针对目前基于深度学习的低重叠度点云配准方法在学习全局点云场景后进行特征匹配时，忽略局部特征间作用的问题，提出了一种结合边缘特征增强的层次注意力点云配准方法。方法 首先，利用边缘自适应KPConv（edge adaptive KPConv, EAKPConv）模块提取源点云、目标点云特征，增强边缘特征识别能力。然后，利用空间差异注意模块（local spatial contrast attention module, LSCAM）聚合局部空间差异捕捉点云的几何细节，利用序列相似度关联模块（sequential similarity association module, SSAM）计算量化两点云间的相似分数，并利用相似分数引导局部匹配。最后，通过LSCAM模块与SSAM模块结合的层次化注意力融合模块（hierarchical attention fusion module, HAFM）整合局部、全局特征，实现全局匹配。结果 在室内场景点云配准数据集3DMatch和三维模型数据集ModelNet-40进行了对比实验，本算法在3DMatch和3DLoMatch上的配准召回率分别达到93.2%、67.3%；在ModelNet-40和ModelLoNet-40取得了最低的旋转误差（分别为1.417和3.141）和平移误差（分别为0.01391和0.072）。此外，本算法在推理效率上比REGTR算法减少了10ms左右。结论 本算法通过自底向上的层次化处理方式显著提升了有限重叠场景点云的配准精度，同时降低了推理时间。

Objective The study of low overlap point cloud registration is a significant obstacle in the realm of computer vision. Within the field of point cloud registration, specifically in the context of deep learning-based approaches, a notable difficulty occurs when confronted with situations where there is limited overlap between the point clouds. Following the acquisition of knowledge from global point cloud scenes for the purpose of feature matching, these methods frequently fail to consider the interactions among local features. Nevertheless, this lack of attention can greatly impede the efficiency of registration in settings where local feature interactions are vital for establishing precise alignment. The intricate interplay between local characteristics, which is crucial for accurately identifying and aligning partially intersecting point clouds, is not adequately represented. The lack of consideration for the interaction between local characteristics not only affects the reliability of point cloud registration in situations with limited overlap, but also restricts the use of deep learning methods in varied and intricate settings. Therefore, it is crucial to create techniques that can include the comprehension of local feature interactions into the deep learning framework for point cloud registration, particularly in situations when there is limited overlap. Methods The present study introduces a novel technique for aligning point clouds with low overlap, using the Edge Adaptive KPConv (EAKPConv) module to enhance the identification of edge characteristics. The integration of local and global features is effectively accomplished by the combination of the Hierarchical Attention Fusion Module (HAFM) and the Local Spatial Comparison Attention Module (LSCAM). Next, LSCAM exploits the attention mechanism"s capacity to consolidate information, enabling the model to give priority to connections with target nodes and position itself in greater proximity to the clustered center of mass. This allows the model to flexibly capture complex details of the point cloud. The SSAM system utilises a hierarchical architecture, in which each tier of local matching modules applies its own similarity metric to quantify the similarities between point clouds. The local features are subsequently modified and transmitted to the subsequent layer of attention modules, so establishing a hierarchical structure. The hierarchical structure allows the model to collect and merge the inputs from local matches at different scales and levels of complexity, resulting in the formation of global feature correspondences. In this model, the multilayer perceptron (MLP) to accurately find the ideal correspondences and successfully complete the alignment procedure. This innovative method greatly improves the accuracy and efficiency of alignment. Results This work provides empirical evidence supporting the improved efficacy of the suggested algorithm, as demonstrated by its consistent performance across multiple datasets. Notably, it achieved impressive registration recall rates of 93.2% on the 3DMatch dataset and 67.3% on the 3DLoMatch dataset. In the experimental evaluation conducted on the ModelNet-40 and ModelLoNet-40 datasets, the method achieved minimal rotational errors of 1.417 and 3.141 degrees, respectively. It also recorded translational errors of 0.01391 and 0.072, respectively. These outcomes highlight the method"s effectiveness in point cloud registration, demonstrating its capability to accurately align point clouds with low rotational and translational discrepancies. This finding indicates a significant enhancement in accuracy when comparing it to the REGTR approach. In contrast to REGTR, the approach presented in this study demonstrates significantly reduced inference times of 27.205 ms on the 3DMatch dataset and 30.991 ms on the ModelNet-40 dataset. The findings of this study emphasize the efficacy of the proposed methodology in effectively addressing the challenging issue of disregarding features in point cloud registration tasks with little overlap. Conclusion This article presents a novel point cloud matching technique that combines edge improvement with hierarchical attention. It proposes a new approach that integrates polynomial kernel functions into the EAKPConv framework to improve the identification of edge features in point clouds. The text discusses the HAFM, which is used to extract specific local information. The module aims to improve feature matching by using similarities in edge features. This approach successfully achieves a harmonious combination of local and global feature matching, hence enhancing the comprehension of point cloud data. By implementing a hierarchical analysis technique, the registration accuracy is greatly enhanced by accurately matching local-global information. Furthermore, increasing the cross-entropy loss function enhances the accuracy of local matching, hence reducing misalignments. This study assesses the algorithm"s performance on ModelNet-40, ModelelloNet-40, 3Dmatch, and 3DLomatch datasets. The results of the studies indicate that the method substantially enhances registration accuracy, particularly in difficult situations with limited data overlap. Furthermore, the method exhibits superior registration efficiency compared to standard approaches.