Research On Vehicle Recognition Method Based On Point Cloud Data

With the continuous development of the automobile field and the progress of technology,intelligent vehicles and manless driving and other technologies have become the current research hotspots.Vehicle recognition as a key technology in the field of manless driving,has rapidly become a hot research direction.Traditional vehicle identification methods mostly rely on image data,and image data as the projection of 3D space.The image data lack of structural description of 3D physical space.Lidar can quickly collect the surface coordinates of space objects,But the data of point cloud is sparse and disordered due to object occlusion and radar linear scanning principle,which makes feature extraction and target vehicle recognition based on point cloud data very difficulty.Aiming at the above problems,this paper studies the vehicle target recognition algorithm in road scene based on point cloud data.We have improved the deficiencies in the existing point cloud vehicle identification methods,and can identify vehicles in the environment quickly and effectively,It provides powerful technical support for unmanned driving and intelligent robots.The main research contents and improvements are as follows:(1)We study two feature coding methods of point cloud data,voxel feature coding and bird-eyes view feature coding.The bird-eyes view features do not have overlapping targets,which can effectively avoid the negative effects caused by object occlusion.By improving the point cloud feature pyramid network and fusing the features of two different coding methods and features of different scales,we proposed a vehicle identification method based on the fusion of dual feature coding.This method can effectively solve the identification error caused by the object shielding.(2)Study the transformation matrix between point cloud and image,we proposed a point cloud vehicle recognition method based on image mapping in combination with mature two-dimensional recognition methods.In this paper,vehicles in the image are identified by YOLO algorithm,and the range of target point cloud is reduced by using the detection box and the transformation matrix between image and point cloud.Moreover,the K-means clustering algorithm is improved to optimize the vehicle recognition results.This method can quickly and effectively identify vehicles in the point cloud scene,and has better accuracy compared with the vehicle identification method based on pure point cloud data.Through experiments on Kitti dataset,The effectiveness of the vehicle recognition method based on dual-feature coding fusion was verified,and the identification accuracy of the three vehicle targets with different occlusion degrees reached 84.50%,79.83% and72.19%,respectively,which reduced the impact of object occlusion on vehicle recognition.And it shows that the point cloud vehicle recognition method based on image mapping is superior.The recognition rates of simple,medium and difficult models in Kitti data set reach 92.34%,88.03% and 81.38%,respectively.