PV-SSD: A Multi-Modal Point Cloud 3D Object Detector Based on Projection Features and Voxel Features

Abstract

3D object detection using LiDAR is critical for autonomous driving. However, the point cloud data in autonomous driving scenarios is sparse. Converting the sparse point cloud into regular data representations (voxels or projection) often leads to information loss due to downsampling or excessive compression of feature information. This kind of information loss will adversely affect detection accuracy, especially for objects with fewer reflective points like cyclists. This paper proposes a multi-modal point cloud 3D object detector based on projection features and voxel features, which consists of two branches. One, called the voxel branch, is used to extract fine-grained local features. Another, called the projection branch, is used to extract projection features from a bird's-eye view and focus on the correlation of local features in the voxel branch. By feeding voxel features into the projection branch, we can compensate for the information loss in the projection branch while focusing on the correlation between neighboring local features in the voxel features. To achieve comprehensive feature fusion of voxel features and projection features, we propose a multi-modal feature fusion module (MSSFA). To further mitigate the loss of crucial features caused by downsampling, we propose a voxel feature extraction method (VR-VFE), which samples feature points based on their importance for the detection task. To validate the effectiveness of our method, we tested it on the KITTI dataset and ONCE dataset. The experimental results show that our method has achieved significant improvement in the detection accuracy of objects with fewer reflection points like cyclists.