基于边缘卷积的点云配准网络  

Point Cloud Registration Network Based on Edge Convolution

作　　者：鲍国[1,2] 刘思谋 许士彪 张秋昭 段浩然[2] BAO Guo;LIU Simou;XU Shibiao;ZHANG Qiuzhao;DUAN Haoran(Air Force Logistics Academy,Xuzhou 221000,China;School of Environment Science and Spatial Informatics,China University of Mining and Technology,Xuzhou 221116,China)

机构地区：[1]空军勤务学院,江苏徐州221000 [2]中国矿业大学环境与测绘学院,江苏徐州221116

出　　处：《金属矿山》2024年第9期167-174,共8页Metal Mine

基　　金：自然资源部国土卫星遥感应用重点实验室开放基金项目(编号:KLSMNR-G202222);徐州市重点研发计划(社会发展)-社会事业项目(编号:KC23295)。

摘　　要：地下巷道结构狭长且支道繁多,在地下巷道中获取的点云需要进行点云配准获得完整数据,传统的点云配准方法对点云初始位置要求高并且计算迭代次数多,在环境复杂且数据量巨大的地下巷道场景点云中配准效果不佳且计算缓慢。因此,基于深度学习技术,以PCRNet为基础并结合边缘卷积网络在局部特征提取中的优势,构建了一种基于边缘卷积的点云直接配准网络DGRNet,该网络在特征提取模块利用边缘卷积核对输入的点云进行特征提取,能更好地对三维点云的复杂特征变化和几何结构进行学习,提高了对场景局部特征的理解能力。试验结果表明:DGRNet网络在物体模型中对比其他网络在整体上有着更好的配准精度,并且在点云噪声影响下能够保持配准精度稳定,有着较好的鲁棒性;DGRNet在巷道点云配准场景中的4种误差均最小,并且对比PCRNet误差分别降低了19.0%、20.1%、24.2%、21.0%。由此可见,DGRNet网络能够进行高精度的点云配准,为复杂的地下巷道场景点云配准提供了一种新方法。The structure of underground roadway is narrow and there are many branches,and the point cloud obtained in underground tunnels needs to be registered to obtain complete data.Traditional point cloud registration methods require high initial position of the point cloud and have multiple calculation iterations,but the registration effect is poor and the calculation is slow in the complex environment and huge amount of data in the scenic spot cloud of underground roadway.Therefore,based on deep learning technology and PCRNet,combined with the advantages of edge convolutional networks in local feature extraction,a point cloud direct registration network DGRNet based on edge convolution is constructed.The network uses edge convolution to check the input point cloud for feature extraction in the feature extraction module,which can better learn the complex feature changes and geometric structure of 3D point clouds,improved understanding of local features in the scene.The experimental results show that the DGRNet network has better overall registration accuracy in object models compared to other networks,and can maintain stable registration accuracy under the influence of point cloud noise,with good robustness.The four error results of DGRNet in the roadway point cloud registration scene are all the smallest,and compared with PCRNet,the error results have decreased by 19.0%,20.1%,24.2%,and 21.0%,respectively.It is indicated that the DGRNet network can perform high-precision point cloud registration,providing a new method for point cloud registration in complex underground scenes.

关 键 词：点云配准 深度学习 三维激光扫描 巷道 PCRNet DGRNet 

分 类 号：P237[天文地球—摄影测量与遥感]