SIA OpenIR  > 机器人学研究室
无人机室内视觉同时定位与地图构建方法与系统研究
Alternative TitleResearch on Simultaneous Localization and Mapping Approach and System for Unmanned Aerial vehicles in Indoor Environment
王化友1,2
Department机器人学研究室
Thesis Advisor何玉庆
Keyword无人机 同时定位与地图构建 特征法 直接法 多相机
Pages83页
Degree Discipline机械电子工程
Degree Name硕士
2019-05-17
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract无人机(Unmanned Aerial Vehicle, UAV)由于具有结构简单、效率高、机动性能好等优点,已经广泛应用于航拍、安防、巡检、环境建模和室内救援等领域。近些年,随着地震等自然灾害的频繁发生,室内救援已经成为了一个急需解决的问题。由于室内自主定位手段缺乏、环境复杂度高,无人机自主完成室内救援任务最大的难点就是无人机的自主定位与室内环境地图构建。解决该问题最典型的方法就是同时定位与地图构建(Simultaneous Localization and Mapping, SLAM)。SLAM技术可以实时估计系统位姿,并同时构建环境地图。目前SLAM系统可以划分为相机定位、位姿优化、闭环检测和地图构建等环节。当前的典型 SLAM 系统还存在实时性差、环境特征与结构依赖性强、容易跟丢、鲁棒性差等不足,本课题针对当前SLAM系统的不足和在实际无人机上应用的难点,开展以下几点研究内容。本文首先针对RGB-D相机构建了一个基于特征的SLAM系统,该系统能够应用在室内环境,并实时运行。该系统采用了基于ORB-SLAM2的框架,主要由三个核心部分组成:追踪、局部建图和闭环检测。追踪部分不像ORB-SLAM2仅仅优化匹配特征点之间的重投影误差,本文的系统选择同时优化匹配特征点之间的重投影误差和逆深度误差来估计相机位姿。局部建图部分负责管理局部关键帧和局部地图点,并通过使用光束平差法(Bundle Adjustment, BA)来优化局部关键帧位姿和局部地图点位置。同时,局部建图部分基于OctoMap建图框架来构建周围环境的占据栅格地图,以用于机器人导航等高级任务,该地图比ORB-SLAM2构建的稀疏地图表达更加有用。不像ORB-SLAM2那样仅仅使用颜色信息,该 SLAM系统的闭环检测同时使用颜色信息和深度一致性来解决错误的数据关联问题,并通过位姿图优化来提高SLAM系统的全局一致性。在TUM RGB-D数据集上的实验结果表明本文的SLAM系统具有比ORB-SLAM2更高的定位精度和鲁棒性。为了满足无人机内外环控制频率的需求,本文基于特征法和直接法的优缺点,将直接法融合到了前一章构建的SLAM系统的追踪线程中,即构建了一个融合特征法与直接法的快速、鲁棒SLAM系统CFD-SLAM (Combining Feature-based Method and Direct Method),该系统能够同时应用于单目、双目和 RGB-D相机。追踪部分融合特征法与直接法,对关键帧和非关键帧分别采用特征法和直接法进行追踪,以提高系统的实时性和在特征缺失环境下的鲁棒性。特征法通过最小化匹配特征点之间的重投影误差来获得关键帧的位姿估计。直接法通过最小化光度误差来获得非关键帧的位姿估计。对于RGB-D相机,逆深度误差被加入到特征法和直接法的优化目标函数中。该SLAM系统还添加了地图保存、地图加载和基于地图的定位功能来解决机器人绑架问题。最后,本文在开源数据集上与典型开源SLAM系统进行了大量对比实验,实验结果证明本文的SLAM系统在保证定位精度的同时,具有较好的实时性和鲁棒性,能够应用于无人机位置控制。为了解决无人机运动过快图像重叠区域不够导致特征无法匹配、单一视角特征缺失等问题,本文构建了一个多相机SLAM系统。该多相机SLAM系统也是基于前面章节构建的SLAM系统,由追踪、局部建图和闭环检测三个核心部分组成。实验结果验证了该多相机SLAM系统的实时性、精确性和单一视角特征缺失环境下的鲁棒性。为了验证本文视觉SLAM系统在实际无人机平台上的性能,本文设计并搭建了一个四旋翼无人机系统,并在室内动捕系统下对本文SLAM系统性能进行了验证。实验结果证明了本文SLAM算法的有效性。
Other AbstractUnmanned aerial vehicle (UAV) has already widely applied to aerial photography, security, inspection, environmental modeling and indoor rescue, because of its simple structure, high efficiency and excellent maneuverability. In these years, with the frequent occurrence of natural disasters such as earthquakes, indoor rescue has become an urgent problem that needs to be solved. Because of the lack of self-positioning means and high environmental complexity in indoor environment, the biggest difficulty for UAV to complete the indoor rescue task independently is self-positioning and indoor environment map construction. The most typical approach to solve the problem is simultaneous localization and mapping (SLAM). SLAM techniques can estimate system system pose in real time, and build an environment map at the same time. Recent SLAM system can be divided into camera localization, pose optimization, loop closure and map construction, etc. Recent standard SLAM system still has some shortcomings, such as poor real-time performance, strong environmental characteristics and structure dependency, easy to lose and poor robustness. This project focuses on the shortcomings of the current SLAM system and the difficulties of the application in real UAV, and carries out the following research contents. This paper presents a feature-based SLAM system for RGB-D cameras that operates in real time, in indoor environments. The system is composed of three central components: tracking, local mapping and loop closure which are based on ORB-SLAM2. The tracking part estimates the pose of a frame via optimizing both the reprojection error and inverse depth error of matching feature points, rather than just optimizing the reprojection error like ORB-SLAM2. The mapping part is in charge of managing local keyframes and map points, and optimizes the pose of all keyframes and the position of all map points by using bundle adjustment, and an occupancy grid map of the surrounding environment is constructed based on the mapping framework OctoMap for high-level tasks, and the map is more useful then the sparse points representation in ORB-SLAM2. The loop closing of ORB-SLAM2 just uses color appearance information, but our system uses both the color appearance information and depth consistency to deal with inaccurate data association, and it improves the global consistency of the SLAM system through pose graph optimization. The experiment results in the TUM RGB-D dataset shows that our system achieves better localization accuracy and robustness than ORB-SLAM2. In order to satisfy the requirement of UAV inner and outer loop control frequency, based on the advantages and disadvantages of feature-based method and direct method, the direct method is integrated into the tracking thread of SLAM system built in the previous chapter, thus a fast and robust SLAM system combining feature-based method and direct method (CFD-SLAM) is proposed, which can be used for monocular, stereo and RGB-D cameras. The tracking part combines feature-based method and direct method, feature-based method and direct method are utilized for keyframes and non-keyframes tracking respectively to improve real-time performance and robustness in low-texture environments. Feature-based method estimates the pose of keyframes by minimizing re-projection error of feature points. Direct method obtains the pose estimation of non-keyframes by minimizing photometric error. For RGB-D cameras, inverse depth error is added into optimization cost function of both feature-based method and direct method. Map saving, map loading and map-based positioning functions are added into the SLAM system to solve the kidnapped robot problem. Finally, experiments on public datasets against other state-of-the-art SLAM systems demonstrate that our system is faster and more robust than the state-of-the-art SLAM system without precision reduction, and can be used for UAV position control. In order to solve the problem of feature can’t be matched caused by not enough image overlapping area of UAV excessive motion and feature missing of single view, a multi-camera SLAM system is constructed, and the multi-camera SLAM system is based on the SLAM system of the front section. It constructs of tracking, local mapping and loop closing three parts. Experiments results verify the real-time performance, accuracy and robustness in the absence of single-view features of the multi-camera SLAM system. In order to test the performance of the visual SLAM system on actual UAV platform, a Quadrotor UAV system is designed and built, and the per formance of the SLAM system is verified in indoor motion capture system. Experiment results demonstrate the effectiveness of the SLAM system.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/25189
Collection机器人学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
王化友. 无人机室内视觉同时定位与地图构建方法与系统研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
无人机室内视觉同时定位与地图构建方法与系(4003KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王化友]'s Articles
Baidu academic
Similar articles in Baidu academic
[王化友]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王化友]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.