Research On 3D Point Cloud Object Detection Algorithm For Autonomous Driving

Autonomous vehicles require perception of the surrounding 3D environment,and3 D object detection algorithms based on Li DAR point clouds are the foundation for obtaining accurate object position information in 3D space.They play an important guiding role in subsequent decision-making,planning,and motion control.Therefore,the development of accurate and efficient 3D point cloud object detection algorithms is of great practical significance for the development of autonomous driving.This thesis aims to study a 3D point cloud object detection algorithm based on multi-view fusion to address the problems in autonomous driving scenarios.Firstly,this thesis deeply investigates various representations of point cloud and summarizes the methods of multi-view feature extraction,interaction,and fusion,forming a unified paradigm of multi-view fusion.Based on this paradigm,a 3D point cloud object detection algorithm suitable for autonomous driving scenarios is designed and implemented.Specifically,this thesis focuses on the following aspects:(1)Aiming at the problem of multi-view fusion in point clouds,this thesis conducts a thorough analysis and summary of feature extraction,interaction,and fusion methods for point cloud multi-views,and constructs a theoretical framework for interaction and fusion of arbitrary number of views.The framework consists of three parts.Firstly,targeted feature extraction networks are adopted according to the structural characteristics of different point cloud views.Then,based on the key-value storage method of hash mapping,feature interaction between multi-views is realized with the original point view as the bridge.Finally,an attention mechanism is used to generate feature weighting vectors,realizing efficient fusion of multi-views.(2)Aiming at the problem of sparse object detection in point clouds in autonomous driving scenarios,a point-ring view fusion-based anchor-free detection algorithm is designed.In autonomous driving scenarios,a large number of sparse objects are the pain point of 3D point cloud anchor-free detection algorithms.This thesis introduces a ring view branch based on point anchor-free detection algorithms,and improves its detection performance on sparse objects through efficient fusion of point-ring views.At the same time,a down-sampling strategy for point-ring view segmentation is designed to obtain more foreground points as the initial center point of the anchor-free detection head.Moreover,an attention-weighted fusion module is designed,generating a weight vector based on channel attention mechanism to retain important sparse object information in feature channels,to fully and effectively fuse point and ring views.(3)Aiming at the problem of 3D point cloud object detection in large-scale autonomous driving scenarios,a two-stage anchor-based detection algorithm based on point-voxel-ring view fusion is designed.3D point cloud object detection in large-scale autonomous driving scenarios is a highly challenging task.Compared with other algorithms,anchor-based algorithms have good detection performance but still cannot meet the requirements of autonomous driving scenarios.This thesis introduces a voxelbased anchor detection algorithm,and adds point and ring view branches,reducing its dependence on voxel resolution through efficient fusion of multi-views,while improving its performance on complex objects.Additionally,a self-attention mechanism feature fusion module is designed,adaptively learning weights based on the correlation between features,significantly improving the expression ability of multi-view features.