CASIA OpenIR  > 智能制造技术与系统研究中心
基于双目视觉与惯导融合的智能车定位应用研究
黄馨
Subtype硕士
Thesis Advisor汤淑明
2020-05-28
Degree Grantor中国科学院大学
Place of Conferral中国科学院大学
Degree Discipline控制工程
Keyword视觉惯导SLAM 双目视觉 惯导定位 智能车 位姿估计
Abstract

近年来,自动驾驶引起了社会的广泛关注。定位技术在自动驾驶领域发挥着重要作用,是智能车安全行驶不可或缺的核心技术之一。现有智能车通过卫星导航系统实现定位,但是在高楼林立的城市环境中,卫星信号差、长时间丢失等情况容易导致定位精度差、定位失败等问题。因此,实现智能车在无卫星信号场景下的精确定位具有重要的理论意义和应用价值。基于视觉SLAM与惯导定位技术的互补性,本文针对城市交通环境中的智能车定位问题展开双目视觉与惯导融合的相关研究,实现应用于智能车平台上的高精度、低成本、实时的组合定位算法。本文的主要工作和创新性成果包括:

1)惯导定位算法研究。本文基于惯导定位原理和IMU模型,研究了IMU状态传播和IMU预积分方法。在离散化传播IMU状态时,分别采用欧拉法、中值法和四阶龙格库塔法进行航位推算。通过实验比较三种数值积分方法对IMU状态传播的影响,为后续的算法研究提供了基础。

2)双目视觉定位算法研究。在实际场景中,由于存在光照变化、动态物体和快速自运动,KLT光流跟踪往往会陷入局部极小值,从而导致定位精度下降。针对该问题,本文提出了一种融合多种约束的轻量自适应光流跟踪方法。该方法设计了轻量双向环形检验,并结合视差约束进行外点剔除。此外,为了平衡特征点的数量和质量,该方法设计了自适应特征选择策略。实验结果表明,提出的方法能够有效去除外点,改善光流跟踪,从而提高算法的位姿估计精度和计算效率。

3)双目视觉与惯导融合定位算法研究。双目视觉惯性SLAM通过融合视觉数据和惯性数据来提升定位性能,但是增加了计算负担。为提高双目视觉惯性SLAM的精度和计算效率,需要建立可靠的特征关联,优化算法结构并结合一定的加速手段。本文采用提出的轻量自适应光流跟踪方法进行外点剔除,以提高位姿估计精度。此外,本文基于视觉惯性边缘化过程中矩阵的稀疏性,设计了二步边缘化策略。该策略通过改进矩阵构建结构和舒尔补计算方法,实现快速边缘化,从而提高算法的计算效率。实验结果表明,本文方法能够有效提升位姿估计精度,且计算效率提高了32%

最后,基于以上研究和方法,本文实现了基于双目视觉与惯导融合的智能车定位应用,通过落地试验验证了算法的可行性。

Other Abstract

In recent years, autonomous driving has attracted widespread attention in the world. Localization technology, which plays an important role in the field of automatic driving, is one of the key technologies for the driving safety of intelligent vehicles. Traditionally, vehicles are located via satellite navigation system. However, it is easy to obtain inaccurate locations and even lead to location failures due to the poor satellite signals in the modern city of skyscrapers. Therefore, it is of great theoretical and pratical value to realize the precise localization for intelligent vehicles in the non-signaled or poor-signaled scenes. Based on the complementarity of visual SLAM and inertial positioning technology, we carry out the research of binocular visual-inertial SLAM (VI-SLAM) on localization in the urban traffic environment. We aim to realize high-precision, low-cost and real-time localization algorithm in intelligent vehicle platforms in this thesis. Its main work and innovative achievements are as follows:

(1) Research on the current inertial positioning algorithms. Based on the principle of INS and IMU model, this thesis studies the methods of IMU state propagation and preintegration. During the discretization calculation of IMU state, the Euler method, midpoint method, and RK4 method are adopted for dead reckoning. The influence of three numerical integration methods on IMU state propagation is compared by experiments, which provides the foundation for the following research.

(2) Research on the binocular visual SLAM algorithm. In the actual scenes, due to illumination changes, dynamic objects, and large displacements, KLT tracker is prone to fall into local minima, leading to the decrease of location accuracy. To solve this problem, this thesis proposes a light and adaptive feature tracking method integrated with multiple constraints. The method introduces a light bi-circular check and combines with disparity constraint to remove outliers. In addition, an adaptive feature selection strategy is designed to balance the quantity and quality of features. Experiments demonstrate that the proposed method can effectively remove outliers, improve optical flow tracking, and increase the accuracy and efficiency of pose estimation.

(3) Research on the binocular VI-SLAM algorithm. The binocular VI-SLAM algorithm boosts the performance of localization by fusing the visual measurements and inertial measurements at the cost of computational burden. In order to improve the accuracy and efficiency of binocular VI-SLAM algorithm, it is necessary to establish reliable feature association, optimize the algorithm structure, and adopt certain acceleration methods. This thesis introduces the proposed light and adaptive feature tracking method to improve the accuracy of pose estimation. In addition, based on the sparsity of matrix during visual-inertial marginalization, this thesis adopts a two-step marginalization strategy. By improving the construction structure of matrix and schur complement method, the strategy realizes fast computation of marginalization, and improves the computation efficiency of algrithm. Experiments demonstrate that our method can improve the accuracy of pose estimation with 32% speedup.

Finally, this thesis implements the application of intelligent vehicle localization based on binocular visual-inertial fusion, and verifies the feasibility of the algorithm through tests in actual scenes.

Pages99
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/39221
Collection智能制造技术与系统研究中心
Recommended Citation
GB/T 7714
黄馨. 基于双目视觉与惯导融合的智能车定位应用研究[D]. 中国科学院大学. 中国科学院大学,2020.
Files in This Item:
File Name/Size DocType Version Access License
修订后论文2.pdf(3660KB)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
[黄馨]'s Articles
Baidu academic
Similar articles in Baidu academic
[黄馨]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[黄馨]'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.