面向家庭服务机器人的室内场景三维重建

CASIA OpenIR > 毕业生 > 硕士学位论文

	面向家庭服务机器人的室内场景三维重建
	孟馨蕊
	2018-05
学位类型	工学硕士
中文摘要	随着老龄化社会的到来，人们对家庭服务机器人的需求越来越大。人们期待家庭服务机器人可以通过人机自然语言交互的方式，实现自主导航并提供取送物品等服务。为了实现上述目标，服务机器人首先需要获得室内环境的三维地图模型，以方便机器人进行路径规划及躲避障碍物。近年来随着视觉传感器技术的发展，一系列RGB-D 相机的出现使得室内三维重建技术产生了巨大的进步。然而，如何在机器人移动过程中准确且高效地重建出室内场景的三维模型，仍是一个具有挑战性的问题。本文对此问题进行了一系列的探索和研究，主要工作有： 1. 设计了一个基于单台深度相机与机器人位姿传感器的SLAM 系统。该系统中，本文首先通过ROS 操作系统获取机器人的姿态；然后基于机器人坐标系与相机坐标系的标定关系，将机器人的姿态转换为相机的姿态；接着使用自适应加权融合的算法将基于机器人位姿传感器的相机姿态与基于视觉的相机姿态进行了融合；另外设计了两种机器人运动模式以消除累积误差。多个实验证明了该算法的可行性与鲁棒性。 2. 构建了一个基于多台深度相机的稠密SLAM 系统。在该系统的预处理阶段，本文提出了两种相机外参数的标定方法：一种为基于手眼标定的方法，另一种为基于视觉SLAM 的方法。在多相机SLAM 阶段，本文首先从多台相机在同一时刻的姿态中选取一个误差最小的姿态作为参考姿态，使用相机间的外参数与参考姿态确定其余相机的最终姿态；然后本文在构成三维模型的面元上设置了“相机编号”的属性，令不同相机编号的面元进行不同的形变，提高了系统对闭合环路的适应性。同时，大量实验表明了该多相机RGB-D SLAM 系统的建模高效性与定位鲁棒性。 3. 实现了一个基于移动机器人的室内语义三维重建系统。系统包括以下模块:使用键盘控制机器人的运动进行数据采集；基于一种融合深度图像信息的卷积神经网络进行语义分割；基于一种多传感的RGB-D SLAM 算法进行语义三维重建。
英文摘要	Aging society is a great challenge to humanity, and service robot seems a promising future technology to alleviate the social demands of seniors, for which how to construct a semantic 3D map of indoor environment appears an indispensible step to facilitate navigation and help the execution of various caring functions, such as fetching objects. Due to cluttered environments, in particular, ubiquitous object occlusion phenomenon, currently available 3D reconstruction techniques in computer vision field seem unfortunately not up for the job. This thesis is to explore new semantic 3D indoor map reconstruction techniques under the RGB-D SLAM paradigm, our main contributions include: 1. A multi-sensor SLAM system based on a single depth camera and a robot pose sensor is proposed. In the system, we first extract the robot's pose data through the ROS operating system. Then we convert the robot's pose into the camera's pose based on the calibrated relationship between the robot coordinate system and the camera coordinate system. After that, an adaptively weighted fusion algorithm is used to fuse the inferred camera pose with the pure vision-based camera pose. And finally, we design two robot movement patterns to reduce the cumulative errors. The performance of the system is validated by a number of experiments. 2. A dense SLAM system based on multiple depth cameras is proposed. In the preprocessing stage of the system, we propose two methods for calibrating extrinsic parameters, one is based on hand-eye calibration and the other is based on visual SLAM. In the multi-camera SLAM stage, firstly we choose the pose with the minimal error as the reference one out of the poses of multiple cameras, and the final pose of the cameras is derived from the calibrated extrinsic parameters and the reference pose. Then we assign an attribute of camera number to the surfels of the 3D model, and the deformation of the surfels with different camera number is carried separately. A number of experiments are carried out to illustrate the modeling efficiency and the localization robustness of our multi-camera RGB-D SLAM system. 3. A 3D semantic indoor map building prototype system is proposed basedon a moving robot. The system units include: controlling the movement of the robot by keyboard; collecting data; a CNN based 2D image semantic segmentation; and semantic 3D reconstruction based on a multi-sensing RGB-D SLAM. In the SLAM part, we assign a label counter to the surfels of the 3D model to fuse potential conflicting object labels, and by which a semantic 3D reconstruction is finally obtained.
关键词	同时定位与建图三维重建 Rgb-d 视觉惯性里程计相机外参标定语义slam 机器人
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/20949
专题	毕业生_硕士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	孟馨蕊. 面向家庭服务机器人的室内场景三维重建[D]. 北京. 中国科学院研究生院,2018.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
Thesis-孟馨蕊-最终版.pdf（15414KB）	学位论文		限制开放	CC BY-NC-SA