Research On 3D Object Detection Based On Lidar In Autonomous Driving Vehicles

The emergence of autonomous vehicles provides new solutions for reducing the occurrence of road traffic accidents and alleviating traffic congestion.As an important sensor in the environmental perception module of autonomous driving technology,vehicle-mounted lidar can accurately obtain the position and size information of targets such as cars,pedestrians,and cyclists,providing a basis for decision-making,planning,and control modules.This thesis focuses on the research of target detection and recognition technology based on vehiclemounted lidar in autonomous driving perception technology,and the specific research content is as follows:(1)A combination filtering point cloud data preprocessing method is designed for lidar point cloud data,which has the characteristics of "denser near and sparser far" and contains noise points.Firstly,the pass-through filter is used to filter out sparse point cloud data that have no contribution to target detection and retain valuable point cloud data.Then,the designed staircase radius filter algorithm is used to filter out outliers in the point cloud data.The experimental results show that the proposed combination filtering algorithm can simplify and optimize point cloud data while retaining target information such as cars,pedestrians,and cyclists.(2)An Attentive-Pillars 3D object detection model based on Point Pillars with attention mechanism is proposed.This network proposes a point cloud feature enhancer,which adds four additional dimensions of point cloud information compared to Point Pillars,which can better characterize the distribution of point clouds in voxelized pillars.At the same time,a new point cloud feature extraction network is integrated,which combines the maximum pooling feature and the weighted feature with attention mechanism learned in each voxel,preserving the local information and context information of the point cloud in each pillar,providing richer point cloud information for subsequent networks and detection heads.In addition,a category confidence normalization scheme between multiple detection heads is proposed,which can effectively avoid the problem of false detection of objects detected by different detection heads at the same position.The experimental results on the KITTI 3D object dataset show that compared with Pointpillars,the proposed Attentive-Pillars can effectively improve the detection accuracy of three types of targets: cars,pedestrians,and cyclists.(3)A ROS-based data collection platform is first used to collect real-time point cloud data in a campus environment.The point cloud data collected in different scenes are preprocessed,and vehicles,pedestrians,and cyclists in the point cloud data are annotated to create a real vehicle point cloud dataset.Comparative experiments are conducted on the Pointpillars and the improved algorithm Attentive-Pillars based on the real vehicle dataset.The experimental results show that compared with Pointpillars,Attentive-Pillars can more accurately complete the detection and recognition tasks of three types of targets: cars,pedestrians,and cyclists,and has better scene transferability and strong generalization ability.In summary,the Attentive-Pillars algorithm proposed in this thesis improves the detection accuracy of three types of targets: cars,pedestrians,and cyclists.It can effectively complete target detection and recognition tasks,while also having a certain degree of robustness,meeting the requirements of perception in autonomous driving environments.