use the GPF algorithm idea to select the lowest part of the representative point of the harness in each area,and then find the effective harness in the two planes.Use the RANSAC algorithm to fit the effective beam points in each plane,and then judge the included angle of the two fitted planes again.If the included angle is greater than the threshold,the front and back point clouds at the intersection of the two planes will be fitted again to complete For road recognition,if the included angle is less than the threshold,there is no need to fit it again.Finally,the validity and real-time performance of the algorithm are verified by the actual vehicle.(3)Available area extraction.In view of the sudden change of the height of the roadside point relative to the road surface point in the pavement point cloud,the primary selected roadside point is obtained by using the height difference of adjacent points.There is a range of the height of the roadside.The primary selection of roadside points is to select all the points with large height differences,and the height difference in a region is not obtained.Moreover,the road surface point cloud includes not only the road surface and the roadside,but also other surrounding areas.Therefore,the grid map is used to select the points that meet the height range of the roadside,and to remove the preselected roadside points that are too low or too high.The improved DBSCAN clustering algorithm is used to cluster the roadside points,and the isolated noise points are proposed.Finally,RANSAC is used to fit the roadside line to obtain the usable area.(4)Obstacle detection in drivable area.According to the characteristics of DBSCAN clustering algorithm,an adaptive threshold clustering algorithm is proposed,which determines the radius of clustering and the number of minimum sample points according to different distances.The square box model is used to characterize the geometric outline of obstacles after clustering.Finally,create a sample data set.Use the extracted features to train the SVM classifier to complete the target classification.(5)Obstacle tracking prediction.After completing the target detection,it is necessary to track the obstacles,use the correlation matrix method to correlate the prediction of the previous frame with the state of the current frame,and use the Hungarian algorithm to obtain the largest matching pair.In order to obtain accurate predictions and solve the problem of large fluctuations in the continuous frame rate of the lidar,the Kalman filter algorithm is used to optimally estimate the velocity and acceleration to obtain an accurate predicted position,thereby improving the matching accuracy.The accuracy of the estimation is verified by the actual vehicle,and the accuracy of tracking is improved.This article uses C++ programming to process algorithms and receive and send messages on the ROS robot operating system.