毕业生知识库

三维点云的配准旨在找到多个部分重叠但未知关联的点云之间的变换关系，其广泛应用于计算机视觉和机器人领域，是同步定位与地图构建（Simultaneous Localization and Mapping, SLAM）和三维重建中的核心技术。三维点云的配准主要包含点云对的配准和点云序列的配准。早期的三维点云配准方法大多是基于迭代最近点算法（Iterative Closest Point, ICP）类的方法，该类方法的基本思想是在空间中搜索源点云中的三维点在目标点云中的最近点，通过建立点之间的对应来求解变换矩阵。但是基于ICP类的方法严重依赖于初始值，鲁棒性不足。近年来，基于特征的点云配准算法由于具备较高的鲁棒性受到了越来越多的关注，其通过提取三维特征并建立特征关联来实现三维点云的配准。基于特征的三维点云配准研究已经取得了显著进展，然而其仍然存在一定不足和挑战，包括：基于特征的点云对配准方法未能充分考虑点云对之间的差异性，在低重叠率场景中的精度和鲁棒性不足，其中的外点剔除网络的精度和泛化性仍有待提升；激光SLAM中基于特征的点云序列配准算法的鲁棒性不足。针对上述问题，本文对基于特征的三维点云配准方法进行了深入研究，并取得了如下研究成果：

1. 提出了一种面向点云对配准的基于交叉注意力机制的点云特征提取网络。现有方法学习点云描述子时往往只考虑从一帧点云中提取对旋转平移和密度具有一定程度不变性的描述子，而忽略了点云对之间的帧间差异。因此，本文提出了一种新颖的基于交叉注意力机制的网络，该网络允许特征随着点云对之间的差异而变化。其通过考虑点云对之间的信息来学习对于点云对区分性和不变性更强的特征描述子。对于输入的点云对P₁和P₂, 首先通过两个共享权重的编码器模块获得其自身特征。其次，同一层编码器输出的特征经过提出的注意力模块获得条件特征。最后，P₁的自身特征跟条件特征通过解码器模块得到P₁最终对应的融合特征。在公开数据集上的实验结果表明本文提出的算法在三维特征匹配和点云对配准上相比之前的方法取得了更高的精度。

2. 提出了一种面向点云对配准的高精度强泛化性的外点剔除网络。为了提升外点剔除网络的精度以及泛化性，本文提出了一种基于特征一致性种子关联选取和二阶特征一致性损失函数的外点剔除网络。对于输入的关联集，所提网络先通过一个非局部模块来将每个关联嵌入高维的几何特征，然后根据这些特征得到与其他关联相关性较高的关联作为种子关联，接着通过二阶空间一致性以及构建的特征找到每个种子对应的一致集，并得到最终的配准结果。此外，提出了一种基于二阶特征一致性的损失函数来训练这个网络。为了验证算法的有效性，将提出的算法在室内外场景的多个数据集上进行了测试，实验结果表明提出的方法取得了更高的精度和更强的泛化性。

3. 提出了一种基于点云分块匹配的低重叠率场景下的点云对配准算法。现有的针对低重叠率场景的方法主要聚焦于特征提取，但是其精度和鲁棒性不足。为了解决此问题，本文提出了一种新颖的框架来实现低重叠率场景下的点云配准算法，该算法基于提取得到的特征，通过点云分块，块匹配和最优变换的选择与优化得到最终的结果。在公开数据集上的实验结果表明了在不同的特征下，提出的方法相比之前的方法在低重叠率场景下均取得了更高的精度。同时，实验结果也显示了在非低重叠率场景下，提出的算法也取得了有竞争力的配准精度。

4. 提出了一个基于主成分分析 （Principal Component Analysis, PCA）特征提取和两阶段点云匹配的激光SLAM方法。目前激光SLAM中点云序列配准算法的精度，效率和鲁棒性难以很好地平衡，且基于特征的方法提取的特征不能很好地适应视角的变化。为了解决上述问题，本文提出了一个新的激光SLAM方法。对于输入的激光点云序列，首先设计了一个基于PCA的近似曲率估计算法来提取更加鲁棒的特征点，这些特征点包含三维点云中的边缘点和平面点；继而基于提取的特征点实施一个两阶段点云匹配策略，来提高定位的精度和建图的连续性。此外，使用一个基于正态分布变换（Normal Distributions Transform, NDT）的回环检测算法来减小系统在大场景中的累计误差。在公开数据集和自采数据集上的实验结果显示了提出算法的有效性。

Registration of 3D point clouds aims to find the transformation relationship between multiple partially overlapping point clouds with unknown correspondences. It is widely used in the fields of computer vision and robotics, which is the core technologies in Simultaneous Localization and Mapping (SLAM) and 3D reconstruction. Registration of 3D point clouds mainly includes registration of a pair of point clouds and registration of a point cloud sequence. Early registration methods of 3D point clouds are mainly based on Iterative Closest Point (ICP), the core idea of these methods is to search for the spacial closest point of each point from the source point cloud in target point cloud, and solve the transformation matrix by establishing correspondence between these points. However, the ICP-based methods heavily rely on initial values and lack robustness. In recent years, feature-based registration of point clouds has gained more attention for their high robustness, which achieves registration of point clouds by extracting 3D features and establishing feature correspondences. Significant progress has been made in the research of feature-based registration of 3D point clouds, but there are still certain shortcomings and challenges. Feature-based point cloud pair registration does not fully consider the differences between point cloud pair, lacks accuracy and robustness in low-overlap scenes, and the accuracy and generalization ability of the outlier rejection network in it still need be improved. The robustness of the feature-based point cloud sequence registration in LiDAR SLAM is insufficient. To address these problems, this thesis carries out an in-depth research on feature-based registration of 3D point clouds, and the following research results have been achieved:

1. This thesis proposes a cross-attention-based feature extraction network for registration of a pair of point clouds. Most existing methods focus on extracting descriptors with varying degrees of invariant to rotation, translation and density while neglecting the differences within point cloud pair. Thus, this thesis proposes a novel cross-attention-based network, which allows features to vary with differences within point cloud pair. It learns the feature descriptors considering information on both point clouds, and enhances the discriminability of the learned descriptors under the circumstance of matching these two point clouds. The input pair of point clouds, P₁ and P₂, are passed through two weight sharing encoder modules to obtain their self-features. Then the self-features from the same layer in encoder are passed through the proposed cross-attention module to obtain conditioned features. Finally, the self-features of P₁ and conditioned features are fused through the decoder module to obtain the final fused features of P₁. Experimental results of both 3D feature matching and point cloud pair registration on public datasets show that the proposed algorithm presents the superior performance over the previous methods.

2. This thesis proposes an accurate outlier rejection network with higher generalization ability for registration of a pair of point clouds. In order to improve the accuracy and generalization ability of the outlier rejection network, this thesis proposes an outlier rejection network with a seed selection method based on feature consistency and a loss function based on second order feature consistency. For input correspondences, each correspondence is embedded into a high-dimensional geometric feature space through a nonlocal module based on spatial consistency. Then the obtained features are passed through a seed selection module to obtain the seed with high correlation with other correspondences. After this, second order spatial consistency and the constructed features are utilized to find the consistent set corresponding to each seed, and obtain the final result. In addition, a new loss function based on the second order feature consistency is proposed. To prove the effectiveness of the algorithm, the proposed algorithm is tested on multiple datasets in indoor and outdoor scenes. Experimental results in both accuracy and generalization ability show that the proposed network presents the superior performance.

3. This thesis proposes a low-overlap point cloud pair registration algorithm based on point cloud blocks matching. Existing methods for low-overlap scenes mainly focus on feature extraction, but their accuracy and robustness are not enough. To solve the problem, this thesis proposes a novel framework to realize low-overlap point cloud registration, which obtain the final result through point cloud partitioning, block matching, outlier rejection and best transformation selection and refine. Experimental results on public dataset show that the proposed algorithm presents the superior performance over the previous methods on low-overlap scenes under different features. The experimental results also show that the proposed algorithm achieves competitive performance in non low-overlap scenes.

4. This thesis proposes a LiDAR SLAM with PCA-based feature extraction and two-stage point cloud matching. Accuracy, robustness and efficiency of point cloud sequence registration in LiDAR SLAM are not well balanced, and there has been not a good feature can adopt well to viewpoint changes. To solve the above problems, this thesis proposes a new LiDAR SLAM. For the input laser point cloud sequence, a PCA-based curvature evaluation method is proposed to extract the edge and planar features firstly. Then based on the extracted features, a two-stage matching of point clouds is used to improve the localization accuracy and mapping consistency. In addition, a NDT-based loop detection is utilized to reduce the cumulative error of the proposed system in large-scale scenes. The experimental results on both public datasets and custom datasets demonstrate the effectiveness of the proposed algorithm.