APR-OIS: A Near-Sensor Point Cloud Pre-Processing Accelerator on FPGA
Yiming Gao, Herman X. Lam · 2025
Raw point clouds generated by 3D sensors (such as Li-DARs, and RGB-D cameras) are typically large in scale, containing an enormous number of points. For this reason, the raw point clouds usually require an expensive down-sampling phase to reduce the number of points while preserving their spatial structure, before being used as input to most point cloud applications. In recent years, the Octree spatial indexing method [2] for the point clouds has been used to accelerate the down-sampling phase. Based on the existing Octree-based down-sampling method and spatial proximity function of Octree nodes, we propose an approximation method, Approximate Octree-Indexed Sampling (APR-OIS), to further accelerate Octree-based down-sampling with a minimal loss of accuracy. It is implemented as an FPGA-based accelerator, functioning as a near-sensor processor