同步定位与建图（SLAM）研究的是在陌生环境下，载体通过所搭载的传感器收集的数据估计出自身的位置和姿态，并建立起周围环境的地图，其也被广泛应用于机器人和自动驾驶等领域。基于激光雷达（LiDAR）的SLAM具有较高的精度和鲁棒性，是当前学界研究与工程应用的热点。然而，LiDAR SLAM在一些场景的实际应用中依然存在精度较低、显著漂移、鲁棒性不足等问题，为解决这些问题，进一步提升SLAM系统的精度和鲁棒性，本文进行了如下研究： 首先，针对经典的LiDAR SLAM方法LOAM因未能从测量原理层面区分观测质量的好坏而存在的室内性能较差的问题，本文提出了一种基于观测加权的LOAM方法（OW-LOAM）。OW-LOAM采用传统的扩展卡尔曼滤波SLAM的思想，对LOAM在位姿估计时构建的每一组点-线观测和点-面观测进行噪声方差估计，并采用噪声方差的倒数来对相应的观测进行加权。该方法在贝叶斯估计理论的支持下实现了观测质量的区分，能够抑制较差观测对系统精度的不利影响。 其次，针对众多LiDAR SLAM方法在室外长距离运行后都会面临的漂移问题，本文提出了一种基于时间加权观测（TWO）的低复杂度直接闭合回环的方法，帮助LiDAR SLAM降低自身漂移。TWO的基本是思想是SLAM在早期漂移较小，建立的地图也更为准确，因而在位姿估计时可以给关联到地图中老旧部分（OPOM）的观测分配更高的权重，以让其帮助回环闭合。再者，为克服纯LiDAR SLAM的局限性，本文提出了一种激光雷达/单目相机/惯性测量单元（IMU）完全紧耦合的里程计方法——FT-LVIO。FT-LVIO基于误差状态迭代卡尔曼滤波框架，将LiDAR和IMU的数据同步到图像帧时刻，以同时利用三种传感器的数据对滤波器进行更新。该方法能够充分利用传感器间的互补特性，提升位姿估计准确性，同时降低单传感器退化对系统的影响。 此外，本文在公开数据集和实验室便携式手持SLAM系统采集的私有数据集上对上述方法进行了大量的实验验证，结果表明所提出的方法均能明显提升SLAM系统的精度和鲁棒性。同时，本文还为手持SLAM系统开发了便于操作的用户图形界面和彩色建图功能，提升了设备的实用性。
Simultaneous Localization and Mapping (SLAM) aims at estimating the position and attitude of a vehicle with the data collected by the onboard sensors and establishing a surrounding map in the unfamiliar environment, which is also widely used in robotics and autonomous driving. SLAM based on Light Detection and Ranging (LiDAR) is usually accurate and robust, so LiDAR SLAM has become a hot topic in both academic research and engineering application. However, in some scenarios, LiDAR SLAM may have low accuracy, present a significant drift, or not be robust enough. To address these problems and further improve the accuracy and robustness of the SLAM system, this paper conducts the following research: First, the classic LiDAR SLAM method LOAM fails to distinguish the observation quality from the measurement principle, resulting in poor indoor performance. To solve the problem, this thesis proposes an observation-weighted LOAM method (termed OW-LOAM). OW-LOAM adopts the idea of traditional extended Kalman filter based SLAM, estimating the noise of each point-to-line and point-to-plane observation constructed by LOAM during the pose estimation and weighing each observation with the inverse of corresponding noise variance. Supported by Bayesian estimation theory, the proposed method realizes the distinction of observation quality and can help suppress the adverse effects of bad observations. Second, many LiDAR SLAM methods can present significant drift after a long journey in the outdoor environment. To address the problem, this thesis proposes a lightweight method based on time-weighted observations (TWO) for directly closing the loop to reduce the drift of LiDAR SLAM. The basic idea of TWO is that LiDAR SLAM is low-drift in the early time, and the map built at that time tends to be more accurate. Thus, higher weights can be assigned to the observations corresponding to the Old Parts Of the Map (OPOM) during the pose estimation, encouraging OPOM to help achieve a loop closure. Third, to overcome the weakness of LiDAR SLAM, this thesis proposes a fully tightly-coupled LiDAR/camera/IMU(Inertial Measurement Unit) fusion odometry method named FT-LVIO. Built atop the framework of error-state iterated Kalman filter, FT-LVIO synchronizes the data from LiDAR and IMU to the timestamps of images and simultaneously utilizes the data from three types of sensors to update the filter. This method can take full advantage of the complimentary characteristics of individual sensors, improve the accuracy of pose estimation, and reduce the effects of the degeneration of the single sensor for the system. In addition, to validate the aforementioned methods, we conduct extensive experiments on both the public datasets and the private datasets gathered by the portable handheld SLAM system of our laboratory. The results show that the proposed methods can significantly improve the accuracy and robustness of the SLAM system. Meanwhile, we develop a friendly graphical user interface and add the color mapping function for the handheld SLAM system, further enhancing its practicability.
