面向点云识别的最近邻搜索硬件加速器

动态图卷积神经网络(dynamic graph convolutional neural network, DGCNN)作为点云识别主流算法之一, 主要由边缘卷积层构成, 而最近邻搜索操作占据边缘卷积层63%的计算时间。文章针对现有的最近邻搜索加速器准确率较低、速度较慢的问题, 设计一种面向点云识别的最近邻搜索硬件加速器。该加速器采用基于点云分割的并行双调流水排序结构进行2轮双调排序, 并用曼哈顿距离替代欧氏距离衡量点与点距离的远近。实验结果表明, 在同样的实验环境配置下, 相较于其他点云最近邻搜索加速器, 文章设计的最近邻搜索加速器速度提升了3.6倍。

As one of the mainstream algorithms for point cloud recognition, dynamic graph convolutional neural network (DGCNN) is mainly composed of edge convolutional layers, and the nearest neighbor search takes up 63% of the computing time of edge convolutional layers. Aiming at the problems of low accuracy and slow speed of existing nearest neighbor search accelerators, this paper proposes a design of nearest neighbor search hardware accelerator for point cloud recognition. In this accelerator, a parallel bitonic flow sorting structure based on point cloud segmentation is adopted for two rounds of bitonic sort, and Manhattan distance is used instead of Euclidean distance to measure the distance between points. Experimental results show that the speed of the proposed nearest neighbor search accelerator is 3.6 times faster than that of the existing point cloud nearest neighbor search accelerator under the same experimental environment configuration.