【作者】 周汝琴；

【导师】 江万寿；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2022， 博士

【摘要】 随着激光雷达(Light Detection and Ranging,Li DAR)设备的轻量化和低成本化,三维激光点云已广泛应用在自动驾驶、机器人、测绘遥感、医疗、工业设计、文物保护等众多领域。然而,在对实物进行数据采集时,许多因素(如投影盲点、视觉死区和复杂曲面)决定了无法通过单个测量设备单次完成整个目标/场景的数据采集。为了获取目标/场景完整的三维点云数据,需要对不同视角下采集到的三维点云进行旋转平移至同一个坐标系下,拼接成一个完整的三维点云,这就是点云配准。点云配准作为三维点云数据处理的一项基础任务,在三维重建、目标识别与跟踪、地图构建与定位、物体抓取等应用中起着重要的作用。本文研究的重点是地面激光点云刚性配准算法。它的核心是根据两个局部重叠的点云重叠区域的特征信息来寻找点对应关系或计算刚性变换参数。然而,在实际的应用中,地面激光点云配准任务常常面临着各种各样的困难,如噪声、非重叠区域、密度变化、对称目标或重复结构,这给现有的点云刚性配准算法带来了极大的挑战。为了有效解决上述问题,本文从传统方法和深度学习方法两方面出发,对点云配准算法展开了一系列研究,具体如下:首先,针对二维投影描述符信息压缩、三维体素描述符对密度和边界变化敏感的问题,提出了一种基于体素化缓冲加权二值描述符(Voxel-based Buffer-weighted Binary Descriptor,VBBD)的配准方法。该描述符具有以下优点:(1)直接获取三维信息,避免了二维投影带来的邻域信息压缩,具有更好的描述性;(2)通过计算缓冲区加权密度对体素进行二值编码,提高了描述符对边界效应、噪声和密度变化的鲁棒性。在此描述符基础上,结合KM(Kuhn-Munkres)算法和RANSAC(Random Sample Consensus)算法分别进行特征匹配和变换参数求解。在小目标、室内和室外场景数据集的一系列实验表明,基于VBBD的粗配准方法能有效地提高配准效率与精度,并具有较强的鲁棒性;其次,考虑到配准有效信息主要集中在重叠区域而非重叠区域信息导致点云配准精度大幅度降低的问题,提出了一种基于注意力机制的重叠区域检测网络Mask Net++。该网络通过计算两个局部重叠点云的二进制掩码向量,同时获得这两个局部重叠点云中描述同一场景/目标相同几何结构的内点(即重叠区域)。在合成的小目标数据集和真实的室外场景数据集的一系列实验表明,Mask Net++能够同时高精度地检测两局部重叠点云的重叠区域。Mask Net++可应用于:(1)点云去噪,高精度地滤除大量的随机噪声;(2)点云配准,作为局部重叠点云配准任务的预处理步骤,将局部重叠点云配准转化为完全重叠点云配准,从而极大提高配准的精度;最后,针对现有的基于深度学习的点云配准方法在局部重叠点云配准任务中表现较差的问题,受重叠区域检测网络Mask Net++的启发,提出了一个基于重叠区域信息交互的姿态回归网络SCANet。该网络在全局特征提取子模块中引入空间自注意力聚合模块(Spatial Self-Attention Aggregation,SSA),有效利用点云不同层次的全局信息;在姿态估计子模块中采用通道互注意回归模块(Channel Cross-Attention Regression,CCR),实现源点云和目标点云重叠区域信息交互的同时,逐步回归到姿态参数。该网络参数较少,直接输出配准变换参数。它是完全可微的,直接处理点云,不涉及点对应关系解算,能有效避免对称目标和重复结构的匹配二义性。在合成的小目标数据集和真实的室外场景数据集的一系列实验表明,与现有的传统配准方法和基于深度学习的配准方法相比,本文提出的顾及重叠区域信息的SCANet在局部重叠点云配准的精度和效率上都达到了先进的水平。更多还原

【Abstract】 With a rapid development of Li DARs(Light Detection and Ranging),3D point cloud has been widely applied in many fields,such as automatic driving,robot,remote sensing,medical treatment,industrial design,and cultural relics protection.However,during data acquisition,many factors,such as projection,visual blind spot and complex surface,determine that the data acquisition of the whole object/scene is impossibly completed by only one single measurement.In order to obtain the complete 3D point cloud of the object/scene,it is necessary to align the3 D point clouds collected from different perspectives to the same coordinate system through rotation and translation,and splice them into a complete 3D point cloud.Such above process is called point cloud registration.As a basic task of 3D point cloud data processing,point cloud registration plays an important role in 3D reconstruction,object recognition and tracking,mapping and localization,object capture,and other applications.The focus of this paper is on the rigid registration of ground-based Li DAR point cloud,which mainly depends on the information of the overlapping area between two partially overlapped point clouds to find correspondences or calculate the rigid transformation parameters.However,in fact,rigid registration of ground-based Li DAR point cloud often faces various difficulties,such as noise,non-overlapping areas,density variation,symmetrical objects or repeating structures,which brings great challenges to existing registration algorithms.To effectively solve above problems,this paper carried out a series of research on the point cloud registration in the aspects of traditional methods and deep learning methods.Details are as follows:Firstly,to deal with the information loss of 2D projection descriptors and the sensitivity to density variation and boundaries of 3D voxel descriptors,a voxel-based buffer-weighted binary descriptor(VBBD)based registration method is proposed.This descriptor has the following advantages:(1)3D information can be directly obtained,which can avoid the information compression caused by 2D projection,so that it has better descriptiveness;(2)the voxel is binarized by the weighted density of the buffer,which improves the robustness to boundary,noise and density variation.On the basis of this descriptor,a KM(Kuhn-Munkres)algorithm and a RANSAC(Random Sample Consensus)algorithm are combined for feature matching and parameters solving,respectively.A set of experimental results on object dataset,indoor dataset and outdoor dataset show that this VBBD based registration method can effectively improve the computational efficiency and is of high robustness;Secondly,considering that effective information of registration is mainly concentrated in the overlapping area and the points in non-overlapping areas greatly decrease the registration accuracy,an attention-based overlapping area detection network,called Mask Net++,is proposed.By simultaneously calculating two binary mask vectors of two partially overlapped point clouds,the network can obtain the inliers(overlapping areas)describing the same geometry of the same object/scene at the same time.A set of experiments on synthetic object dataset and real-world outdoor dataset demonstrate that the Mask Net++ can detect overlapping areas in high accuracy.Mask Net++ can be widely applied in(1)point cloud denoising,which can effectively filter a large number of random noise in high accuracy;(2)point cloud registration,which can greatly improve the performance of partial-to-partial registration by acting as a preprocess to estimate overlapping areas when deal with partial point clouds;Finally,to deal with the problem that the existing deep learning based registration methods perform poorly in partial-to-partial point cloud registration,inspired by Mask Net++,a pose regression network based on overlapping area information interaction,named SCANet,is proposed.The spatial self-attention aggregation module(SSA)is introduced into the feature extraction subnetwork to effectively extract the global information of different levels,and the channel cross-attention regression module(CCR)is used in the pose estimation subnetwork for the information exchange between the two global features while pose regression.The network has few parameters,directly outputs the transformation parameters.It is fully differentiable,directly processes point clouds,and does not involve the solution of point correspondence,which can effectively avoid the matching ambiguity of symmetrical objects and repeated structures.A set of experiments on synthetic object dataset and real-world outdoor dataset show that the SCANet with overlapping area information interation achieves the state-of-the-art performance in accuracy and efficiency compared with the existing traditional registration methods and deep learning based registration methods.更多还原