Research On 3D Point Object Detection Technology For Unmanned Vehicles Based On Deep-Learning

LIDAR-based 3D point cloud target detection is one of the important technologies in driverless environment sensing by obtaining accurate position and orientation information of multiple targets such as cars,pedestrians and cyclists in space to provide important reference for subsequent decision and control modules.In the traditional method of 3D point cloud object detection,it usually requires manual design of complicated models,rules and parameters,which is not ideal to achieving the target of self-learning and self-design in replacing manual with a machine.While the 3D point cloud object detection based on deep learning is gaining popularity in academia and industry for its ability to perform the task of detecting targets in point cloud scenes in an almost end-to-end manner.This paper conducts research on 3D point cloud object detection technology for unmanned vehicle based on deep learning,and the specific work is as follows:First,an adaptive sampling feature extraction network ASCA-Pointpillars based on correlated point attention is proposed to address the problem that Pointpillars algorithm is not accurate enough to detect objects at a long distance from lidar.At first,a distance-based point cloud adaptive scale voxel sampling algorithm is proposed to address the loss of spatial information for distant objects in the process of point cloud voxelization sampling due to the use of uniform scale voxel sampling method.It can change the size of the sampled voxels according to the distribution characteristics of the point cloud,thus making up for the shortage of uniform scale voxel sampling.In the second place,a point cloud feature augmentation algorithm based on the correlative point attention module is proposed for the problem that the features after the feature extraction process of point cloud data within a single column voxel lack the information association between points in the feature encoding process.It enables point cloud features within the same column voxel to establish connections with each other,thus enhancing the feature learning capability of the algorithm for point clouds within column voxels and providing richer feature information for downstream classification and regression tasks.Experimental results on the KITTI 3D Object dataset demonstrate that the proposed ASCAPointpillars algorithm improves the average precision of car,pedestrian,and cyclist detection by 4.58%,3.25%,and 3.72%,respectively,compared with the original algorithm Pointpillars.Second,in order to alleviate the problem that the 3D point cloud object detection algorithm is weak for a smaller number of categories due to the unbalanced number of categories in the3 D point cloud dataset,a random sampling data augmentation algorithm RS-Aug is proposed considering the semantic information of the scene.The algorithm can randomly sample from the point cloud category sample database constructed from the existing point cloud data set during the model training process,and obtain the semantic information of the obstacles in the current frame of the point cloud scene through the optimized ground point segmentation algorithm,non-ground point clustering algorithm and enclosing frame fitting algorithm,and then combine the semantic information to reasonably add the sampled targets to the current point cloud scene,so as to balance the number of different categories of detection targets in the data set and improve the detection performance of the 3D point cloud object detection algorithm.According to the experimental results on the KITTI 3D Object 3D point cloud dataset,the 3D point cloud object detection algorithm Pointpillars,after incorporating the proposed point cloud data augmentation algorithm RS-Aug in this paper during the model training process,improves the average precision of detecting a relatively small number of pedestrians and cyclists relative to the car category by 1.33% and 1.27%,respectively.In addition,the average precision of ASCA-Pointpillars after combining the RS-Aug data augmentation algorithm proposed in this paper is improved by 1.59% and 2.46% in detecting pedestrian and cyclist categories,respectively.Third,the detection performance of the improved 3D point cloud object detection algorithm combining the RS-Aug and ASCA-Pointpillars algorithms proposed in this paper is verified based on the self-made real vehicle point cloud dataset.Firstly,the point cloud dataset is made based on the real vehicle platform by collecting,pre-processing and labeling the point cloud data to create a real vehicle point cloud dataset.Then,in order to better apply the deep learning model proposed in this paper to practice,the deep transfer learning method is introduced in the process of model training,and the knowledge learned from model training in large public data sets is transferred to the detection tasks in practical application scenarios with insufficient training data,and the detection effect of the comprehensive improved algorithm proposed in this paper is verified based on the self-made real vehicle point cloud dataset in the server and embedded device Nvidia Jetson AGX Xavier.The performance of the proposed algorithm for detecting cars,pedestrians and cyclists can be effectively strengthened by introducing deep transfer learning methods.In addition,the integrated improved model trained on the server is ported to the embedded device Xavier and accelerated with Tensor RT optimization to perform the real-time object detection task for the surrounding environment.