基于RGB-D的同步定位与构图算法研究
发布时间：2018-11-03 06:52
【摘要】：SLAM(Simultaneous Localization and Mapping)作为机器人实现自主化的基础,能够在未知环境下实现机器人定位与环境地图构建的同步进行,近年来受到了越来越多的关注。按照选用的传感器不同可将SLAM分为三种:基于声呐、基于雷达和基于视觉的SLAM。本文主要研究视觉SLAM中的RGB-D SLAM,即以RGB-D相机为视觉传感器的同步定位与构图。KINECT作为RGB-D传感器价格便宜,且所获得的图片质量也较好,因此本文在KINECT的基础上进行研究与实验。本文首先对SLAM的发展现状与研究内容进行了分析介绍,将SLAM分为以图像处理为目的的前端和以优化为目的的后端,并对前端与后端的各个环节进行了详细阐述与推导。随后,针对各环节存在的问题提出了相应的改进方法。在特征提取环节,通过对主流提取算法的仿真分析,提出以FAST-SIFT的结合方式进行特征提取,不仅加快了计算速度,而且保证了提取的稳定性。另外,打破了传统算法中将特征点视为一体的方式,对特征点进行分类,并说明了平面点的优势,继而提出以“先面后点”的方式提取出更多的平面点。在特征匹配中,同样采取先匹配面后匹配面上的点的方式,通过平面参数对点云深度进行修正,减小了深度噪声的影响。随后又提出一套完整严格的误匹配剔除规则,避免了因为误匹配导致的不良结果。在运动估计环节中,用经过修正后的点云进行配准,提出基于PROSAC((Progressive Sample Consensus))的ICP(Iterative Closest Point)算法,相比与传统算法更为快速、准确。最后,在后端算法中,制订了合理的环回检测策略,并以G2O为仿真工具,验证了添加了回环后的结果更为精确。本文首先以网上公开数据集freiburg2_pioneer_slam作为仿真数据,对改进算法进行了仿真分析,随后采用手持KINECT的方式,在实验室环境中进行实验。最终仿真与实验结果表明了改进算法的优越性。
[Abstract]:As the basis of autonomous robot, SLAM (Simultaneous Localization and Mapping) can realize the synchronization of robot localization and environment map construction in unknown environment. In recent years, more and more attention has been paid to SLAM (Simultaneous Localization and Mapping). SLAM can be divided into three types according to the sensors selected: sonar based, radar based and visual based SLAM. This paper mainly studies the RGB-D SLAM, in visual SLAM, that is, the synchronous location and composition of RGB-D camera as visual sensor. KINECT is cheap as RGB-D sensor, and the image quality is good. Therefore, this paper based on the KINECT research and experiment. In this paper, the development status and research contents of SLAM are analyzed and introduced. The SLAM is divided into the front end for image processing and the back end for optimization, and each link between the front end and the back end is described and deduced in detail. Then, according to the existing problems of each link, the corresponding improvement methods are put forward. In the part of feature extraction, through the simulation analysis of the mainstream extraction algorithm, it is proposed that the combination of FAST-SIFT can not only speed up the calculation speed, but also ensure the stability of the extraction. In addition, it breaks the traditional method of treating feature points as a whole, classifies feature points, explains the advantages of plane points, and then proposes to extract more plane points by "first face and then point". In feature matching, the point on the surface is matched first and then the point on the surface is matched. The depth of the point cloud is modified by plane parameters to reduce the influence of depth noise. Then a complete set of strict rules for eliminating mismatch is proposed to avoid the bad results caused by mismatch. In motion estimation, the modified point cloud is used to register, and the ICP (Iterative Closest Point) algorithm based on PROSAC (Progressive Sample Consensus) is proposed, which is faster and more accurate than the traditional algorithm. Finally, in the back-end algorithm, a reasonable loopback detection strategy is worked out, and G2O is used as a simulation tool to verify that the results after adding the loop are more accurate. In this paper, the improved algorithm is simulated and analyzed by using the open data set (freiburg2_pioneer_slam) on the Internet as simulation data, and then the experiment is carried out in the laboratory environment by handheld KINECT. The simulation and experimental results show the superiority of the improved algorithm.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
本文编号：2307020
[Abstract]:As the basis of autonomous robot, SLAM (Simultaneous Localization and Mapping) can realize the synchronization of robot localization and environment map construction in unknown environment. In recent years, more and more attention has been paid to SLAM (Simultaneous Localization and Mapping). SLAM can be divided into three types according to the sensors selected: sonar based, radar based and visual based SLAM. This paper mainly studies the RGB-D SLAM, in visual SLAM, that is, the synchronous location and composition of RGB-D camera as visual sensor. KINECT is cheap as RGB-D sensor, and the image quality is good. Therefore, this paper based on the KINECT research and experiment. In this paper, the development status and research contents of SLAM are analyzed and introduced. The SLAM is divided into the front end for image processing and the back end for optimization, and each link between the front end and the back end is described and deduced in detail. Then, according to the existing problems of each link, the corresponding improvement methods are put forward. In the part of feature extraction, through the simulation analysis of the mainstream extraction algorithm, it is proposed that the combination of FAST-SIFT can not only speed up the calculation speed, but also ensure the stability of the extraction. In addition, it breaks the traditional method of treating feature points as a whole, classifies feature points, explains the advantages of plane points, and then proposes to extract more plane points by "first face and then point". In feature matching, the point on the surface is matched first and then the point on the surface is matched. The depth of the point cloud is modified by plane parameters to reduce the influence of depth noise. Then a complete set of strict rules for eliminating mismatch is proposed to avoid the bad results caused by mismatch. In motion estimation, the modified point cloud is used to register, and the ICP (Iterative Closest Point) algorithm based on PROSAC (Progressive Sample Consensus) is proposed, which is faster and more accurate than the traditional algorithm. Finally, in the back-end algorithm, a reasonable loopback detection strategy is worked out, and G2O is used as a simulation tool to verify that the results after adding the loop are more accurate. In this paper, the improved algorithm is simulated and analyzed by using the open data set (freiburg2_pioneer_slam) on the Internet as simulation data, and then the experiment is carried out in the laboratory environment by handheld KINECT. The simulation and experimental results show the superiority of the improved algorithm.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
【参考文献】
相关期刊论文 前3条
1 刘浩敏;章国锋;鲍虎军;;基于单目视觉的同时定位与地图构建方法综述[J];计算机辅助设计与图形学学报;2016年06期
2 闵华松;杨杰;;融合IMU的RGBD-SLAM算法改进研究[J];计算机工程与设计;2015年01期
3 石杏喜;赵春霞;郭剑辉;;基于PF/CUKF/EKF的移动机器人SLAM框架算法[J];电子学报;2009年08期
相关硕士学位论文 前3条
1 辛冠希;基于RGB-D摄像机的同步定位与建图研究[D];哈尔滨工业大学;2016年
2 徐冬云;基于Kinect的视觉同步定位与建图研究[D];哈尔滨工业大学;2016年
3 李洪臣;单目视觉移动机器人SLAM方法建模与仿真分析[D];电子科技大学;2014年
，本文编号：2307020
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2307020.html
