CASIA OpenIR  > 毕业生  > 博士学位论文
融合多源传感器信息的机器人定位与建图方法研究
施鹏程
2024-05-17
Pages154
Subtype博士
Abstract

同时定位与地图构建(Simultaneous Localization and Mapping, SLAM)作为自主机器人应用的基本先决条件,近年来取得了蓬勃的发展,优秀的研究成果逐步实现从实验室向工业界的落地。层出不穷的数据集和实际场景的多样性,对SLAM系统的定位精度和鲁棒性提出了更高的要求。一些复杂的外界环境,比如快速光照变化、弱纹理、高遮挡、场景结构退化、剧烈运动等,会导致SLAM系统性能下降甚至运行失败。特别是对于使用单一传感器的算法,由于其固有的缺陷和局限性,很难适应多样化的环境。融合多源信息的SLAM方法能够结合不同传感器的优势,使其在复杂多样现实场景中适应性和鲁棒性地提高成为可能。本文围绕激光雷达、惯性测量单元和相机三种模态的传感器信息源,针对视觉-惯性、激光雷达-惯性和激光雷达-惯性-视觉定位与建图三种具体的任务展开研究。本文主要的研究内容与贡献总结如下:

一、提出一种基于非线性因子估计位姿图优化(Pose Graph Optimization,PGO)的视觉-惯性定位与建图方法。本方法综合考虑了位姿图中序列约束(时间上相近)和闭环约束(空间上相似)的协方差估计,以更准确地表示系统真实不确定性。对于序列约束,本方法引入绝对位姿和相对位姿因子,并利用非线性因子提取方法从累积的里程计中获取最优的协方差矩阵;对于闭环约束,本方法引入尺度因子,并提出动态尺度估计方法以近似其协方差的尺度。得益于所提尺度估计算法,本方法在面对由动态物体干扰造成的假阳性闭环场景时,展现出较强的鲁棒性。所提方法的有效性在公开数据集、动态物体干扰数据集和自采数据集上通过实验进行了验证。

二、提出一种基于不变扩展卡尔曼滤波(Invariant Extended Kalman Filter,InvEKF)的激光雷达-惯性定位与建图方法。本方法为激光雷达-惯性定位与建图设计了一种不变EKF状态估计器。考虑到系统会引入IMU偏置和传感器噪声,本方法利用不变观测器原理和李群理论,得到近似对数-线性的误差动力学方程,以改善系统中存在的非一致性问题。将该不变EKF状态估计器应用到激光雷达-惯性定位与建图中,本文分别实现了以机器人为中心(Inv-LIO1)和以世界为中心(Inv-LIO2)的两个独立系统,并对其可观性进行了理论分析。Inv-LIO1中局部坐标系下的状态在每帧激光雷达时间移动,以防止滤波器发散;Inv-LIO2直接在全局坐标系下融合激光雷达与IMU测量。此外,Inv-LIO1的里程计和建图模块分别运行在两个不同的线程中,而Inv-LIO2则采用一种集成里程计和建图模块的架构。在室内仿真数据集、室外公共数据集和自采数据集上的实验表明了所提方法的有效性。

三、提出一种基于多状态颜色一致性约束的激光雷达-惯性-视觉定位与建图方法。本方法紧耦合了激光雷达-惯性里程计(LiDAR-Inertial Odometry,LIO)和视觉-惯性里程计(Visual-Inertial Odometry,VIO)两个子系统,并定义了带有颜色信息的全局地图表示形式。LIO子系统中点云经过运动补偿后,直接用于构建点到面的残差;VIO子系统利用全局地图中点的深度信息,根据滑动窗口中多个相机状态观测到同一地图点颜色的不变性,构建光度误差约束;最后通过不变EKF状态估计器进行系统状态更新。本文还对系统的可观性进行了理论分析。所提方法在公共数据集和自采数据集上进行实验和鲁棒性测试,验证了方法的有效性。

Other Abstract

Simultaneous localization and mapping (SLAM) serves as a fundamental prerequisite for autonomous robot applications and has witnessed remarkable progress in recent years, with outstanding research results gradually transitioning from laboratories to industrial applications. The continuously emerging datasets and the diversity of real-world scenarios have imposed higher demands on the localization accuracy and robustness of SLAM systems. Certain complex environmental conditions, such as rapid illumination changes, texture-less regions, severe occlusions, scene structure degradations, and aggressive motions, can lead to performance degradation or even failure of SLAM systems. Particularly for algorithms relying on one single sensor modality, their inherent limitations make it challenging to adapt to diverse environments. The multi-sensor fusion SLAM have the potential to leverage the strengths of different sensors, thereby improving their adaptability and robustness in complex and diverse real-world scenarios. This paper focuses on the three sensor modalities of LiDAR, IMU, and cameras as information sources, and investigates visual-inertial, LiDAR-inertial, and LiDAR-inertial-visual localization and mapping tasks. The main research content and contributions of this paper are summarized as follows:

1. A visual-inertial localization and mapping method based on nonlinear factor estimation of pose graph optimization (PGO) is proposed. This method comprehensively considers the covariance estimation of sequential constraints (close in time) and loop-closing constraints (spatially similar) in the pose graph to more accurately represent the true system uncertainty. For sequential constraints, this method introduces absolute pose and relative pose factors, and utilizes a nonlinear factor extraction method to obtain the optimal covariance matrix from the accumulated VIO. For loop-closing constraints, this method introduces a scale factor and proposes a dynamic scale estimation method to approximate the scale of their covariances. Thanks to the proposed scale estimation algorithm, this method exhibits strong robustness when facing false positive loop-closure caused by dynamic object interference. The effectiveness of the proposed method has been verified through experiments on public datasets, dynamic object interference datasets, and self-collected datasets.

2. A LiDAR-inertial localization and mapping method based on the invariant extended Kalman filter (InvEKF) is proposed. This method designs an invariant EKF state estimator for LiDAR-inertial localization and mapping. Considering that the system introduces IMU biases and sensor noise, this method employs the invariant observer principle and Lie group theory to derive an approximate log-linear error dynamics equation, mitigating the inconsistency issues within the system. By applying this invariant EKF state estimator to LiDAR-inertial localization and mapping, two independent systems are implemented: one robo-centric (Inv-LIO1) and one world-centric (Inv-LIO2), with their observability properties theoretically analyzed. In Inv-LIO1, the state in the local coordinate frame shifted at each LiDAR frame time to prevent filter divergence. In contrast, Inv-LIO2 directly fuses LiDAR and IMU measurements in the global coordinate frame. Furthermore, the odometry and mapping modules of Inv-LIO1 run in two separate threads, whereas Inv-LIO2 adopts an architecture that integrates the odometry and mapping modules. Experiments on indoor simulation datasets, outdoor public datasets, and self-collected datasets demonstrate the effectiveness of the proposed method.

3. A LiDAR-inertial-visual localization and mapping method based on multi-state color consistency constraints is proposed. This method tightly couples the LiDAR-Inertial Odometry (LIO) and Visual-Inertial Odometry (VIO) subsystems and defines a global map representation form that incorporates color information. In the LIO subsystem, the motion-compensated point cloud is directly employed to construct point-to-plane residuals for optimizing the system state. The VIO subsystem directly utilizes the depth information of points in the global map and constructs photometric error constraints based on the invariance of the color observed by multiple camera states for the same map point within the sliding window. The system state is then updated through an invariant EKF state estimator. This paper also theoretically analyzes the observability  property of the system. The proposed method is experimentally validated and tested for robustness on public datasets and self-collected datasets.

Keyword同时定位与建图 多传感器融合定位 状态估计 视觉-惯性里程计 激光雷达-惯性里程计 不确定性估计
MOST Discipline Catalogue工学
Language中文
IS Representative Paper
Sub direction classification无人系统
planning direction of the national heavy laboratory环境多维感知
Paper associated data
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56544
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
施鹏程. 融合多源传感器信息的机器人定位与建图方法研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
施鹏程学位论文.pdf(33494KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[施鹏程]'s Articles
Baidu academic
Similar articles in Baidu academic
[施鹏程]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[施鹏程]'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.