两点云的稳健配准 [英] Robust registration of two point clouds

问题描述

我需要找到两个3D点云之间的变换和旋转差异.为此，我正在看点云库，因为它似乎很理想.

在干净的测试数据上，我可以使用迭代最近点(ICP)，但是得到的结果却很奇怪(尽管我可能实施不正确).我有 pcl :: estimateRigidTransformation 工作，而且看起来更好，尽管我认为处理嘈杂的数据会更糟.

我的问题是:

两朵云会很吵，尽管它们应该包含相同的点，但还是会有一些差异.最好的解决方法是什么?

我是否应该在这两个云中找到相应的功能，然后再使用 estimateTransform ?还是应该查看 RANSAC 函数以删除异常值?比 estimateRigidTransform 更好的方法是 ICP 吗?

解决方案

设置健壮的点云注册算法可能是一项具有挑战性的任务，需要选择各种不同的选项，超参数和技术来正确获得强结果.

但是

如您所见，成对注册应通过不同的计算步骤来运行以达到最佳效果.单个步骤是:

1. 数据采集:将输入云和参考云输入到算法中.

2. 估算关键点:(兴趣点)是点云中具有以下特征的点:

   1. 它有一个明确的定义，最好是在数学上有根据的定义，
   2. 它在图像空间中具有明确的位置，
   3. 兴趣点周围的局部图像结构具有丰富的局部信息内容

   点云中的这些显着点非常有用，因为它们的总和构成了点云的特征，并有助于区分点云的不同部分.

     pcl :: NarfKeypointpcl :: ISSKeypoint3D<PointInT，PointOutT，NormalT＞pcl :: HarrisKeypoint3D<PointInT，PointOutT，NormalT＞pcl :: HarrisKeypoint6D<PointInT，PointOutT，NormalT＞pcl :: SIFTKeypoint<PointInT，PointOutT>pcl :: SUSANKeypoint<PointInT，PointOutT，NormalT，IntensityT> 

   详细信息: PCL关键点-文档

3. 描述关键点-功能描述符:在检测到关键点后，我们继续为每个关键点计算一个描述符.本地描述符"是点的本地邻域的紧凑表示.与描述完整对象或点云的全局描述符相反，局部描述符试图仅在点周围的局部邻域中类似于形状和外观，因此非常适合于用匹配表示它.(Dirk Holz等人)

     pcl :: FPFHEstimation<PointInT，PointNT，PointOutT>pcl :: NormalEstimation<PointInT，PointOutT>pcl :: NormalEstimationOMP<PointInT，PointOutT>pcl :: OURCVFHEstimation<PointInT，PointNT，PointOutT>pcl :: PrincipalCurvaturesEstimation<PointInT，PointNT，PointOutT>pcl :: IntensitySpinEstimation<PointInT，PointOutT> 

   详细信息: PCL功能-文档

4. 对应估计:下一个任务是找到在点云中找到的关键点之间的对应关系.通常，一个人利用计算出的局部特征描述符，并将它们中的每个特征描述符与另一点云中的对应特征进行匹配.但是，由于来自相似场景的两次扫描不一定具有相同数量的特征描述符，因为一个云可以拥有比另一个云更多的数据，因此我们需要运行一个单独的对应拒绝过程.

     pcl :: registration :: CorrespondenceEstimation<PointSource，PointTarget，标量>pcl :: registration :: CorrespondenceEstimationBackProjection<PointSource，PointTarget，NormalT，Scalar>pcl :: registration :: CorrespondenceEstimationNormalShooting<PointSource，PointTarget，NormalT，Scalar> 

5. 通讯拒绝:执行通讯拒绝的最常见方法之一是使用 PCL模块注册-文档

6. 变换估计:在计算了两个点云之间的鲁棒对应之后，使用绝对方向算法来计算应用于输入云的6DOF(6自由度)变换匹配参考点云.这样做有很多不同的算法方法，但是PCL包括基于奇异值分解的实现.>(SVD).计算出一个4x4的矩阵，该矩阵描述了匹配点云所需的旋转和平移.

     pcl :: registration :: TransformationEstimationSVD<PointSource，PointTarget，标量> 

   详细信息: PCL模块注册-文档

进一步阅读:

• PCL点云注册
• 通过点云库注册
• PCL-功能如何工作

I need to find the transformation and rotation difference between two 3D point clouds. For this I am looking at Point Cloud Library, as it seems ideal.

On clean test data I have Iterative Closest Point (ICP) working, but getting strange results (although I may have implemented it incorrectly). I have pcl::estimateRigidTransformation working, and it seems better, although I assume will deal worse with noisy data.

My questions are:

The two clouds will be noisy, and although they should contain the same points, there will be some discrepancies. What is the best way to deal with this?

Should I find corresponding features in the two clouds to begin with and THEN use estimateTransform? Or should I look at a RANSAC function to remove outliers? Is ICP a better way to go than estimateRigidTransform?

解决方案

Setting up a robust point cloud registration algorithm can be a challenging task with a variaty of different options, hyperparameters and techniques to be set correctly to obtain strong results.

However the Point Cloud Library comes with a whole set of preimplemented function to solve this kind of task. The only thing left to do is to understand what each blocks is doing and to then set up a so called ICP Pipeline consisting of these blocks stacked on each other.

An ICP pipeline can follow two different paths:

1. Iterative registration algorithm

The easier path starts right away applying an Iterative Closest Point Algorithm on the Input-Cloud (IC) to math it with the fixed Reference-Cloud (RC) by always using the closest point method. The ICP takes an optimistic asumption that the two point clouds are close enough (good prior of rotation R and translation T) and the registration will converge without further initial alignment.

This path of course can get stucked in a local minimum and therefore perform very poorly as it is prone to get fooled by any kind of inaccuracies in the given input data.

2. Feature-based registration algorithm

To overcome this people have worked on developing all kinds of methods and ideas to overcome bad performing registration. In contrast to a merely iterative registration algorithm a feature-based registration first tires to find higher lever correspondences between the two point clouds to speed up the process and to improve the accuracy. The methods are capsuled and then embedded in the registration pipleline to form a complete registration model.

The following picture from the PCL documentation shows such a Registation pipeline:

As you can see a pairwise registration should run trough different computational steps to perform best. The single steps are:

1. Data acquisition: An input cloud and a reference cloud are fed into the algorithm.

2. Estimating Keypoints: A keypoint (interest point) is a point within the point cloud that has the following characteristics:

   1. it has a clear, preferably mathematically well-founded, definition,
   2. it has a well-defined position in image space,
   3. the local image structure around the interest point is rich in terms of local information contents

   Such salient points in a point cloud are so usefull because the sum of them characterizes a point cloud and helps making different parts of it distinguishable.

   pcl::NarfKeypoint
   pcl::ISSKeypoint3D< PointInT, PointOutT, NormalT >
   pcl::HarrisKeypoint3D< PointInT, PointOutT, NormalT >
   pcl::HarrisKeypoint6D< PointInT, PointOutT, NormalT >
   pcl::SIFTKeypoint< PointInT, PointOutT >
   pcl::SUSANKeypoint< PointInT, PointOutT, NormalT, IntensityT >
   

   Detailed Information: PCL Keypoint - Documentation

3. Describing keypoints - Feature descriptors: After detecting keypoints we go on to compute a descriptor for every one of them. "A local descriptor a compact representation of a point’s local neighborhood. In contrast to global descriptors describing a complete object or point cloud, local descriptors try to resemble shape and appearance only in a local neighborhood around a point and thus are very suitable for representing it in terms of matching." (Dirk Holz et al.)

   pcl::FPFHEstimation< PointInT, PointNT, PointOutT >
   pcl::NormalEstimation< PointInT, PointOutT >
   pcl::NormalEstimationOMP< PointInT, PointOutT >
   pcl::OURCVFHEstimation< PointInT, PointNT, PointOutT >
   pcl::PrincipalCurvaturesEstimation< PointInT, PointNT, PointOutT >
   pcl::IntensitySpinEstimation< PointInT, PointOutT >
   

   Detailed information: PCL Features - Documentation

4. Correspondence Estimation: The next task is to find correspondences between the keypoints found in the point clouds. Usually one takes advantage of the computed local featur-descriptors and match each one of them to his corresponding counterpart in the other point cloud. However due to the fact that two scans from a similar scene don't necessarily have the same number of feature-descriptors as one cloud can have more data then the other, we need to run a seperated correspondence rejection process.

   pcl::registration::CorrespondenceEstimation< PointSource, PointTarget, Scalar >
   pcl::registration::CorrespondenceEstimationBackProjection< PointSource, PointTarget, NormalT, Scalar >
   pcl::registration::CorrespondenceEstimationNormalShooting< PointSource, PointTarget, NormalT, Scalar >
   

5. Correspondence rejection: One of the most common approaches to perform correspondence rejection is to use RANSAC (Random Sample Consensus). But PCL comes with more rejection algorithms that a worth it giving them a closer look:

   pcl::registration::CorrespondenceRejectorSampleConsensus< PointT >
   pcl::registration::CorrespondenceRejectorDistance
   pcl::registration::CorrespondenceRejectorFeatures::FeatureContainer< FeatureT >
   pcl::registration::CorrespondenceRejectorPoly< SourceT, TargetT >
   

   Detailed information: PCL Module registration - Documentation

6. Transformation Estimation: After robust correspondences between the two point clouds are computed an Absolute Orientation Algorithm is used to calculate a 6DOF (6 degrees of freedom) transformation which is applied on the input cloud to match the reference point cloud. There are many different algorithmic approaches to do so, PCL however includes an implementation based on the Singular Value Decomposition(SVD). A 4x4 matrix is computed that describes the rotation and translation needed to match the point clouds.

   pcl::registration::TransformationEstimationSVD< PointSource, PointTarget, Scalar >
   

   Detailed information: PCL Module registration - Documentation

Further reading:

• PCL Point Cloud Registration
• Registration with the Point Cloud Library
• PCL - How features work

这篇关于两点云的稳健配准的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！