基于 GPU 的并行 ICP 点云配准算法研究

针对传统串行精配准算法在海量点云数据配准时计算效率低的问题, 文章利用图形处理器(graphics processing unit, GPU)的多线程计算能力将传统算法并行化, 基于GPU实现并行化的统一计算设备架构迭代最近点(compute unified device architecture iterative closest point, CUDAICP)算法。首先采用粗配准方法对源点云进行旋转平移, 得到源点云的初始位置, 再将其与目标点云输入CUDAICP算法进行精配准; 对房间点云、带有楼梯的房间点云2种场景点云数据进行配准实验。结果表明: 在粗配准中, 采样一致性初始配准(sample consensus initial alignment, SAC-IA)算法在不同场景下具有较好的效果; 在精配准中, CUDAICP算法与传统迭代最近点(iterative closest point, ICP)算法相比, 在保证精度的同时, 速度提升最高可达8.2倍。

In response to the problem of low efficiency of traditional serial fine registration algorithms in calculating massive point cloud data, this paper parallelizes the traditional algorithm by leveraging the multi-thread computing capability of graphics processing unit (GPU). The parallelized algorithm, called compute unified device architecture iterative closest point (CUDAICP) algorithm, is implemented on GPU. Firstly, coarse registration methods are used to perform initial rotation and translation on the source point cloud, obtaining its initial position. Then, the target point cloud is input into the CUDAICP algorithm for fine registration. Experiments on room point clouds and room point clouds with stairs demonstrate that the sample consensus initial alignment (SAC-IA) coarse registration algorithm has good performance in different scenarios. In terms of fine registration, the CUDAICP algorithm achieves up to 8.2 times faster speed compared to the traditional iterative closest point (ICP) algorithm while maintaining accuracy.