Research On Obstacle Detection Of 3D Lidar By Using A Multi-model Fusion Method Under Challenging Conditions

In recent years,with the rising development of artificial intelligence technology,the autopilot technology based on intelligent networked transportation equipment is developing rapidly,but automatic driving technology is faced with many technical problems,one of which is the poor perception performance,which is difficult to ensure the safety of automatic driving.Many research work related to obstacle detection based on LIDAR point cloud mostly focus on using a single traditional detection model or deep learning detection model to detect obstacles,without considering the complex environment a single model will encounter like the large road fluctuations and traffic participants,which leads to the robust deficiency of the perception system.Therefore,this paper takes the perception module in automatic driving system as the research object,combined with the multi-model fusion strategy,focusing on studying the obstacle detection of LIDAR point cloud in complex environment,which will lay the foundation for the robustness of the perception module.The main work of this paper is as follows:First of all,in terms of the traditional detection methods of LIDAR point cloud,it is difficult to achieve the balance between real-time and accuracy in complex environment,and the point cloud segmentation and clustering algorithms rarely meet the characteristics of LIDAR point cloud data.Considering the uneven ground in complex environment,the method of ground segmentation based on depth map could improve the accuracy of point cloud segmentation algorithm.At the same time,according to the characteristics of laser point cloud data,that is,the farther,the sparser,the vertical resolution and the unbalanced horizontal resolution,the curved voxel is introduced to divide the non-ground point cloud,and then the connected area analysis of the 27 neighboring curved voxels is carried out to improve the cluster.By optimizing the rectangular frame estimation algorithm based on the L-shape point cloud cluster,the real-time performance of the algorithm will be improved.Secondly,as the existing deep learning point cloud detection method using trained network to inference,it is difficult to meet the real-time capability of the perception system.Thus,the deep learning detection network Point Pillars is trained,and the network and weights obtained after training are converted into ONNX file format.In addition,these ONNX files will be parsed based on Tensor RT to optimize the network structure and parameter accuracy as a way to establish an inference engine.It is expected that the real-time performance will be optimized based on the Tensor RT.Since the single-model detection sensing system is not robust,a multi-model fusion detection system is designed in this paper.Traditional detection method and deep learning method are used to obstacle detect by inputting the same frame of LIDAR point cloud data,the detection results are merged to improve the safety redundancy of the perception system.Finally,based on the LIDAR point cloud data collected by a vehicle platform,the verification is carried out,and multiple sets of lidar point cloud data are received through the robot operating system to verify the performance of LIDAR point cloud obstacle detection algorithm based on multi-model fusion in complex environments.The experimental results indicate that the multi-model fusion strategy proposed in this paper plays an important role.The obstacle detection algorithm based on multi-model fusion significantly improves the security redundancy of the perception system in challenging conditions.