Research On Object Detection Method Based On Lidar Point Cloud

In recent years,the application of deep learning in autonomous driving has become a hot research field.Object detection is one of the key technologies of autonomous driving systems.In autonomous driving scenarios,LiDAR and optical cameras are usually used to realize the perception of the surrounding environment of the vehicle.LiDAR can provide threedimensional spatial information,and optical cameras can provide the semantic information of the surrounding scenes,both of which play an irreplaceable role.Current deep learning methods have achieved excellent results in the field of two-dimensional image object detection,but the application of deep learning in lidar point cloud object detection still faces great challenges:(1)Existing algorithms have the problem of slow detection speed in large-scale scenarios,and cannot meet the requirements of real-time detection.(2)Due to the sparseness of the point cloud,the existing network has a poor detection effect on some long-distance or occluded targets.In response to the above problems,this paper studies the existing point cloud object detection algorithm based on deep learning,and proposes corresponding improvement measures on the basis of the existing algorithm to improve the detection speed of the network and the detection accuracy of difficult targets.The work done in this paper is as follows:(1)Research on the existing two-stage point cloud object detection algorithm based on point features.In view of the high time cost of current point cloud object detection algorithms based on point features in the point cloud down-sampling stage,a two-stage object detection algorithm based on random sampling and local feature aggregation is proposed on the basis of existing algorithms.In the process of point cloud feature extraction,random sampling is used to down sample the point cloud,which greatly reduces the time consumed by the network in the feature extraction stage.At the same time,in order to solve the problem that random sampling may lose important key points and cause the network detection accuracy to decrease,a local feature aggregation module is added.This module encodes the features of neighboring points and expands the receptive field of the center point to preserve the original point features,thereby solving the problem of information loss caused by random sampling.(2)Aiming at the problem of poor detection effect of existing algorithms on occluded and distant targets,a point cloud object detection algorithm based on regional pyramids is proposed.By processing RoI voxelization and constructing regional pyramids of different scales to capture a wider range of points of interest,the point cloud Transformer module is introduced to enhance the learning of the local features of the grid center point.In view of the sparseness of the point cloud,a ball query radius prediction module is added to the network,so that the model can adaptively adjust the query range according to the density of the point cloud.Finally,an optimized angle loss function is used to solve the angle confrontation problem of high overlap frames.(3)The effectiveness of the proposed algorithm is experimentally verified,and the performance of the model is evaluated and tested under the KITTI data set.At the same time,a corresponding ablation experiment was designed to verify the effectiveness of each module in the model.The experimental results show that the object detection algorithm based on random sampling and local feature aggregation guarantees the detection accuracy of the original algorithm,and the detection speed is increased by two times.Compared with the algorithm proposed in Chapter 3,the object detection algorithm based on regional pyramid improves the detection accuracy of medium and difficult targets by 3.77%and 1.70%respectively.