Planar object detection from 3D point clouds based on pyramid voxel representation
Planar object detection from 3D point clouds based on pyramid voxel representation
复制标题
基于金字塔体素表示的 3D 点云平面物体检测
DOI：
10.1007/s11042-016-4192-6
复制
发表时间：
2017-11-01
影响因子：
3.6
通讯作者：
Bai, Dongfang
中科院分区：
文献类型：
--
作者：
Hu, Zhaozheng;Bai, Dongfang
平面检测是许多计算机视觉应用中的基本任务。针对现有方法计算量大的问题，提出了一种快速可靠的三维点云平面检测方法。3D空间首先被划分为金字塔体素。每个3D点被分配给每个金字塔层处的一个体素，使得所有3D点由金字塔体素表示。对于每个体素，我们应用特征值分解来分析内部的3D点，并提出了一个快速平面检测的索引。特别是，平面指数的有效计算，没有显式的特征值分解，以提高计算。基于体素之间的几何关系，对检测到的平面体素进行分析和合并，用于平面目标检测。该方法对三维点云采用逐体素处理而不是逐点处理，从而大大提高了计算效率，并具有良好的检测效果。该方法已通过Kinect 1.0的RGB-D传感器在室内和室外环境中采集的实际三维点云进行了验证。实验结果表明，该方法可以快速检测单个和多个平面目标在这两种环境中。方法的精密度和准确度分别为97.1%和94.5%。与现有方法相比（例如，Hough变换，RANSAC），所提出的方法可以大大提高计算效率在几个数量级。
Planar detection is a fundamental task in many computer vision applications. This paper proposed a fast and reliable plane detection method from 3D point clouds to address the high computation cost problem in existing methods. The 3D space is first partitioned into pyramid voxels. Each 3D point is assigned to one voxel at each pyramid layer so that all the 3D points are represented by pyramid voxels. For each voxel, we apply the Eigen value decomposition to analyze 3D points inside and propose an index for fast plane detection. Especially, the plane index is efficiently computed with no explicit Eigen value decomposition to enhance the computation. The detected planar voxels are analyzed and merged for planar object detection based on geometric relationship between voxels. The proposed method uses voxel-wise instead of point-wise processing of the 3D point clouds so that it can greatly enhance the computation efficiency yet with good detection results. The proposed method has been validated with actual 3D point clouds collected by RGB-D sensor of Kinect 1.0 in both indoor and outdoor environments. The results demonstrate that the proposed method can quickly detect single and multiple planar objects in both environments. The precision and the accuracy of the proposed method are 97.1% and 94.5%, respectively. Compared to existing methods (e.g., Hough Transform, RANSAC), the proposed method can greatly enhance the computation efficiency in several orders of magnitudes.