毕业生知识库

     视觉同时定位与地图构建 (Visual Simultaneous Localization and Mapping, VSLAM) 是近年来自主无人系统领域的研究热点，但单目 SLAM 在实际应用中不能提供绝对尺度的定位信息，且在快速运动、纹理较少或光照快速变化的场景中容易定位失败。因此，探索多传感器和视觉之间的信息融合方式以实现鲁棒而高精度的定位具有重要的研究意义和实用价值。本文在考虑各传感器特点，以及图像特征点和特征直线等纹理信息的基础上，开展了基于多传感器信息融合的视觉 SLAM 方法研究。论文的主要工作可总结如下：

     第一，针对单目 SLAM 无法估计绝对尺度和在弱纹理区域定位精度较低的问题，提出了一种基于直线特征结构化约束的双目~SLAM~系统，通过融合双目图像中的直线信息来解决弱纹理区域的鲁棒定位问题。首先，针对双目直线 SLAM 系统中存在的匹配速度慢的问题提出了基于直线相交点的快速匹配方法，极大地提升了直线匹配的速度。其次，针对人造环境中存在的大量平行直线，构建了一个基于平行直线约束的误差函数用于提升直线 SLAM 系统的定位精度。最后，在此基础上，搭建了基于直线特征的双目 SLAM 系统，并在 {\em it3f} 数据集上和众多开源算法进行了对比实验，所提方法获得了比其他方法精度更高的定位结果。

     第二，针对单目 SLAM 无法估计绝对尺度和相机快速运动下鲁棒性不高的问题，提出了一种基于深度信息辅助的 RGBD 视觉 SLAM 方法。首先，针对 RGBD 深度传感器的噪声特性，对深度传感器噪声如何影响基于直接法的视觉 SLAM 问题进行了研究，推导了深度噪声对光度误差函数不确定度的影响。以此为基础，分析出图像中靠近物体边缘的像素由于深度不确定大而对误差函数的影响会较大的结论，从而提出了基于边缘抑制的改进型 DVO (Dense Visual Odometry) 算法。在 TUM RGBD 数据集上对算法进行了评估，获得了比原 DVO 算法更精确的定位结果，验证了算法的有效性。

    第三，针对快速运动及光照变化环境下单目 SLAM 系统鲁棒性不高，以及点特征地图几何结构信息不丰富的问题，提出了一种基于滑动窗口图优化技术的紧耦合视觉惯导定位和建图方法。通过充分利用图像中的纹理信息，联合优化特征点和特征直线的重投影误差以及 IMU 预积分误差来获得系统轨迹和地图的最优估计。该方法是已知范围内的第一个图优化框架下融合点和直线特征的视觉惯导系统。在多个公开数据集上对该算法进行了评测，结果表明提出的融合点和直线特征的视觉惯导定位系统获得了比当前开源的其他方法更高的定位精度，验证了算法的有效性。

    第四，提出了一种融合机器人里程计信息和单目视觉信息的定位方法，用于解决基于单目系统的室内移动机器人鲁棒定位问题。首先，利用里程计和相机之间的刚体约束，构建了基于里程计信息的虚拟视觉测量用于约束不同时刻相机间的相对运动，并依据里程计测量模型推导了虚拟视觉测量的协方差矩阵。在此基础上，构建了一个用于标定相机和里程计外参数矩阵以及用于融合两传感器信息进行定位的因子图模型。最后，分别用仿真数据和机器人硬件平台对所述算法进行了验证，实验结果表明融合里程计信息的单目定位系统能提升单目视觉定位算法的精度及鲁棒性。

    最后对论文中的工作进行了总结，并讨论了可在其基础上进行的拓展工作。

Visual Simultaneous Localization and Mapping (VSLAM) is a research hotspot in recent years in primary unmanned systems. However, monocular SLAM cannot provide absolute-scale location information in practical applications, and location failures easily occur in scenes where there are fewer textures, changing illuminations, or rapid movements of cameras. Therefore, exploring information from multiple sensors and cameras to achieve robust and highly precise positioning is of great significance for research and has large practical values. In this paper, the visual SLAM method fusing multi-sensor information is studied according to characteristics of each sensor and the texture information such as points and straight lines in an image. The main contributions of this dissertation can be summarized as follows:

Firstly, in order to solve the scale problem of monocular SLAM and the problem of low accuracy in the region with fewer textures, a stereo SLAM system based on linear feature structuring constraints is proposed. The problem of low accuracy in the region with less texture is solved by merging the straight line information in a stereo image. A fast matching method based on the intersection of straight lines is proposed to increase matching speeds in the stereo-line-feature-based SLAM system. When a large number of parallel lines exist in an artificial environment, an error function based on parallel linear constraints is constructed to improve the positioning accuracy of the line-based SLAM system. A stereo SLAM system based on line features is built, and contrast experiments are conducted on it3f datasets using many open source algorithms. The proposed method obtains higher accuracy than other methods.

Secondly, in order to solve the scale problem of monocular SLAM and the problem of low robustness under fast motion, a depth-information-assisted RGBD visual SLAM method is proposed. Aiming at the noise characteristics of the RGBD depth sensor, how the depth sensor noise affects the visual SLAM problem based on the direct method is studied, and how depth noises affect the uncertainty of the photometric error function is deduced. Based on this, it is concluded that pixels near the edge of the object in an image will have a greater impact on the error function due to greater depth uncertainties, and an improved DVO (Dense Visual Odometry) algorithm with edge suppression is proposed. The algorithm is evaluated on the TUM RGBD dataset, and a more accurate positioning result than the original DVO algorithm is obtained, and the effectiveness of the algorithm is validated.

Thirdly, aiming at the problem of low robustness of the monocular SLAM system in the fast movement and the environment under changing illuminations and the lack of rich geometry information in a point feature map, a tightly-coupled visual inertial navigation system based on sliding window optimization is proposed. By making full use of the texture information in an image, the re-projection error of feature points and feature lines and the IMU pre-integration errors are combined to obtain an optimal estimation of the system trajectory and the map. To the best of our knowledge, this method is the first optimization-based visual inertial system fusing features of points and straight lines. The proposed algorithm is evaluated on multiple public datasets. Extensive experiment results show that the proposed visual inertial positioning system with feature fusions of points and straight lines obtains higher positioning accuracy than other current open-source methods and verifies the effectiveness of the algorithm.

Fourthly, a positioning method for integrating robot odometer information and the monocular visual information is proposed to solve the problem of robust positioning of indoor mobile robots based on the monocular system. Using the rigid body constraints between the odometer and the camera, a virtual visual constraint based on odometer information is constructed to constrain the relative motion among cameras at different times, and the covariance matrix of the virtual visual constraint is deduced based on the odometry measurement model. Based on this, a factorial model is built to calibrate the external parameters matrix between the camera frame and the odometer frame to fuse the two sensor information. The algorithm is validated with simulation data and the robot platform respectively. Experimental results show that the monocular positioning system which integrates the odometer information can improve the accuracy and robustness of the monocular visual localization algorithm.

Finally, conclusions of the research are given, and the future work is addressed.