With the characteristics of accurate 3D point cloud data and robustness to illumination change, 3D LiDAR provides the basis for high-quality localization and mapping. It is significant in both research and applications. This thesis concerns the research on 3D LiDAR localization and mapping. The main contents are as follows:
Firstly, the research background and its significance of this thesis are given. The research development of 3D LiDAR localization and mapping, and fused localization and mapping of 3D LiDAR with other sensors is reviewed. The content and structure of this thesis are introduced.
Secondly, aiming at the problem of low efficiency of scan-to-map matching in typical 3D LiDAR SLAM (simultaneous localization and mapping), a new method DGP-SLAM is proposed based on DGP (directed geometry point) and sparse frame. On one hand, DGP features are sparsely distributed in space, and the dense geometric point feature map is replaced by a sparse DGP feature map, which improves the efficiency of data association in the scan-to-map matching module. Besides, a strict data association strategy is designed to improve the accuracy of data association. On the other hand, sparse fusion frames further increase the sparsity of the map, and the DGP feature propagation strategy based on the sparse frame improves the quality of features in the map. Also, DGP feature-based loop detection and pose graph optimization are conducted for global consistency. The proposed method can effectively improve the efficiency of scan-to-map matching, and the experiments on the datasets verify the effectiveness of the proposed method.
Thirdly, a hierarchical 3D LiDAR odometry method HELO based on scan-to-map matching and fixed-lag smoothing is proposed. The high-level fixed-lag smoothing module takes the pose estimation result of the low-level scan-to-map matching module as its input. Through multi-frame pose optimization, the accumulated error is reduced and a high-precision pose is obtained in real time. Concretely, a feature-centric map feature management scheme is designed to manage features in both low-level and high-level modules, which builds the local map and provides multi-frame data associations for above modules, respectively. The results of data associations are then combined with the one-to-many measurement residual strategy to construct feature-wise multi-frame pose constraints, and orientation residuals are also introduced to improve the accuracy of pose optimization. The experimental results on datasets show that the proposed method can effectively reduce the accumulated error while ensuring real-time performance.
Fourthly, a nested 3D LiDAR odometry method TFAS with the tight fusion of filtering and smoothing is proposed. The filtering and smoothing modules are tightly integrated into two nested windows, and they share the features and poses in the windows. On one hand, the memory consumption is reduced, and on the other hand, the optimization result of the smoothing module provides feedback for the filtering module to improve its accuracy. In the filtering module, the strategy of the refined pose optimization is presented to enhance the robustness to outlier data association, and multi-residual constraint is designed to improve the pose optimization accuracy. Meanwhile, the DGP feature triangulation strategy is introduced in the smoothing module and the quality of features is improved. The effectiveness of the proposed method is experimentally verified on the datasets.
Fifthly, a LiDAR-inertial odometry based on the DGP feature is proposed. The sparse DGP features ensure the efficiency of the proposed odometry. The IMU measurement is used to remove the rotation and translation distortions of the point cloud and predict the state under the framework of multi-state iterative extended Kalman filter. As a result, the performance under severe motion is improved. The strategies of feature tracking based on the stable local map as well as the adaptive keyframe selection and culling are designed, which improve the direction estimation accuracy of DGP features in low-beam 3D LiDAR. Also, the redundant keyframes in the map are reduced. In addition, a LiDAR-inertial initialization strategy is presented to realize the system initialization with different motion states. The experimental results on the datasets verify the robustness of the proposed method.
Finally, the conclusions are given and future work is addressed.