RGB-D Sensor based Plane Extraction and Scan Matching Method in 3D Indoor Environments

Abstract

The 3-D mapping of indoor environment is an important requirement for the mobile robots to autonomously execute various tasks including localization, navigation and target tracking. The commercial launch of low-cost and light-weight RGB-D sensors such as the Microsoft Kinect provides an attractive, powerful alternative for the generation of 3-D map of indoor environments. For the 3-D mapping systems using RGB-D sensors, scan matching is the most common way to find the relative pose offset of the robot between two successive range sensor samples. In this thesis, a novel scan matching method based on plane features is presented which is available for the3-D mapping of indoor environments using hand-held RGB-D sensors, including the plane extraction and the matching method based on the plane parameter space transform, and the estimation of the sensor pose based on the extracted plane features.

While extracting planes, the local neighborhood of a point in the Cartesian space is estimated as a local planar patch which is projected to the plane parameter space. The Statistical Information Grid (STING) structure is constructed to organize the points in plane parameter space. The plane features are then extracted by searching the STING structure from top to bottom for the grid which contains most of the points that are densely distributed. The points in certain grid are most likely from a same plane in the Cartesian space, and the grid parameter is used to formulate the extracted plane features. To match the extracted planes, a graph structure is created with the extracted planes as nodes, and the geometrical relationship between planes as edges. Similarity between nodes from different graphs is defined considering the plane color and geometrical relationship between planes, which is employed to establish the correspondence of the extracted planes.

The estimation method of RGB-D sensor pose is proposed using the corresponding planes extracted from successive frames. Three cases are considered according to the number of constraints provided by the corresponding planes for the 6-DOF transformation in 3D space, including6-DOF constraints, 5-DOF constraints and 3-DOF constraints, and specific solutions to the three cases are proposed respectively. Especially for the latter two under-constraint cases, the estimation of the sensor pose relies only on the extracted plane features and points in plane parameter space, rather than the assistance of other features.

Key words: hand-held RGB-D sensor;3-D mapping of indoor environment; scan matching; plane feature

摘要

室内三维环境地图创建是机器人在室内环境中完成定位、导航和跟踪等任务的前提。RGB-D传感器，如Microsoft Kinect出现后，很快就因其重量轻、成本低和信息量丰富等特点，在室内三维环境建图中得到了重要的应用。在用RGB-D传感器完成室内三维环境建图时，扫描匹配方法是完成其位姿估计的常用方法。针对基于手持RGB-D传感器的室内三维环境建图，本文提出一种基于平面特征的扫描匹配方法，主要包括基于平面参数空间变换的平面提取与匹配方法，以及基于平面特征的RGB-D传感器位姿估计方法。

在进行平面提取时，首先对笛卡尔空间中的扫描点进行局部平面拟合，得到各个扫描点处的局部平面并将其投影到平面参数空间中，形成平面参数空间数据点集合，然后在平面参数空间中构建统计信息网格(Statistical Information Grid, STING)结构，通过对统计信息网格结构进行自顶向下的搜索，找到包含数据点数目足够多，且分布足够集中的网格，即判断该网格内的数据点来源于笛卡尔空间中的同一平面，并用该网格的参数描述提取出的平面特征。而在平面匹配时，首先以提取出的平面为节点，以平面之间的几何关系为边，建立平面关联关系图，然后结合平面颜色信息以及平面之间的几何关系信息，对两帧平面关联关系图的节点之间的相似度进行定义，并根据此相似度确定平面之间的匹配关系。

在利用两帧扫描数据中提取出的具有匹配关系的平面进行RGB-D传感器位姿估计时，根据匹配的平面能够为三维空间位姿变换求解提供的约束个数，可以将其分为六自由度约束、五自由度约束和三自由度约束三种情况，并分别给出具体的求解方案。针对后两种约束不足的情况，本文提出一种新的扫描匹配方法，该方法不依赖于其他特征，而是使用两帧平面参数空间中数据点集合提供所需的约束，进而完成RGB-D传感器的位姿估计和环境的三维建图。

关键词：手持RGB-D传感器；室内三维环境建图；扫描匹配；平面特征

Recommended citation: Q. Sun. RGB-D Sensor based Plane Extraction and Scan Matching Method in 3D Indoor Environments. M.S. thesis, Nankai University, Tianjin, China, 2016. http://sunqinxuan.github.io/files/publications-2016-05-25-M.S.Thesis.pdf