SIA OpenIR  > 光电信息技术研究室
基于RGB-D相机的稠密即时定位与地图构建方法研究
Alternative TitleDense Visual SLAM for RGB-D Cameras
付兴银
Department光电信息技术研究室
Thesis Advisor朱枫
Keyword稠密即时定位地图构建 面元 Icp 关键帧 Rgb-d
Pages110页
Degree Discipline模式识别与智能系统
Degree Name博士
2018-11-27
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract

随着移动机器人、无人车、无人机、以及增强现实(Augmented Reality,简称 AR)和虚拟现实(Virtual Reality,简称 VR)等行业的发展,即时定位与地图构建(Simultaneous Localization And Mapping,简称 SLAM)在工业界的需求越来越大。即时定位与地图构建技术可以实时定位系统的位姿,并且同时构建环境的地图。通常情况下,稀疏特征点 SLAM 只能定位系统的位姿,而稠密 SLAM 系统重建的环境的稠密地图可以用作机器人导航、避障、以及用于 AR/VR 行业环境和物体建模等。目前、基于特征点的稀疏 SLAM 系统已经趋于成熟,而稠密的 SLAM 算法还存在许多不足,目前稠密 SLAM 系统按照流程可以划分为相机定位、融合 RGB-D 图像稠密重建、闭环检测和位姿优化和地图更新等环节,本课题针对稠密 SLAM 系统在各个环节的不足,提出以下几点研究内容。(1)改进的基于 ICP 算法的深度图像配准算法。迭代最近点算法(Iterative Closest Point, 简称 ICP)在点云配准领域应用广泛。本文针对 ICP 算法在几何特征少的环境中配准不精确并且鲁棒性低的问题,提出基于几何方差矩阵提取深度图像配准不稳定的方向,根据不稳定方向计算匹配点对约束配准不稳定方向的贡献,通过增大对不稳定方向约束大的点在配准时的权重,提高配准的鲁棒性和精度。同时,利用目前 GPU 架构的特性,提升对匹配点贡献做累加效率。(2)基于面元模型的大范围实时定位与地图构建方法。目前,在稠密 SLAM 系统中,Truncated Signed Distance Function 模型(简称 TSDF)和 Surfel 模型(面元模型)是两种常用的模型。目前基于面元模型的实时重建系统,分配固定大小的存储空间存储重建的模型点,这使得系统能够重建的区域的大小有限;并且随着重建区域的扩展,系统重建的模型点数量增多,算法执行效率会逐渐降低。当系统检测到闭环时,如果当前跟踪累计误差较大,闭环约束通常不能有效地更新模型点。针对基于面元模型重建大范围场景时的上述问题,我们提出在稠密 SLAM 系统中维护两个地图:局部地图和全局地图。新获取的 RGB-D 图像通过和局部地图配准计算位姿,当局部地图中的模型点和相机距离较远并且时间戳超过设定阈值时,将局部地图中的模型点移动到全局地图。系统采用从局部地图和全局地图来回迁移点的策略,扩展重建的区域,并且保持系统的帧率恒定。当系统检测到闭环时,首先做位姿图优化,位姿图优化后,根据位姿图优化的结果优化变形图(Embedded deformation graph),最后采用优化后的变形图更新重建的模型点。在优化变形图时,为了使得大范围模型点的更新更加鲁棒,我们提出一种新的变形图优化策略,我们将优化前后的位姿图作为约束;并且为了使得闭环处闭环前后重建的模型点能够对齐,我们从回环帧抽取三维点作为约束。(3)基于特征点和关键帧的稠密即时定位与地图构建方法。目前,基于特征的稀疏 SLAM 算法已经趋于成熟,基于特征的 SLAM 算法通常具备跟踪精度高、闭环检测和重定位稳定和计算量小等优点。近年来,稠密 SLAM 系统多采用的稠密配准算法,稠密配准算法在处理运动模糊和低纹理场景时往往优于稀疏的 SLAM 算法,稠密 SLAM 系统重建输出的稠密地图也更加实用。结合稀疏 SLAM 和稠密 SLAM 系统的优点,我们提出基于特征点和关键帧的稠密 SLAM 系统。在相机跟踪时,融合稀疏特征点配准的重投影误差和稠密配准误差计算位姿,我们在 g2o 优化库中新建稠密配准约束边,借助 g2o 优化库联合优化稀疏配准误差和稠密配准误差计算位姿。我们在稀疏的 SLAM 系统中,添加稠密建图线程,使得系统可以重建稠密地图。系统不断地检测闭环,在检测到闭环后,执行位姿图优化和全局的 BA(Bundle Adjustment)优化;从优化前的位姿图中抽取关键帧构建变形图的节点,并且将优化前后的位姿图作为约束优化变形图,最后根据优化后的变形图更新重建的稠密地图。我们还提出根据计算的当前关键帧的位姿,从稠密地图投影得到深度图,采用投影的深度图剔除稀疏特征点地图中的外点,投影的深度图还用来计算稀疏特征配准时特征的权值。本文针对实时重建系统位姿配准不精确、不鲁棒,目前算法难以重建大范围场景等问题进行研究,对于基于 ICP 算法的相机跟踪、和基于面元模型重建大范围场景提出新的算法和系统框架。并且融合稀疏 SLAM 系统和稠密 SLAM 系统的优点,提出新的稠密重建系统,我们的重建系统在跟踪精度,重建精度和可扩展性等方面,都优于目前稀疏和稠密 SLAM 系统。

Other Abstract

With the development of unmanned systems and Augmented Reality (AR) and Virtual Reality (VR) industries, Simultaneous Localization And Mapping (SLAM) is growing popular in the industry. SLAM systems are able to localize its location in the environment and simultaneously construct the model of the scene. Generally, sparse SLAM maintains a sparse feature map in the system and cannot produce a dense map. The dense model generated by the dense SLAM system is widely used in robot navigation and 3D scanning in AR/VR. Currently, feature-based sparse SLAM has reached a premature age, while dense SLAM systems still have many problems. A dense SLAM system can be divided into camera localization, dense mapping, loop closure detection and optimization, model updating, etc. In order to make the dense SLAM algorithm step forward, the research contents that we propose as listed below. (1) Improved camera tracking algorithm based on ICP algorithm. Iterative Closest Point (ICP) algorithm is widely used in point cloud registration. The registration result may not be satisfactory in the scene without rich geometric features. We propose to detect the uncertainty when minimizing the point-to-plane distance error of corresponding points, and increasing the weight of the points which constraints the unstable transformations. Our system achieves more accurate and robust camera tracking results. Furthermore, we improve the camera tracking efficiency with GPU Shuffle instructions during warp reduction. (2) Large-scale dense mapping system with surfels. Currently, Truncated Signed Distance Function (TSDF) and surfel are two commonly used models in the dense mapping system. The state-of-the-art dense mapping systems using surfels to fuse RGB-D sequence usually work fine with small-sized regions. But the produced dense model may not be satisfactory when system tracking drift gets large. Furthermore, the system, for example, ElasticFusion, would slow down if the reconstructed dense surface becomes large. In our system, we propose to maintain two maps: local map and global map. The new RGB-D images are integrated into the local map. And the old surfels in the local map that generated in the former times and have a large distance from the current camera center are moved to the global map. The amount the updated surfels in the local map when every new frame arrives are kept bounded, Therefore, the reconstructed scene can be extended to be very large, and the system efficiency is kept high. We optimize pose graph to distribute camera tracking drift when new loop closure is detected, and maintain the consistency between the pose graph and dense model with embedded deformation graph. To make the large-scale dense map update more robust and accurate, we propose to build the constraints with the pose graph and the 3D points sampled from loop closure frames when optimizing embedded deformation graph. (3) Dense mapping system with features and keyframes. At present, the feature-based sparse SLAM algorithm has reached a premature age. In recent years, the dense registration algorithm is popular in dense mapping systems. The dense registration algorithm usually works better than the sparse registration method in dealing with motion blur and low-texture scenes. The dense map produced by the dense SLAM system is also more useful. Combining the advantages of sparse SLAM and dense SLAM systems, we propose a new dense SLAM system. We optimize the combination of the re-projection error of sparse feature points and the dense registration error to calculate camera pose. To reconstruct dense maps, we add dense mapping threads. Camera tracking drift is distributed with pose graph optimization and global BA if new loop closure is detected. We construct embedded deformation graph to updated the dense model. The pose graph before and after optimization is used to build the constraints to optimize the parameters. We propose to use the projected depth map to remove the outers in the sparse feature point map. The projected depth map is also used to calculate the weight of the feature when minimizing the reprojection error. In this paper, we aim to make the dense mapping algorithm step forward. We improve the ICP algorithm and the dense mapping system with surfels. We combine the advantages of sparse and dense SLAM systems and propose a new dense mapping system. Our system achieves state-of-the-art results both in camera tracking and reconstruction accuracy.

Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/23643
Collection光电信息技术研究室
Recommended Citation
GB/T 7714
付兴银. 基于RGB-D相机的稠密即时定位与地图构建方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
Files in This Item:
File Name/Size DocType Version Access License
基于RGB-D相机的稠密即时定位与地图构(17284KB)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
[付兴银]'s Articles
Baidu academic
Similar articles in Baidu academic
[付兴银]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[付兴银]'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.