3D object detection has been used in many fields, such as virtual reality, automatic driving and target tracking. 3D object detection methods usually use point clouds as input, but point clouds are disordered and rotationally invariant. To solve this problem, voxel-based methods convert point clouds into voxels. However, the raw point clouds contain a large number of background points which are not relevant to the target or not helpful for subsequent detection, and voxel-based methods invariably feed them directly into the network for 3D object detection. Besides, the voxel-based approach is to model the point cloud data to generate a plenty of voxels with the same dimension. But random sampling during voxel partition will make voxel representation weaker and reduce the performance of classifier as well as box regressor. We, therefore, propose a plug-to-play module, which contains a shape-awared filter(SAF) and a semantic-ranked sampler(SRS). SAF can effectively remove some of the background points in the raw point clouds, to accelerate inference indirectly. And, SRS can enhance the expression of voxel feature by retaining the points of high confidence. Finally, we remove previous orientation classifier and propose a new loss method named ADIoU loss to improve the orientation estimation performance. Experiments on the KITTI car detection bench-mark demonstrate that our method shows faster inference speed and higher detection accuracy compared with SOTA methods.
展开▼
