英文摘要 | Localization is a crucial prerequisite for autonomous operations of robots. With the advantages of accurate distance measurement and robustness to illumination change, 3D LiDAR (Light Detection and Ranging) provides the basis for high-quality localization of robots. Therefore, the research on 3D LiDAR localization is significant in theory with widespread applications. This thesis conducts the research on localization based on 3D LiDAR. The main contents are as follows:
Firstly, the research background and significance of this thesis are given. The research developments of 3D LiDAR localization, 3D LiDAR-inertial localization, and 3D LiDAR place recognition are reviewed. The content and structure of this thesis are introduced.
Secondly, aiming at the contradiction between the accuracy and efficiency in distribution-based 3D LiDAR odometry, a new method HDLO based on tightly associated distribution and maximum likelihood estimation (MLE) is proposed. A sparse data association from source distribution to target distribution is designed. For each pair of associated distributions, the union of corresponding point sets is utilized to represent the reference distribution, which associates the matched distributions tightly and thereby improves the representation of local structures. On this basis, a cost function that takes the point-to-reference distribution distances as constraints is constructed using MLE, and a computational decoupling strategy is presented. Through the pre-calculation of distribution parameters, the optimization complexity is decoupled from the number of points. As a result, the accurate pose and real-time performance are both obtained. Meanwhile, the aforementioned data association and optimization scheme is generalized to the situation of multiple frames. With the distribution-based inter-frame cross constraints, multi-frame poses within the fixed-size window are jointly optimized. The cumulative error is then reduced and the local consistency of pose estimation is improved. The average relative translation and average relative rotation errors of HDLO on the KITTI dataset are 0.50% and 0.16 degrees/100 meters, respectively. This demonstrates the effectiveness of the proposed method.
Thirdly, a hierarchical tightly-coupled 3D LiDAR-inertial odometry based on distribution is proposed by introducing IMU into the aforementioned HDLO to improve the robustness and accuracy of localization, which is termed as HD-LIO. To solve the degradation problem of point cloud distribution constraints in the tight coupling process of IMU and LiDAR, a loss function that dynamically changes with the distribution parameters is designed according to the propagation of point cloud measurement noise in the point-to-distribution distance observation equation. By adjusting the loss term, the anti-degradation point cloud distribution constraints are generated, ensuring the stability of pose estimation under the aggressive motions. On this basis, filtering and smoothing are integrated to achieve localization in a hierarchical tightly-coupled manner. With the iterative extended Kalman filter, the low level constructs joint constraints of IMU prior and anti-degradation point cloud distribution to estimate the pose of the current frame in real time. The high level combines prior, IMU pre-integration, and point cloud observation constraints to perform fixed-lag smoothing on multi-frame poses. In this way, the efficiency is ensured and a coarse-to-fine odometry estimation with robustness is realized. The effectiveness of the proposed method is verified on the datasets. The average relative translation and average relative rotation errors of HD-LIO on the NC and ENC datasets are 0.49% and 1.97 degrees/100 meters, respectively, which proves the effectiveness of the proposed method.
Fourthly, a LiDAR place recognition method CSINet is proposed, which is based on range image and cyclic column-shift-invariant attention. The existing range image-based solutions resort to the invariance of NetVLAD for the robustness to point cloud rotation, which restricts the change of scales during feature extraction. To address this problem, a cyclic column-shift-invariant (CSI) attention is designed. It weights outputs of average pooling and maximum pooling on the input feature, which is then cyclically convoluted to generate attention map equivariant to the input. By weighting the input feature based on matrix multiplication, the invariance to the cyclic shift of range image along the column direction is achieved while capturing global contextual information, which provides a prerequisite for the variation of feature scales. Then, a multi-scale feature enhancement module based on spatial downsampling and spatial-channel mixing is presented to mine the information of different scales on the output features of CSI attention. This enhances the discriminability of the global feature vector and facilitates the place recognition performance. The experiment results on the NCLT, KITTI, and Ford datasets verify the effectiveness of the proposed method.
Fifthly, a 3D LiDAR-inertial localization and mapping framework ADLI-SLAM based on adaptive distribution is designed, which takes the aforementioned 3D LiDAR-inertial odometry HDLO as the frontend and combines the loop closure optimization backend consisting of the proposed LiDAR place recognition method CSINet and pose graph. For the frontend, a scene-adaptive distribution extraction strategy is added at its low-level filtering to guarantee sufficient distribution constraints and reinforce localization robustness in challenging environments such as stairs. Also, its high-level smoothing additionally considers point cloud distribution parameters, and thus implements the bundle optimization of pose and distribution by minimizing the determinant of the distribution covariance matrix in voxels. It further improves the estimation precision. The backend applies the place recognition method CSINet to detect the candidate loop closure frame corresponding to the latest point cloud in the fixed-lag smoothing, and generates loop closure factor after the consistent check. Further, the loop closure factor is combined with odometry factors from the smoothing to execute loop closure optimization within the global pose graph, which improves the global consistency of localization and mapping. The performance of the proposed method is verified on public datasets.
Finally, the conclusions are given and future work is presented. |
修改评论