Research On Automatic Lidar Target Detection Technology For Open Scene

With the application of self-driving cars in the field of transportation,the research on the key technologies of autonomous driving is also getting deeper.As an important part of autonomous driving technology,environmental perception is the “eye” of intelligent vehicles.At present,the research based on the three-dimensional lidar target detection system has achieved many results,but the target detection in open scenes still has the following problems:First,the perception environment is relatively simple,mainly based on the scene with sparsely distributed targets,and multi-type targets interact There are few researches on target detection in the scene;the second is that the data volume of the 3D lidar point cloud is extremely large,and the computing power required for directly applying the original point cloud to the target detection is high,and a reasonable streamlined method needs to be designed;Uniformity leads to poor adaptability of traditional density-based clustering algorithms.In view of the above problems,this paper is oriented to open traffic scenarios,focusing on four key technologies:point cloud simplification,ground point filtering,dynamic interest area division,and target clustering.First,starting from the detection requirements and environmental characteristics,the overall structure of the lidar target detection system is designed,and a three-layer four-module detection architecture is proposed.Among them,three layers are input layer,perception layer and output layer;four modules are point cloud simplification,background separation,region segmentation and target object extraction.The architecture gradually strips out the effective target point cloud through the first three modules.Experimental results show that the simplified method can reduce the number of invalid point clouds while ensuring the number of point clouds required for target extraction.The number of point clouds after the reduction only accounts for 12.88% of the original point clouds.Then,for the problem of poor clustering effect of DBSCAN algorithm on uneven density data,on the basis of studying the spatial distribution law of lidar point cloud,a point cloud density characteristic model is established,and an adaptive ellipse clustering method is proposed.The clustering threshold of this method changes dynamically according to the density of the point cloud,so as to solve the problem of the point cloud being dense and distant due to the scanning characteristics of the lidar,and ensure the accuracy of the target clustering.And further optimize the key parameters of the adaptive clustering algorithm through simulation experiments.In order to verify the performance of the proposed target detection system,a real vehicle verification platform was built and experimental verification was conducted in the campus environment.The results show that the lidar target detection system designed in this paper can quickly and accurately extract the target in a complex traffic environment,the target detection accuracy rate reaches 85.66%,and finally output roadside information and target state parameters.