LiDAR-based 3D object detection is a new technology that has emerged in recent years. However, in real scenes, the objects are usually scanned with incomplete shape points by LiDAR due to the occlusion, thus they suffer from insufficient structure information, which damages the detection performance. To solve this issue, we propose a Transformer-based Shape Enhancement Network TSENet for occluded object detection. Specifically, we first conduct a spherical voxelization operation to extract the incomplete point clouds of the occluded objects, which are sent to the shape enhancement module to predict the complete shape points. And then we adopt 3D sparse convolutions in the backbone to extract the enhanced voxel feature from the recovered shapes, which is fused with the voxel features from the original incomplete point clouds. Then we project the fused features into the bird's eye view to generate the initial proposals, from which we extract the RoI features and send them to the detection head to generate the refined detection boxes for the occluded objects. Extensive experiments on the KITTI and Waymo datasets demonstrate that the detection accuracy of our TSENet significantly outperforms other popular methods.
Paper
Full text
Object Detection of Occlusion Point Cloud based on Transformer
Semantic Scholar · Computer Science · 2023
Abstract
LiDAR-based 3D object detection is a new technology that has emerged in recent years. However, in real scenes, the objects are usually scanned with incomplete shape points by LiDAR due to the occlusion, thus they suffer from insufficient structure information, which damages the detection performance. To solve this issue, we propose a Transformer-based Shape Enhancement Network TSENet for occluded object detection. Specifically, we first conduct a spherical voxelization operation to extract the incomplete point clouds of the occluded objects, which are sent to the shape enhancement module to predict the complete shape points. And then we adopt 3D sparse convolutions in the backbone to extract the enhanced voxel feature from the recovered shapes, which is fused with the voxel features from the original incomplete point clouds. Then we project the fused features into the bird's eye view to generate the initial proposals, from which we extract the RoI features and send them to the detection head to generate the refined detection boxes for the occluded objects. Extensive experiments on the KITTI and Waymo datasets demonstrate that the detection accuracy of our TSENet significantly outperforms other popular methods.