基于深度交互性多尺度感受野特征学习的三维点云配准网络  

Three-Dimensional Point Cloud Registration Network Based on Deep Interactive Multi-Scale Receptive Field Feature Learning

作　　者：周涵 王旭初[1,2] 袁越 Zhou Han;Wang Xuchu;Yuan Yue(College of Optoelectronic Engineering,Chongqing University,Chongqing 400040,China;Key Laboratory of Optoelectronic Technology&Systems(Chongqing University),Ministry of Education,Chongqing 400040,China)

机构地区：[1]重庆大学光电工程学院,重庆400040 [2]重庆大学光电技术与系统教育部重点实验室,重庆400040

出　　处：《光学学报》2024年第14期169-180,共12页Acta Optica Sinica

基　　金：重庆大学科技项目(H20200677)。

摘　　要：提出了一种基于深度交互性多尺度感受野特征学习的三维点云配准网络(DIM-RFNet),通过结合点云结构的上下文一致性寻找潜在相似结构特征以实现点云准确配准。首先通过点云的分组构建邻域片并提取邻域片特征,使用上下文编码-解码器来提取多尺度感受野特征,并降低内存开销和提高不同结构特征的区别度。为了使任务集中在重叠区域、减小异常值的不利影响,并进一步扩大多尺度特征感受野,DIM-RFNet构建一种新的重叠关系编码-解码器,通过预测重叠关系置信度得到特征表示、重叠得分、匹配得分。将该方法在ModelNet40、ModelLoNet、3DMatch、3DLoMatch和OdometryKITTI数据集上与相关方法进行比较,结果表明,该方法显著提高了特征提取能力,并在低重叠率场景中取得有竞争力的配准性能。Objective We aim to enhance the performance of point cloud registration tasks. In recent years, attention mechanisms have shown great potential in 3D vision tasks such as point cloud registration. Currently, the lack of depth interaction during the feature extraction stage may result in the loss of important latent similar structures, thus degrading the performance under low-overlap scenarios. To this end, we propose a 3D point cloud registration network called DIMRFNet based on deep interactive multi-scale receptive field features, which combines structural context consistency to identify latent similar structure features for efficient point cloud registration.Methods The proposed DIM-RFNet model includes two stages. In the coarse registration stage, the sampled point cloud is input into the neighborhood patch feature extraction module to obtain the neighborhood patches and feature information matrix. Then, the information is fed into the context structure encoder, embedding the neighborhood patches into a highdimensional space and aggregating different-scale features. These features are further input into a transformer to update the high-dimensional features. The context structure decoder continuously expresses the neighborhood patches and corresponding high-dimensional features using a multi-layer perceptron(MLP), ultimately outputting a set of key points and their dimension-reduced structural features. In the fine registration stage, the key points and features obtained during the coarse registration stage are input into the overlap relation encoder, which employs structural feature cross-attention and self-attention to predict pairs of points with overlapping relations, leading to an overlap relation confidence matrix. The top K pairs with the highest overlap relation confidence are selected and input into the overlap relation decoder, which outputs feature representation, overlap score, and match score.Results and Discussions Our method is extensively evaluated on synthetic datasets ModelNet40 and ModelLoNe

关 键 词：三维点云配准 注意力机制 结构特征 特征交互 重叠关系 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]