VOS-NN: An Efficient Bitonic-Based Voxel-Enhanced Octree Nearest Neighbor Search Accelerator for Large-Scale Point Cloud

Abstract

K-Nearest Neighbor (KNN) search is essential for large-scale point cloud perception in autonomous driving and robotics. However, existing accelerators are hindered by inefficient data structures, high memory consumption, and limited scalability, leading to significant performance bottlenecks in real-time applications. To address these challenges, we present VOS-NN, a scalable and high-performance FPGA-based KNN search framework that integrates a novel Voxel-enhanced Octree Structure (VOS) with a dataflow-driven accelerator architecture. First, the proposed VOS combines adaptive voxel partitioning and octree subdivision with a hybrid memory layout, enabling efficient, fine-grained organization of large-scale point clouds and improved memory locality for rapid neighbor search. Second, we develop a pipelined VOS-NN accelerator that incorporates spatial locality-aware candidate reuse and a parallel bitonic sorting network tailored to the VOS structure, achieving low-latency, high-throughput top-K selection. Experimental results on ZCU104 and XCU50 FPGAs with the KITTI dataset demonstrate that VOS-NN delivers up to 25 $\times $ speedup over CPU implementations and $9.06\times $ acceleration over state-of-the-art KNN accelerators, while maintaining real-time performance for reference sets exceeding three million points. The code is available at: https://github.com/WuYaOTG/VOSNN