Research On Three Dimensional Object Detection Based On Attention Fusion And Cross-module

The three-dimensional object detection technology in environmental perception systems is an important part of the autonomous driving field,and the high accuracy of point cloud data makes it an important research direction for three-dimensional object detection algorithms.However,current three-dimensional object detection algorithms based on point cloud cannot fully use the rich feature information of point clouds,the extraction ability of target object features is not ideal,and the perceptual detection performance of objects at different scales is also unbalanced.In this paper,we propose a three-dimensional object detection algorithm based on attention fusion and a three-dimensional object detection algorithm based on cross-module attention to address the above problems,so that the network can make full use of the high-precision information collection capability of lidar to build an efficient and accurate environment perception system,and conduct principle analysis,algorithm verification and experimental evaluation.The main research of this paper is as follows:(1)Research on 3D object detection algorithm based on attention fusion.To address the problem that lidar point cloud features cannot be adequately represented in object detection,this paper proposes an attention fusion 3D object detection network based on the SECOND object detection network.The network assigns the target point cloud to a voxel grid and converts the features to a 2D representation by a 3D feature extraction network,uses attention fusion for further feature learning in the spatial feature extraction part,and finally adjusts the obtained results by a loss function to obtain the location of the refined bounding box.The performance evaluation on the KITTI dataset shows that the network enhances the ability to extract feature information for distant targets,small-scale targets,and the presence of occluded targets,and is able to reduce the impact of background features on the detection accuracy of the algorithm.(2)Research on 3D object detection algorithm based on cross-module attention.To address the problem of unbalanced detection performance of targets at different scales,this paper proposes a 3D object detection network with cross-module attention.The cross-module attention approach is used to improve the 3D feature extraction network by transferring the attention information from the front and back attention modules,so as to give full play to the attention modules and further improve the perception and detection performance of the network and reduce false and missed detections.And the two-dimensional feature extraction backbone network part is improved using a multi-scale feature fusion method,which enables the network to better understand and represent the shape and structure information of the target through the deep fusion of spatial and semantic features in the feature mapping,thus improving the network feature expression capability.The performance evaluation on the KITTI dataset and a series of comparative ablation experiments show that the cross-module attention-linking approach with the multi-scale feature fusion approach can effectively improve the detection accuracy of the algorithm,and the inference speed of the algorithm reaches 33 frames per second,which is also competitive in terms of real-time performance.