3D OBJECT DETECTION NETWORK COMBINED WITH POINT CLOUD COMPLETION

The 3D object detection methods based on point cloud have been widely applied in the autonomous driving field. However, the shapes of the objects in the scene scanned by the LiDAR sensor are usually incomplete due to occlusion and distance, resulting in difficulty in locating these objects. To address this issue, we propose a novel two -stage 3D object detection network, which combines the point cloud completion task to recover the overall shapes of partial objects and then fuse the recovered structure information into the spatial knowledge via the attention mechanism for 3D box refinement. Concretely, we construct the voxel-based backbone with sparse convolutions to extract the voxel features containing spatial information for generating the raw proposals in stage 1. In stage 2, we first declare a Point-Voxel point cloud Completion (PVC) model, which incorporates the coarse -grained structure information captured from voxels with the fine-grained point -wise features for recovering complete shapes. And then we adopt the attention fusion mechanism to merge the shape features achieved by the PVC model into the voxel features extracted from stage 1, thus, enhancing the critical structural features of objects to promote detection accuracy. Extensive experiments on the KITTI and Waymo datasets demonstrate that the proposed method can effectively improve the accuracy of 3D object detection.