Optimization of the Use of Spherical Targets for Point Cloud Registration Using Monte Carlo Simulation

Registrations based on the manual placement of spherical targets are still being employed by many professionals in the industry.However,the placement of those targets usually relies solely on personal experience without scientific evidence supported by numerical analysis.This paper presents a comprehensive investigation,based on Monte Carlo simulation,into determining the optimal number and positions for efficient target placement in typical scenes consisting of a pair of facades.It demonstrates new check-up statistical rules and geometrical constraints that can effectively extract and analyze massive simulations of unregistered point clouds and their corresponding registrations.More than 6×10^(7) sets of the registrations were simulated,whereas more than IOO registrations with real data were used to verify the results of simulation.The results indicated that using five spherical targets is the best choice for the registration of a large typical registration site consisting of two vertical facades and a ground,when there is only a box set of spherical targets available.As a result,the users can avoid placing extra targets to achieve insignificant improvements in registration accuracy.The results also suggest that the higher registration accuracy can be obtained when the ratio between the facade-to-target distance and target-to-scanner distance is approximately 3:2.Therefore,the targets should be placed closer to the scanner rather than in the middle between the facades and the scanner,contradicting to the traditional thought. Besides,the results reveal that the accuracy can be increased by setting the largest projected triangular area of the targets to be large.

Automatic Extraction of the Sparse Prior Correspondences for Non-Rigid Point Cloud Registration

Non-rigid registration of point clouds is still far from stable,especially for the largely deformed one.Sparse initial correspondences are often adopted to facilitate the process.However,there are few studies on how to build them automatically.Therefore,in this paper,we propose a robust method to compute such priors automatically,where a global and local combined strategy is adopted.These priors in different degrees of deformation are obtained by the locally geometrical-consistent point matches from the globally structural-consistent region correspondences.To further utilize the matches,this paper also proposes a novel registration method based on the Coherent Point Drift framework.This method takes both the spatial proximity and local structural consistency of the priors as supervision of the registration process and thus obtains a robust alignment for clouds with significantly different deformations.Qualitative and quantitative experiments demonstrate the advantages of the proposed method.

Point Reg Net: Invariant Features for Point Cloud Registration Using in Image-Guided Radiation Therapy

In image-guided radiation therapy, extracting features from medical point cloud is the key technique for multimodality registration. This novel framework, denoted Control Point Net (CPN), provides an alternative to the common applications of manually designed keypoint descriptors for coarse point cloud registration. The CPN directly consumes a point cloud, divides it into equally spaced 3D voxels and transforms the points within each voxel into a unified feature representation through voxel feature encoding (VFE) layer. Then all volumetric representations are aggregated by Weighted Extraction Layer which selectively extracts features and synthesize into global descriptors and coordinates of control points. Utilizing global descriptors instead of local features allows the available geometrical data to be better exploited to improve the robustness and precision. Specifically, CPN unifies feature extraction and clustering into a single network, omitting time-consuming feature matching procedure. The algorithm is tested on point cloud datasets generated from CT images. Experiments and comparisons with the state-of-the-art descriptors demonstrate that CPN is highly discriminative, efficient, and robust to noise and density changes.

Deep learning based point cloud registration:an overview

Point cloud registration aims to find a rigid transformation for aligning one point cloud to another.Such registration is a fundamental problem in computer vision and robotics,and has been widely used in various applications,including 3D reconstruction,simultaneous localization and mapping,and autonomous driving.Over the last decades,numerous researchers have devoted themselves to tackling this challenging problem.The success of deep learning in high-level vision tasks has recently been extended to different geometric vision tasks.Various types of deep learning based point cloud registration methods have been proposed to exploit different aspects of the problem.However,a comprehensive overview of these approaches remains missing.To this end,in this paper,we summarize the recent progress in this area and present a comprehensive overview regarding deep learning based point cloud registration.We classify the popular approaches into different categories such as correspondences-based and correspondences-free approaches,with effective modules,i.e.,feature extractor,matching,outlier rejection,and motion estimation modules.Furthermore,we discuss the merits and demerits of such approaches in detail.Finally,we provide a systematic and compact framework for currently proposed methods and discuss directions of future research.

Automated registration of wide-baseline point clouds in forests using discrete overlap search

Forest is one of the most challenging environments to be recorded in a three-dimensional(3D)digitized geometrical representation,because of the size and the complexity of the environment and the data-acquisition constraints brought by on-site conditions.Previous studies have indicated that the data-acquisition pattern can have more influence on the registration results than other factors.In practice,the ideal short-baseline observations,i.e.,the dense collection mode,is rarely feasible,considering the low accessibility in forest environments and the commonly limited labor and time resources.The wide-baseline observations that cover a forest site using a few folds less observations than short-baseline observations,are therefore more preferable and commonly applied.Nevertheless,the wide-baseline approach is more challenging for data registration since it typically lacks the required sufficient overlaps between datasets.Until now,a robust automated registration solution that is independent of special hardware requirements has still been missing.That is,the registration accuracy is still far from the required level,and the information extractable from the merged point cloud using automated registration could not match that from the merged point cloud using manual registration.This paper proposes a discrete overlap search(DOS)method to find correspondences in the point clouds to solve the low-overlap problem in the wide-baseline point clouds.The proposed automatic method uses potential correspondences from both original data and selected feature points to reconstruct rough observation geometries without external knowledge and to retrieve precise registration parameters at data-level.An extensive experiment was carried out with 24 forest datasets of different conditions categorized in three difficulty levels.The performance of the proposed method was evaluated using various accuracy criteria,as well as based on data acquired from different hardware,platforms,viewing perspectives,and at different points of time.The proposed method achieved a 3D registration accuracy at a 0.50-cm level in all difficulty categories using static terrestrial acquisitions.In the terrestrial-aerial registration,data sets were collected from different sensors and at different points of time with scene changes,and a registration accuracy at the raw data geometric accuracy level was achieved.These results represent the highest automated registration accuracy and the strictest evaluation so far.The proposed method is applicable in multiple scenarios,such as 1)the global positioning of individual under-canopy observations,which is one of the main challenges in applying terrestrial observations lacking a global context,2)the fusion of point clouds acquired from terrestrial and aerial perspectives,which is required in order to achieve a complete forest observation,3)mobile mapping using a new stop-and-go approach,which solves the problems of lacking mobility and slow data collection in static terrestrial measurements as well as the data-quality issue in the continuous mobile approach.Furthermore,this work proposes a new error estimate that units all parameter-level errors into a single quantity and compensates for the downsides of the widely used parameter-and object-level error estimates;it also proposes a new deterministic point sets registration method as an alternative to the popular sampling methods.

Disordered Multi-view Registration Method Based on the Soft Trimmed Deep Network

Compared with the pair-wise registration of point clouds,multi-view point cloud registration is much less studied.In this dissertation,a disordered multi-view point cloud registration method based on the soft trimmed deep network is proposed.In this method,firstly,the expression ability of feature extraction module is improved and the registration accuracy is increased by enhancing feature extraction network with the point pair feature.Secondly,neighborhood and angle similarities are used to measure the consistency of candidate points to surrounding neighborhoods.By combining distance consistency and high dimensional feature consistency,our network introduces the confidence estimation module of registration,so the point cloud trimmed problem can be converted to candidate for the degree of confidence estimation problem,achieving the pair-wise registration of partially overlapping point clouds.Thirdly,the results from pair-wise registration are fed into the model fusion to achieve the rough registration of multi-view point clouds.Finally,the hierarchical clustering is used to iteratively optimize the clustering center model by gradually increasing the number of clustering categories and performing clustering and registration alternately.This method achieves rough point cloud registration quickly in the early stage,improves the accuracy of multi-view point cloud registration in the later stage,and makes full use of global information to achieve robust and accurate multi-view registration without initial value.

OAAFormer:Robust and Efficient Point Cloud Registration Through Overlapping-Aware Attention in Transformer

In the domain of point cloud registration,the coarse-to-fine feature matching paradigm has received significant attention due to its impressive performance.This paradigm involves a two-step process:first,the extraction of multilevel features,and subsequently,the propagation of correspondences from coarse to fine levels.However,this approach faces two notable limitations.Firstly,the use of the Dual Softmax operation may promote one-to-one correspondences between superpoints,inadvertently excluding valuable correspondences.Secondly,it is crucial to closely examine the overlapping areas between point clouds,as only correspondences within these regions decisively determine the actual transformation.Considering these issues,we propose OAAFormer to enhance correspondence quality.On the one hand,we introduce a soft matching mechanism to facilitate the propagation of potentially valuable correspondences from coarse to fine levels.On the other hand,we integrate an overlapping region detection module to minimize mismatches to the greatest extent possible.Furthermore,we introduce a region-wise attention module with linear complexity during the fine-level matching phase,designed to enhance the discriminative capabilities of the extracted features.Tests on the challenging 3DLoMatch benchmark demonstrate that our approach leads to a substantial increase of about 7%in the inlier ratio,as well as an enhancement of 2%-4%in registration recall.Finally,to accelerate the prediction process,we replace the Conventional Random Sample Consensus(RANSAC)algorithm with the selection of a limited yet representative set of high-confidence correspondences,resulting in a 100 times speedup while still maintaining comparable registration performance.

Decoupled deep hough voting for point cloud registration

Estimating rigid transformation using noisy correspondences is critical to feature-based point cloud registration.Recently,a series of studies have attempted to combine traditional robust model fitting with deep learning.Among them,DHVR proposed a hough voting-based method,achieving new state-of-the-art performance.However,we find voting on rotation and translation simultaneously hinders achieving better performance.Therefore,we proposed a new hough voting-based method,which decouples rotation and translation space.Specifically,we first utilize hough voting and a neural network to estimate rotation.Then based on good initialization on rotation,we can easily obtain accurate rigid transformation.Extensive experiments on 3DMatch and 3DLoMatch datasets show that our method achieves comparable performances over the state-of-the-art methods.We further demonstrate the generalization of our method by experimenting on KITTI dataset.

Nearest Neighbor Sampling of Point Sets Using Rays

We propose a new framework for the sampling,compression,and analysis of distributions of point sets and other geometric objects embedded in Euclidean spaces.Our approach involves constructing a tensor called the RaySense sketch,which captures nearest neighbors from the underlying geometry of points along a set of rays.We explore various operations that can be performed on the RaySense sketch,leading to different properties and potential applications.Statistical information about the data set can be extracted from the sketch,independent of the ray set.Line integrals on point sets can be efficiently computed using the sketch.We also present several examples illustrating applications of the proposed strategy in practical scenarios.

ALGORITHM OF PRETREATMENT ON AUTOMOBILE BODY POINT CLOUD

As point cloud of one whole vehicle body has the traits of large geometric dimension, huge data and rigorous reverse precision, one pretreatment algorithm on automobile body point cloud is put forward. The basic idea of the registration algorithm based on the skeleton points is to construct the skeleton points of the whole vehicle model and the mark points of the separate point cloud, to search the mapped relationship between skeleton points and mark points using congruence triangle method and to match the whole vehicle point cloud using the improved iterative closed point （ICP） algorithm. The data reduction algorithm, based on average square root of distance, condenses data by three steps, computing datasets＇ average square root of distance in sampling cube grid, sorting order according to the value computed from the first step, choosing sampling percentage. The accuracy of the two algorithms above is proved by a registration and reduction example of whole vehicle point cloud of a certain light truck.

RANSACs for 3D Rigid Registration:A Comparative Evaluation

Estimating an accurate six-degree-of-freedom(6-Do F)pose from correspondences with outliers remains a critical issue to 3D rigid registration.Random sample consensus(RANSAC)and its variants are popular solutions to this problem.Although there have been a number of RANSAC-fashion estimators,two issues remain unsolved.First,it is unclear which estimator is more appropriate to a particular application.Second,the impacts of different sampling strategies,hypothesis generation methods,hypothesis evaluation metrics,and stop criteria on the overall estimators remain ambiguous.This work fills these gaps by first considering six existing RANSAC-fashion methods and then proposing eight variants for a comprehensive evaluation.The objective is to thoroughly compare estimators in the RANSAC family,and evaluate the effects of each key stage on the eventual 6-Do F pose estimation performance.Experiments have been carried out on four standard datasets with different application scenarios,data modalities,and nuisances.They provide us with input correspondence sets with a variety of inlier ratios,spatial distributions,and scales.Based on the experimental results,we summarize remarkable outcomes and valuable findings,so as to give practical instructions to real-world applications,and highlight current bottlenecks and potential solutions in this research realm.

Multi-view ladar data registration in obscure environment

Multi-view laser radar （ladar） data registration in obscure environments is an important research field of obscured target detection from air to ground. There are few overlap regions of the observational data in different views because of the occluder, so the multi-view data registration is rather difficult. Through indepth analyses of the typical methods and problems, it is obtained that the sequence registration is more appropriate, but needs to improve the registration accuracy. On this basis, a multi-view data registration algorithm based on aggregating the adjacent frames, which are already registered, is proposed. It increases the overlap region between the pending registration frames by aggregation and further improves the registration accuracy. The experiment results show that the proposed algorithm can effectively register the multi-view ladar data in the obscure environment, and it also has a greater robustness and a higher registration accuracy compared with the sequence registration under the condition of equivalent operating efficiency.

Improvement of iterative closest point with edges of projected image

Background There are many regularly shaped objects in artificial environments.It is difficult to distinguish the poses of these objects when only geometric information is used.With the development of sensor technologies,inclusion of other information can be used to solve this problem.Methods We propose an algorithm to register point clouds by integrating color information.The key idea of the algorithm is to jointly optimize the dense and edge terms.The dense term was built in a manner similar to that of the iterative closest point algorithm.To build the edge term,we extracted the edges of the images obtained by projecting point clouds.The edge term prevents the point clouds from sliding during registration.We used this loosely coupled method to fuse geometric and color information.Results The results of the experiments showed that the edge image approach improves precision,and the algorithm is robust.

异常点云干扰下的车身构件鲁棒性配准方法

点云配准是大型车身构件位姿参数测量的关键方法,但现有算法在大量异常点云干扰下难以配准至有效位姿,从而导致匹配失真,进而无法保证后续机器人作业质量。针对此问题,提出一种能够有效抑制异常点云干扰的车身构件鲁棒性配准算法——鲁棒函数加权方差最小化(RFWVM)算法。建立鲁棒函数加权目标函数,通过施加随迭代次数可变的动态权重来抑制配准过程中异常点云的影响,并由高斯-牛顿法迭代完成刚性转换矩阵的求解。以高铁白车身侧墙、汽车车门框为研究对象的试验结果表明,较经典的最近点迭代(ICP)算法、方差最小化(VMM)算法、加权正负余量方差最小化(WPMAVM)算法和去伪加权方差最小化(DPWVM)算法,所提出的RFWVM算法配准精度更高,能够有效抑制各种异常点云对配准结果的影响,并具有更好的稳定性和鲁棒性,能够有效实现各类车身构件点云的精确配准。

基于SIFT特征点提取的ICP配准算法

为解决传统迭代最近点(ICP)算法对点云配准的起始点对选择不佳而导致配准时间长、效率低的问题,提出一种基于尺度不变特征变换(SIFT)特征点提取的ICP点云配准算法(ST-ICP)。首先使用SIFT算法进行原始点云与目标点云的SIFT特征点提取,根据提取特征点完成快速点特征直方图(FPFH)特征运算,通过采样一致性初始配准算法(SAC-IA)搜索对应点对、求解变换矩阵,再进一步运用ICP算法进行点云精细配准。实验结果表明:与ICP算法相比较,ST-ICP算法的配准误差在迭代次数为5次时减小了1.019 cm,迭代次数为10次时减小了0.443 cm;在配准误差达到10^(-2) cm级别时,ST-ICP算法所用时间比传统ICP算法减少了12.829 s。ST-ICP算法优化了对应点对的选择,提升了配准精度和配准效率。

一种基于广义最大相关熵准则的运动平均方法

本文提出一种基于广义最大相关熵准则的运动平均方法,有效提高了多视角点云配准的效率和精度.首先,借助李代数优化框架,导出了残差与待求变量二者之间的线性显式关系,将该残差与广义最大相关熵准则结合,提出了考虑所有相对运动的变量解算模型;其次,借助半二次优化技术和交替方向乘子法,给出了模型的一种数值解算方法;最后,提出了一种具有自适应性的核宽选取方法,以使算法在相对运动集包含较多外点的情况下仍能获得较为准确的配准结果.以真实模型和环境数据集为对象开展了对比实验,基于双视角点云配准算法分别获得了5个测试点集在30%、20%点云面积重叠率阈值下的双视角配准结果,共形成10组相对运动集.其中,相对运动集误差随重叠率阈值的下降逐渐增加.所提方法在10组运动集上均获得了较佳的配准结果,实验结果表明本文所提方法有效提高了运动平均方法的计算精度与效率.此外,当运动集包含大量外点时,采用本文所提方法仍能获得较为准确的配准结果,表明本文所提方法具有较强的鲁棒性.

基于边缘卷积的点云配准网络

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。

基于RGB-D双目视觉的苗期玉米三维模型重构方法研究

以玉米幼苗为对象,研究基于RGB-D双目视觉的苗期玉米三维模型重构方法,实现了部分重构参数的优化。首先,针对目标进行固定步距角环绕图像采集,依据RGB图像中目标区域分割结果,对深度图像进行目标区域深度数据分割,并采用改进后的均值滤波对苗期玉米区域内深度数据孔洞进行自适应填充;其次,针对苗期玉米各角度的深度点云数据,采用先粗后精完成多角度点云配准与融合;最后,对比两种体素精简方法对点云的精简平滑效果,实现苗期玉米三维模型的重构。通过试验对比步距角对苗期玉米模型的重构效率与精度,结果表明:采用八叉树滤波精简效果较好,60°步距角建模误差最小,重构的模型与苗期玉米株高精度误差为4.4 mm,茎粗平均精度误差为0.62 mm,能满足苗期玉米的三维重构形态测量需求。

露天矿无人机遥感边坡地表形变提取方法研究

针对当前露天矿边坡监测过程中存在的设备留有监测死角、点位布控缺乏依据、地质隐患解译困难、无人机(Unmanned Aerial Vehicle,UAV)影像点云重构复杂度高等问题,提出了一种基于无人机遥感的边坡地表形变提取方法。首先,通过分析UAV激光点云与影像特点构建点云序列;其次,利用融合尺度不变特征变换(Scale Invariant Feature Transform,SIFT)与圆柱形邻域搜索的改进迭代最近点(Iterative Closest Point,ICP)算法,实现点云序列的精准高效配准,提高边坡形变提取精度;最终,借助数字高程模型(Digital Elevation Model,DEM)叠加分析与可视化,精准定位边坡重点形变区域,直观提取边坡形变位置和大小,并结合正射影像图像特征进行形变区域分析与解译。以南芬露天矿为工程应用实例,研究表明:边坡形变模型标准偏差为0.032 m,对比全球定位系统-实时动态差分(Global Positioning System-Real Time Kinematic,GPS-RTK)实测形变值,形变中误差为0.012 m,能够快速实现大尺度复杂边坡地表扫描与形变提取,从而为地质灾害隐患分析、盲区边坡形变监测与地面监测设备科学布控提供技术支撑。

多尺度特征融合的点云配准算法研究

现有点云配准算法提取的特征不够丰富,导致配准精度很难进一步提升。针对该问题,本文提出一种基于深度学习的多尺度特征融合点云配准算法。首先,利用EdgeConv提取多个不同尺度的特征,该特征能够保持局部几何结构特性;接着,引入非线性极化注意力对其输出特征进行筛选,从而提高特征信息的有效性;然后,将以上多尺度特征进行融合并再次利用EdgeConv提取其特征,从而提高特征的表达能力;在刚体姿态估计阶段,采用线性李代数处理旋转变换以充分挖掘点云中的变换信息;最后,根据配准过程中提取点云特征的变化,动态调整损失函数各组成部分的权重,获得更准确的模型预测结果。在ModelNet40数据集上进行实验,本文算法在训练集和测试集样本种类相同时的旋转误差为1.8267,位移误差为0.0010;在训练集和测试集的样本种类不相同时(泛化实验)的旋转误差为2.9794,位移误差为0.0010。实验结果表明,本文算法的配准精度相比当前主流算法均有提高且泛化性能较好。