| With the emergence of artificial intelligence algorithms and the rapid development of fields such as autonomous driving and robotics,3D object detection algorithms have become a crucial component of perception systems.However,single-sensor 3D object detection methods suffer from certain limitations,such as the lack of necessary 3D depth information for image-based methods and the tendency for point cloud-based methods to suffer from noise interference,false detection,and missed detection,particularly for small objects.To address these issues,this paper proposes a multi-modal fusion approach that uses monocular cameras and lidar sensors to detect 3D objects,optimizes existing 3D object detection methods,and improves the accuracy and robustness of the algorithm.Specifically,this paper makes the following contributions and innovations:(1)To balance detection accuracy and efficiency in 3D object detection algorithms,this paper analyzes the PV-RCNN model and proposes a radar point cloud-based algorithm.This algorithm uses a new query method-voxel index query-to directly extract 3D voxel features of the key point neighborhood and convert them into bird’s-eye view(BEV)representation.Using a 2D backbone network and RPN,the algorithm generates 3D region proposals and optimizes the voxel feature extraction module to aggregate voxel features and extract features from3 D region proposal regions.Compared with direct detection on the bird’s-eye view,this method avoids the problem of missing 3D structural features.Experimental results show that the proposed method achieves higher accuracy rates and faster detection speeds on the KITTI 3D dataset.(2)To address the problem of point cloud sparsity in small-size target detection and long-distance target point cloud,this paper proposes a 3D target detection framework called VPG based on multi-modal fusion.The framework consists of two main modules: the instance segmentation module and the point cloud recognition module.The otherwise sparse 3D point cloud is augmented by generating dense 3D virtual points through a set of 2D detectors,which can be applied to other lidar-based 3D detectors.Based on this,the paper verifies and analyzes VPG using the voxel-indexed RCNN model on the Su Cences dataset,and VPG achieves improved detection accuracy for all object categories.(3)An ablation experiment is designed to verify the effectiveness of the proposed VPG model,using the VI-RCNN model as a benchmark.By varying the input,voxelization,or an optional second order,the improvement effect of each module of VPG on the point cloud algorithm is obtained.In addition,a comparative experiment is designed between YOLOv5 and voxel-indexed RCNN in 2D target detection to verify the effectiveness of YOLOv5 for small and long-distance targets.Finally,a comparative experiment is conducted using image resolution and detection target distance as variables.Through the comparative analysis of these experiments,the accuracy of the virtual point generation model in 3D target detection is found to be significantly improved compared with the benchmark model,particularly for small and long-distance targets. |
