基于Kinect传感器的三维点云配准优化

3D point cloud registration optimization based on Kinect

针对目前三维点云配准中传统ICP(Iterative Closest Point)算法存在的速度慢、精度低的问题。采用微软Kinect2.0深度传感器从真实的场景中获取目标物体的点云数据,通过点云分割、滤波、下采样等预处理工作,确保点云配准质量。在点云的粗配准中,使用特征点采样一致性算法,使点云获得更好的初始位置,为精配准创造了良好的初始条件。在点云的精配准中,提出一种利用线性最小二乘法优化的点到面ICP算法。实验结果表明,改进后算法的均方根误差为0.788 mm,时间为56.31 ms。与基于尺度不变特征变换的ICP算法和特征点采样一致性改进ICP算法相比,改进后的算法配准精度分别提高了30.9%和33.6%,速度提高了18.9%和32.1%。

Aiming at the slow speed and low accuracy of the traditional ICP(Iterative Closest Point) algorithm in 3D point cloud registration. The Microsoft Kinect2. 0 depth sensor was used to obtain the point cloud data of the target object from the real scene. Then, preprocessing such as point cloud segmentation, filtering and down sampling were carried out to ensure the quality of point cloud registration. In the coarse registration of the point cloud, the feature point sampling consistency algorithm was used to make the point cloud obtain a better initial position and create a good initial condition for the fine registration. Finally, a point-to-surface ICP algorithm optimized by linear least squares was proposed in the fine registration. The experimental results show that the root mean square error of the improved algorithm is 0. 788 mm and the time is 56. 31 ms. Compared with the ICP algorithm based on SIFT feature points and the improved ICP algorithm based on feature point sampling consistency, the registration accuracy of the improved algorithm is increased by 30. 9% and 33. 6%, and the speed is increased by 18. 9% and 32. 1%, respectively.