SIA OpenIR  > 水下机器人研究室
基于单目直接法的水下视觉惯性里程计方法研究
Alternative TitleUnderwater Visual-Inertial Odometry Based on Monocular Direct Method
赵洪全
Department水下机器人研究室
Thesis Advisor徐会希
Keyword自主水下机器人 水下定位 视觉惯性里程计 单目直接法 低速运动
Pages75页
Degree Discipline机械制造及其自动化
Degree Name硕士
2021-05-21
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract自主水下机器人(Autonomous Underwater vehicle,AUV)是人类探索海洋空间的重要工具,而强大的自主定位能力是实现AUV完全自主运动的关键。当AUV执行水下对接、水下设备检查等近距离作业任务时,需要有更高的定位精度,而基于视觉传感器与惯性测量单元(Inertial Measurement Unit,IMU)融合的视觉惯性里程计(Visual-Inertial Odometry,VIO)能够利用相机捕捉到水下环境中细致的纹理,同时IMU又能够在缺失视觉特征的场景中提供短期较好的位姿估计,相较于其他的水下自主定位方式,VIO方法在水下近距离作业时有更大的优势。VIO是移动机器人领域新兴的一个研究方向,并且受到越来越多学者的关注。但是目前对VIO系统的研究与应用主要集中在陆上环境,并且已经取得了一定的进展,而对水下VIO系统的研究相对较少,其主要原因在于VIO系统在水下场景中的应用存在一些局限性,比如水下是一个弱纹理环境,存在大量重复纹理,视觉传感器在水中的作用范围有限,水下图像存在较严重的退化问题,另外,AUV还通常以较低的速度航行。为将VIO系统有效地应用到水下场景中,本文提出一个适用于AUV执行近距离作业任务时的水下VIO系统,其具体的研究内容如下:1)针对水下环境中存在大量重复性弱纹理特征并且缺乏角点的特点,在该VIO系统的视觉部分使用了直接法进行数据关联。由于直接法只要求图像中有明显的梯度变化,所以直接法还可以利用图像中除角点外的其他特征,该策略使得直接法可以工作在缺乏角点并且存在大量重复纹理的水下环境。此外,为保证提取到足够多有效的特征点,在特征提取过程中还采用了动态调整特征点提取数量和提取阈值的方法。2)本VIO系统采用了基于直接法的数据关联方式。在进行数据关联时,首先对图像施加了一个光度仿射变换函数,该函数可以对图像进行实时的曝光补偿。之后,在系统的粗跟踪阶段,将视觉信息和惯性信息进行了简单的融合,其中IMU传感器的测量值既可以作为先验信息为位姿的粗跟踪提供初始值,又能够在视觉信息无效时提供短期较精确的位姿信息。最后,由于直接法无法通过三角测量的方式计算特征点的深度,所以采用了极线搜索和块匹配的方法对其进行估计。3)针对水下机器人低速运动的特点,在VIO系统初始化阶段主要进行视觉状态量的初始化,而将尺度等惯性状态量加入到联合优化中进行估计。为降低低速运动状态下IMU预积分量的误差,在关键帧的选择策略和边缘化策略中综合考虑了图像的空间分布与时间差异。此外,为保持VIO系统的一致性同时降低尺度状态量的线性化误差,在联合优化过程中使用了改进的动态边缘化策略,该策略对变化的尺度施加了更严格的判断标准,并增加了对视觉信息可靠性的判断。4)利用AQUALOC水下数据集对本文的VIO系统进行了实验研究与分析。首先,通过实验对本VIO系统中一些重要的参数进行了调节,确定了该系统运行时最佳的参数取值。然后,将直接法与特征点法在三种水下场景中进行了对比实验,实验表明本系统采用的直接法在相对恶劣的水下环境中有更好的性能。最后,对本VIO系统在低速运动状态下运行的有效性进行了实验验证,实验表明本系统在低速运动状态下仍能够正常运行,并能实现快速初始化。
Other AbstractAutonomous Underwater Vehicle (AUV) is an important tool for human to explore the ocean space, and the powerful ability of autonomous localization is the key to realize the completely autonomous motion of AUV. When AUV performs close range tasks such as underwater docking and underwater equipment inspection, higher localization accuracy is required. Visual-Inertial Odometry (VIO), based on the fusion of visual sensor and Inertial Measurement Unit (IMU), is capable of capturing fine textures in underwater environments with cameras, and IMU can provide better pose estimation in short term in the scene with lack of features. Therefore, compared with other underwater autonomous localization methods, VIO has more advantages in underwater close range operation. VIO is a new research direction in the field of mobile robot and has attracted more and more scholars' attention. However, the current research and application of VIO mainly focus on the land environment, and some progress has been made, while the research on underwater VIO is relatively less. The main reason is that there are some limitations in the application of VIO in underwater scene, such as underwater is a weak texture environment, there are a lot of repeated textures, and the scope of vision sensor in the water is limited, underwater image has serious degradation problems, in addition, AUV usually navigates at lower speeds. In order to effectively apply VIO to underwater scenes, this paper proposes an underwater VIO suitable for AUV to perform short-range operations tasks. The specific research contents are as follows: 1) In view of the large number of repetitive weak texture features and lack of corner points in the underwater environment, the direct method is used for data association in the visual part of this VIO. Since direct method only requires obvious gradient changes in the image, it can make use of other features besides corners in the image. This strategy enables VIO to work in the underwater environment with lack of corners and a large number of repeated textures. In addition, a feature point extraction method that dynamically adjusts the gradient threshold is used in the feature extraction process to get enough feature points. 2) The proposed VIO adopts a data association method based on direct method. When performing data association, a photometric affine transformation function is first applied to the image, which can perform real-time exposure compensation on the image. Afterwards, in the coarse tracking stage of the system, visual information and inertia information are simply fused, in which the measurement value of IMU sensor can not only be used as a priori information to provide initial value for coarse pose tracking, but also provide short-term accurate pose information when the visual information is invalid. Finally, because direct method cannot calculate the depth of the feature points by triangulation, epipolar search and patch matching are used to estimate it. 3) According to the characteristics of low-speed motion of underwater vehicle, the initialization stage of the system mainly initializes the visual state variables, and the inertial state variables such as scale are estimated during joint optimization. To control the error of the IMU preintegration under low-speed motion, the spatial distribution and time difference of the image are comprehensively considered in the keyframe selection strategy and the marginalization strategy. In addition, in order to maintain the consistency of the VIO while reducing the linearization error of scale, an improved dynamic marginalization strategy is used in the joint optimization process, which imposes more strict judgment criteria on the changing scale and adds judgment on the reliability of visual information. 4) The experimental research and analysis of proposed VIO are carried out by using AQUALOC underwater dataset. Firstly, some important parameters of proposed VIO are adjusted by running the dataset, and the optimal parameter value is determined. Secondly, direct method and feature point method are compared in three kinds of underwater scenes. The experiment shows that direct method has better performance in relatively bad underwater environment. Finally, the effectiveness of this VIO at low speed motion is verified by experiments. The experiment verifies that this VIO can still operate normally in this state and can achieve rapid initialization.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/28954
Collection水下机器人研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
赵洪全. 基于单目直接法的水下视觉惯性里程计方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2021.
Files in This Item:
File Name/Size DocType Version Access License
基于单目直接法的水下视觉惯性里程计方法研(3404KB)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
[赵洪全]'s Articles
Baidu academic
Similar articles in Baidu academic
[赵洪全]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[赵洪全]'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.