CASIA OpenIR  > 毕业生  > 硕士学位论文
多视几何与数据驱动相结合的深度估计算法研究
陈天
Subtype硕士
Thesis Advisor张大朋 ; 刘希龙
2019-06-13
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Name工程硕士
Degree Discipline控制工程
Keyword环境感知 计算机三维视觉 视觉感知 深度估计
Abstract

随着视觉感知和深度学习技术的发展,计算机三维视觉在增强现实、自动驾 驶等领域的应用取得了长足的进步。其中,基于视觉的立体信息感知是其中的重 要基础性通用技术模块之一。本文面向不同场景和基于不同的图像数据获取方 式,以获得稠密、精确的场景深度为立足点,开展了多视几何与数据驱动相结合 的深度估计算法研究。

本文分别概述了基于多视几何理论的深度估计算法和基于数据驱动的深度 估计算法研究现状,并分别介绍了这两类深度估计算法的优缺点,介绍了不同输 入图像数据类型的深度估计方法和原理,并对研究背景和意义做了介绍。本文针 对数据驱动的深度估计方法存在的固有缺陷,结合视觉几何方法,做了以下研 究。

首先,设计了 RGB-D 数据驱动的单目深度估计算法框架,将深度梯度图和 深度图进行联合优化,梯度损失在边缘区域响应较大,使得模型专注于图像中物 体的边缘区域。为了进一步提升边缘区域的预测精度,本文在预训练阶段使用数 据归一化方法及边缘区域加权处理方法。同时在模型中引入注意力模块,使得网 络在每层的局部感受域内融合空间和通道信息来构建信息特征,提升网络提取 的特征指向性,进而提高模型的精确度。

其次,设计了双目数据驱动的单目深度估计算法模型框架,根据双目视差重 建理论设计无监督训练模式,避免使用深度真值进行模型训练。使用逆深度进行 三维深度的回归预测,提高室外场景的预测精度。提出了局部平面优化算法,基 于深度信息得到空间法向量,在优化过程中将像素点的法向量差异作为平面约 束条件,提高大平面区域像素的深度预测精度。

最后,针对单目深度估计模型在场景深度预测中存在的尺度不定性问题,设 计了基于多视几何的预测深度和数据驱动预测深度的线性融合模型。多视图存 在的对极几何约束可以较为准确地估计共同观测到的特征点的深度,利用这些 稀疏特征点可以对不同帧的网络输出深度图进行尺度校正。同时,由于图像中的 平面结构不受尺度影响,使用网络输出的深度图进行几何转化可以大致估计中 图像场景中的平面信息,利用这些平面信息对几何深度进行优化。

Other Abstract

With the development of visual perception and deep learning technology, great progress has been made in 3D computer vision for augmented reality or automatic driv- ing etc. In 3D computer vision, visual based stereo information perception is one of the most important basic general technical modules. This paper is aiming to obtain dense and accurate scene depth in different scenarios. Based on different image data acquisi- tion methods, the depth estimation algorithms combining multi-view geometry and data driving are studied.

This paper reviews the state of the art of depth estimation algorithms based on multi-view geometry theory and data-driven methods, and introduces the advantages and disadvantages of these two types of depth estimation algorithms, and also introduces the principles of depth estimation methods when handling different types of image data. In this paper, the following researchs are done to overcome the inherent defects of the data-driven depth estimation methods, by combining visual geometry algorithms.

Firstly, a novel RGBD data-driven monocular depth estimation framework is in- troduced. In the framework, depth gradient map and depth map are jointly optimized. The gradient loss has large impact in the edge region, which makes the model focus on the edge region of the image. In order to improve the prediction accuracy of the edge region further, this paper uses normalization and edge weighting in the pretrain- ing stage. At the same time, the attention mechanism is used in the model, which fuse the space and channel information into the network in which information features are constructed in the local receptive domain of each layer. In this way, the feature extrac- tion directionality is improved in the network, and finally improves the accuracy of the model.

Secondly, a novel binocular data-driven monocular depth estimation framework is introduced. The unsupervised training paradigm is designed according to the binocular disparity reconstruction, which makes the depth ground truth free to carry out model training. In addition, the inverse depth is used to perform 3D depth regression predic- tion to improve the accuracy of outdoor scenes. A local plane optimization mothod is proposed. The space normal vector is computed by depth information. In the optimization process, the normal vector difference of the pixel points is used as the plane constraint condition to improve the depth prediction accuracy of the pixels in the large plane region.

Finally, a fusion model to cooperate the multi-view geometry depth estimation with data-driven depth prediction is introduced to reduce the scale uncertainty in learning based monocular depth prediction. In the fusion model, the epipolar geometry constraint of multi-view can estimate accurate depth, while to realize real time, only depth of sparse feature points can be estimated. By using these depth of sparse feature points, the scale of network predicted depth maps can be estimated. In the other hand, 3D shape information from the image is independent to the scale of depth, using the depth map, the planes in the scene can be extracted and the geometric depth can be optimized by using the plane and corresponding homography.

Subject Area计算机感知
MOST Discipline Catalogue工学 ; 工学::控制科学与工程
Pages92
Funding ProjectNational Natural Science Foundation of China[61503376] ; National Natural Science Foundation of China[NSFC 61633020]
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23851
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
陈天. 多视几何与数据驱动相结合的深度估计算法研究[D]. 中国科学院自动化研究所. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
6.3纸质版.pdf(5263KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[陈天]'s Articles
Baidu academic
Similar articles in Baidu academic
[陈天]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[陈天]'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.