Research On Vehicle Target Recognition Algorithm Based On Lidar 3D Range Image

At present,the research of vehicle target recognition is mainly concentrated in the field of two-dimensional images.Due to the limited ability of two-dimensional images to express objects,the development of target recognition based on two-dimensional images also encountered bottlenecks.The 3D image of the scene obtained by lidar ranging is known as three-dimensional range image.In the field of image processing,the three-dimensional range image of lidar is also called 3D laser point cloud.Through the 3D laser point cloud,the 3D data information of the scene can be extracted.Through the 3d data information of the scene,the target object can be better perceived,so as to better identify the target.Therefore,the paper builds a vehicle target recognition system based on the 3D laser point cloud to identify the type of the vehicle target in complex scene,and at the same time studies the algorithms involved in each module of the system.In order to accomplish the objectives of the project,the thesis mainly studies the following contents:Firstly,the paper implements a system for acquiring point cloud data.In this system,the communication scheme between the client and the scanning device is designed.At the same time,the 3D range image data collected by the radar is converted into specifications.Secondly,the point cloud data preprocessing system is designed and implemented.This system contains the following four parts:coordinate system conversion,background point removal,noise point removal,scene segmentation.In this process,a new background point removal method based on the scanning characteristics of the lidar is proposed to remove complex background point clouds.Also,three algorithms for noise point removal are implemented.By comparing the efficiency and effectiveness of the three algorithms,the paper finally chose the statistical filter.Finally,two segmentation algorithms are implemented,and the region growth algorithm is determined to achieve the segmentation of scene point clouds.Thirdly,the paper recognizes and classifies vehicle targets in different scenarios.Vehicle target recognition is performed for scenes where the vehicle is not blocked,the vehicle is blocked by a single object,and the vehicle is blocked by multiple objects.The thesis compares the performance of 8 traditional recognition schemes obtained by combining four 3D feature extraction algorithms and two classification models.A new feature extraction algorithm combining the improved viewpoint feature and size feature is proposed.Based on this algorithm and K-nearest neighbor search model,vehicle target recognition in various scenarios is performed.The experimental results show that the recognition accuracy of this scheme is higher than the traditional scheme in different scenarios.Finally,the algorithm of each module of the system is encapsulated and the interactive design is added to obtain a highly available vehicle identification application software for vehicle target recognition.