基于GPU的实时三维点云数据配准研究

在三维重建中，不同摄像机坐标系下点云配准耗时过多。为此，提出一种基于图形处理单元(GPU)的实时三维点云数据配准算法。利用投影映射法获取匹配点对，使用点到切平面距离最小化方法计算变换矩阵，通过GPU多线程并行处理大规模图像数据。实验结果表明，对于分别包含307 200个数据的2帧点云，在保持原有配准效果的基础上，该算法的最优耗时仅为基于CPU的最近邻迭代算法的11.9%。

Research on Real-time 3D Point Cloud Data Registration Based on GPU

In 3D reconstruction, point cloud registration time of different camera coordinate system is too long. In order to sovle this problem, this paper proposes a real-time 3D point cloud data registration algorithm based on Graphic Processing Unit(GPU). The projection mapping method is used to find corresponding point pairs, the transformation matrix is obtained by minimizing the distance from the point to the tangent plane, and uses the GPU to multi-thread and parallelly process mass image data. Experimental results show that, in order to keep original registration effect, the optimal time-consuming of this algorithm is about 11.9% compared with Iterative Closest Point algorithm based on GPU(ICP-GPU) in containing 307 200 data 2 frame point cloud.