PVSA : A general and elegant sampling algorithm for Voxel-based 3D object detection

Perceiving the environment is vital for autonomous vehicles as it serves as the foundation for decision making and path planning. LiDAR is a widely employed sensor, which produces a voluminous and sparsely populated point cloud. For voxel-based 3D object detection methods, the initial step involves the division of the raw point cloud into voxels, the process known as voxelization. Nevertheless, once the number of point clouds contained within a voxel reaches the certain threshold, the allocation of additional point clouds to that voxel ceases. This leads to a greater degree of information loss. Scholars primarily focus on the subsequent stages following voxelization, such as feature extraction and utilization. We first focus on the sampling issue during the voxelization. In the paper, we propose a general and elegant Points in Voxel Sampling Algorithm module named PVSA. During the voxelization, the assignment of all points into their respective voxels continues even after the maximum number of points in a voxel has been reached. For voxels in which the number of internal point clouds exceeds the certain threshold, the farthest distance sampling method is utilized as it ensures a genuine and uniform distribution of the point cloud within the voxel. We conducted an evaluation of the proposed module using the Kitti dataset. Experimental findings suggest that the incorporation of the PVSA module enhances the object detection capabilities of the voxel-based model, particularly in the identification of samll targets like pedestrians. The incorporation of PVSA modules significantly enhances Pillarnet's capacity to recognize pedestrians, resulting in a 46.2% pt improvement in performance at a distance of 20 meters. On average, there is an enhancement of 1.43% pt.