结合通道优先卷积注意力的点云配准网络  

Point Cloud Registration Network Combining Channel Priority Convolution Attention

作　　者：张艺旋 韩燮[1,2,3] 赵融 王松 程郴[4] ZHANG Yixuan;HAN Xie;ZHAO Rong;WANG Song;CHENG Chen(School of Computer Science and Technology,North University of China,Taiyuan 030051,China;Shanxi Provincial Key Laboratory of Machine Vision and Virtual Reality,North University of China,Taiyuan 030051,China;Shanxi Province’s Vision Information Processing and Intelligent Robot Engineering Research Center,North University of China,Taiyuan 030051,China;China Institute for Radiation Protection,Taiyuan 030006,China)

机构地区：[1]中北大学计算机科学与技术学院,山西太原030051 [2]中北大学机器视觉与虚拟现实山西省重点实验室,山西太原030051 [3]中北大学山西省视觉处理及智能机器人工程研究中心,山西太原030051 [4]中国辐射防护研究院,山西太原030006

出　　处：《中北大学学报（自然科学版）》2024年第6期753-763,共11页Journal of North University of China(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62272426);山西省科技重大专项计划“揭榜挂帅”项目(202201150401021)。

摘　　要：点云配准网络在处理大规模点云和捕获局部细节特征时存在局限性,导致其对点云重叠区域配准的精度不足,本文提出了一种新的点云配准网络CR-RORNet来解决此问题。该网络结合了通道优先卷积注意力和ResPointNet模块,克服了现有方法在应对复杂场景和不规则点云时的不足,提升了复杂点云模型的配准效果。首先,在粗配准阶段设计了ResPointNet模块,通过引入残差连接机制强化了点云模型全局特征和多层次特征的提取与融合。其次,在动态图卷积神经网络中将通道优先卷积注意力机制CPCA(Channel Prior Convolutional Attention)与跨阶段梯度聚合机制进行结合,CPCA机制利用通道先验信息加强了网络对重要特征通道和区域的关注,在处理点云重叠部分时,能有效增强网络模型对点云局部细节特征的捕捉能力并抑制低置信度区域的影响,从而显著提升配准的效果;跨阶段梯度聚合机制融合了点云模型不同深度层次的梯度信息,确保在处理微小零部件或大范围场景点云模型时,网络能充分理解点云的结构和局部细节,并使学习到的特征具有良好的表达力,从而实现了复杂场景下点云数据的高精度配准。实验表明,CR-RORNet在自采数据集上的表现优于其他点云配准方法,相比基线RORNet,CR-RORNet在RMSE(t)误差降低了39.5%,在MSE(R)误差降低了5.1%。公开数据集ModelNet40中的实验结果表明,该网络具有良好的泛化性能。The limitations of point cloud registration networks in processing large-scale point clouds and capturing local detail features,resulted in insufficient registration accuracy for overlapping areas of point clouds,so a new point cloud registration network CR-RORNet was proposed.This network combined channel first convolutional attention and ResPointNet module,and overcame the shortcomings of existing methods in dealing with complex scenes and irregular point clouds,and improved the registration performance for point clouds with significant initial pose differences.Firstly,in the coarse registration stage,the ResPointNet module was designed to enhance the extraction and fusion of global and multi-level features of the point cloud model by introducing residual connection mechanism.Secondly,in the dynamic graph convolutional neural network,channel prior convolutional attention CPCA and cross stage gradient aggregation mechanism were combined.The CPCA mechanism utilized channel prior information to strengthen the network’s attention to important feature channels and regions.When dealing with overlapping parts of point clouds,it could effectively enhance the network model’s ability to capture local detail features of point clouds and suppress the influence of low confidence regions,significantly improving the registration effect;the cross stage gradient aggregation mechanism integrated gradient information from different depth levels of point cloud models,and ensured that the network could fully understand the structure and local details of the point cloud when dealing with small components or largescale scene point cloud models.And the learned features had good expressive power,so high-precision registration of point cloud data in complex scenes was realized.Experimental results show that CR-RORNet performs better than other point cloud registration methods on self collected datasets.Compared to baseline RORNet,CR-RORNet reduces RMSE(t)error by 39.5% and MSE(R)error by 5.1%.Experiment results on the publicly avail

关 键 词：点云配准 深度学习 注意力机制 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]