| The development of automatic driving technology has led to increased attention on 3D object detection,which is a critical component of 3D scene perception and understanding in this field.However,3D point cloud data,which is non-Euclidean in nature,presents unique challenges for traditional object detection methods that are based on grid data.As a result,researchers have begun to investigate methods based on graph neural networks(GNNs)to address these challenges.This thesis focuses on the use of graph neural networks for 3D point cloud object detection,with the goal of improving both accuracy and robustness.This is of great significance for 3D scene perception and understanding in the process of autonomous driving,and for promoting safer autonomous driving.The study explores methods to enhance the accuracy of object detection in 3D point clouds,while also examining ways to defend against attacks on graph neural networks.The main work content and designed algorithms are as follows:(1)A 3D object detection algorithm based on graph sampling and graph attention mechanism is proposed.Firstly,the feasibility of introducing attention mechanism when using graph to model point cloud data is analyzed from the perspective of entropy distribution,and the importance of different nodes is different after using graph structure to model point cloud data.However,previous research is based on the results directly,thinking that attention mechanism is a method that can improve the performance of feature extraction.Then,graph sampling is added to point cloud 3D object detection based on graph neural network,and a shared attention coefficient sampler is designed.For images with higher density than the reference model,the detection accuracy of small objects such as pedestrians and cyclists is improved without increasing the training cost.Then,by embedding multi-level attention mechanism,the detection accuracy of small and complex objects is improved while maintaining the model size.The 3D object detection of vehicles,pedestrians and cyclists is carried out on the KITTI data set,which effectively improves the detection accuracy of objects at the level of Moderate and Hard.(2)A robust 3D point cloud object detection algorithm based on graph purification is proposed.The graph purification and attention mechanism are combined to improve the robustness of graph neural network in 3D object detection.Firstly,the point cloud data is preprocessed by image purification filtering and noise reduction to reduce the disturbance noise brought by counter attack to the point cloud image;Then,we use the double attention mechanism to encapsulate the node-level attention and subgraph level attention in the large image before and after graph sampling,so as to reduce the weight of the characteristics of the attacked node in information aggregation.The graph structure attack simulation and defense experiments are carried out on the point cloud graph on the KITTI dataset,and the robustness of the designed model is verified.In this thesis,the proposed algorithm is trained and evaluated on the KITTI dataset,and the dataset collected in the actual scene is used for experimental verification,and compared with the classic 3D object detection algorithms.The experimental results further prove the effectiveness of the proposed method. |
