当前深度图像配准方法介绍
2017-06-23 09:59
429 查看
转自：http://blog.csdn.net/daydayyup/article/details/21381239
本文主要分析了A review of recent range image registration methods with accuracy evaluation这篇文章中所介绍的各种粗匹配和精确匹配的方法
l A review of recentrange image registration methods with accuracy evaluation
Abstract
l Main work:A survey of the most common techniques
l Two problems:
l Obtain a rough registration to get a motion between different views
l Obtain a fine registration from an initial approximation
Introduction
l Applications:Reverse engineering,mould fabrication,artifact reproduction,etc.
l Range finders:Time-of-flight lasers,laser scanning,stereovision,pattern projection
Classification of methods
l Goal:Find the Euclidean motion between a set of range images.
l In terms of whether initial info. is required,we divide the process into two methods:coarse registration and fine registration.
Coarse registration
Methods:
l point signature
l spin image
l PCA
l RANSAC-based Darces
l Line-based
l Genetic Alg
l Principal Curvature
Criteria:
l the kind of correspo-ndences used
l the method used to compute the motion
l robustness
l registration strategy
Fine registration
Criteria:
l the registration strategy
l the use of an efficient search method
l the way of computing the motion in each iteration
l robustness
Methods:
l ICP
l Chen
l Signed Distance Fields
l Genetic Alg
l Registration strategy:multi-view;pair-wise
l The use of k-d tree to speed up or not
l The distance to minimize:point-to-point (problem:a given point in the first view may not exist in the second view);point-to-plane (theoretically less iterations)
Coarse registration
l The most common correspondences are points, curves,and surfaces.
Point signature(P-Ps)
l Point signature is a point descriptor used to search for correspondences.
l Advantage:the matching process is very fast and efficient
l Main drawback:the intersection of a sphere to the surface is not very easy,espectially when the surface is represented as a cloud of points or a triangulated surface(interpolation);sensible to noise
Spin image(P-Ps)
l Some spin image are computed in the first view and search for the best correspondences in the second view,then remove the outliers.
l Advantage:good results;the spin image feature is very robust except in case of symmetries or repeated regions.
l Main problem:the spin image strongly depends on the resolution of the method (solution:face-based spin image,interpolation) .
Principal component analysis(P-P)
l Using the direction of the main axis of the volu-me given by the cloud of points to align the sequence of range images.
l Advantage:this method is very fast.
l Drawback:it can only be used with effectiveness when there tis a sufficient number of points.In practice ,a 50% overlapping is critical.It's also limited in coping with surfaces that contain symmetries.In most cases this method is far from the expected.
Ransac-based darces(P-Ps)
l This method is based on finding the best three point-correspondences between two range ima-ges to obtain an estimation of the Euclidean motion.Normal vector and the principal curvatu-re is also used.
l The result obtained is very good because of its robustness even on the presence of outliers. It can only be used when the number of points in each view is relatively small,and where time is not critical.
Algebraic surface model(S)
l This method estimate the motion between sufraces represented as a polynomial model.It's based on obtaining a function of the distance between the polynomial model and the points.
l No need of points or curve correspondences.It's faster than others.A normal vector at each point is required.Good results need less than 15% of non-overlapping region.
Line-based algorithm
l This alg. proposed to use (straight or curve)lines to find pairs of correspondences.
l It is applied to large and structured environments in which planar regions and straight lines can be easily found.
l It obtains good results and the main drawback is the difficulty the segment the straight segme-nts as well as the supervisor.
l Improvement:Matching bitangent curves
l The main idea is that all bitangent points are co-incident in the dual space.
l Advantage:The range image is previously trans-formed into the dual space before the search for possible matches starts,which decreases the computing time and increases the robustness.
Genetic algorithm
l Defining a vector that contains the n index of correspondences between both range images. Also a fitness function is required to measure the quality of each potential solution.
l In most genetic approaches,the results obtained are quite good but the computing time is expen-sive,specially in the presence of a large number of points.
Principal curvature
l This approach proposed that use of the differe-ntial of points characterized by the principal curvature (max & min)in the matching process.
l The result is considered to be a good estimation. The main problem is that it is not robust.As the correspondence is not validated,it might be a false matching.
Fine registration methods
l The term fine registration use an initial guess to iterate and converge to a more accurate solutio-n,in which a distance function is minimized.
ICP
l The goal is to obtain an accurate solution by minimizing the distance between point-corresp-ondences,known as closest point.
l ICP obtains good results even in the presence of Gaussian noise.But it can not cope with non-overlapping regions because the outliers are never removed.The convergence is not guaran-teed.
l Modifications:
l Nearest Neighbor Problem is applied to facilita-te the search of closest points.This method dec-reases the computing time drastically.
l Muti-resolution Scheme ICP:It's for fast registra-tion,whose main idea is to solve the first few iterations using down sampled points and to pr-ogressively increase the resolution.
l ICP using invariant features(ICPIF):The matchi-ng process uses a weighted feature distance. Different invariant features were proposed,such as curvature,moment,spherical harmonics,amo-ng which the spherical harmonics provided the best convergence rate.
l Color and curvature features:As color is used as a matching constraint,symmetric objects can be well registered.
l RICP method making use of the Least Median of Squares:Aiming at finding a registration with-out outliers.It's robust!
l Picky ICP:It is based on outlier thresholding.The main difference with respects to the previous methods is that at every iteration only the pairs of correspondences with the smallest distances are used in the motion compution.
Method of Chen
l This is an alternative to the ICP algorithm,which was based on minimizing the distance between points and planes.
l To improve the precision of the algorithm,the Normal Space Sampling is proposed.The main idea is to select more points in the region where the normal is different from the other parts of the surface.
l This method is more robust to local minimum, and in general,results are better.Despite no robust method are applied ,this method is less sensible in the presence of non-overlapping regions.
l Chen's approach usually requires less iterations than the ICP approach.
Matching signed distance fields
l The method is a robust one so the outliers are removed,and all the views of a given object are registered at the same time(muti-view).
l Advantage:It registers all the views at the same time,which implies a more robust solution and avoids the error propagation problem present in pair-wise registration methods.And the final result is directly an integrated model.
l Drawback:It cannot be used in real time.
Genetic algorithm
l The goal of this method is to find a chromoso-me composed of the 6 parameters of the motion that aligns a pair of range images accurately.
l A similar method related to the original one has a more robust fitness function and the initial guess is not required.
l Overall,genetic alg. could avoid local minima especially when the initial motion is not provid-ed or it is given with low precision.It also works well in the presence of noise and ouliers given by non-overlapping regions.
l The main problem is the time required to converge.
本文主要分析了A review of recent range image registration methods with accuracy evaluation这篇文章中所介绍的各种粗匹配和精确匹配的方法
l A review of recentrange image registration methods with accuracy evaluation
Abstract
l Main work:A survey of the most common techniques
l Two problems:
l Obtain a rough registration to get a motion between different views
l Obtain a fine registration from an initial approximation
Introduction
l Applications:Reverse engineering,mould fabrication,artifact reproduction,etc.
l Range finders:Time-of-flight lasers,laser scanning,stereovision,pattern projection
Classification of methods
l Goal:Find the Euclidean motion between a set of range images.
l In terms of whether initial info. is required,we divide the process into two methods:coarse registration and fine registration.
Coarse registration
Methods:
l point signature
l spin image
l PCA
l RANSAC-based Darces
l Line-based
l Genetic Alg
l Principal Curvature
Criteria:
l the kind of correspo-ndences used
l the method used to compute the motion
l robustness
l registration strategy
Fine registration
Criteria:
l the registration strategy
l the use of an efficient search method
l the way of computing the motion in each iteration
l robustness
Methods:
l ICP
l Chen
l Signed Distance Fields
l Genetic Alg
l Registration strategy:multi-view;pair-wise
l The use of k-d tree to speed up or not
l The distance to minimize:point-to-point (problem:a given point in the first view may not exist in the second view);point-to-plane (theoretically less iterations)
Coarse registration
l The most common correspondences are points, curves,and surfaces.
Point signature(P-Ps)
l Point signature is a point descriptor used to search for correspondences.
l Advantage:the matching process is very fast and efficient
l Main drawback:the intersection of a sphere to the surface is not very easy,espectially when the surface is represented as a cloud of points or a triangulated surface(interpolation);sensible to noise
Spin image(P-Ps)
l Some spin image are computed in the first view and search for the best correspondences in the second view,then remove the outliers.
l Advantage:good results;the spin image feature is very robust except in case of symmetries or repeated regions.
l Main problem:the spin image strongly depends on the resolution of the method (solution:face-based spin image,interpolation) .
Principal component analysis(P-P)
l Using the direction of the main axis of the volu-me given by the cloud of points to align the sequence of range images.
l Advantage:this method is very fast.
l Drawback:it can only be used with effectiveness when there tis a sufficient number of points.In practice ,a 50% overlapping is critical.It's also limited in coping with surfaces that contain symmetries.In most cases this method is far from the expected.
Ransac-based darces(P-Ps)
l This method is based on finding the best three point-correspondences between two range ima-ges to obtain an estimation of the Euclidean motion.Normal vector and the principal curvatu-re is also used.
l The result obtained is very good because of its robustness even on the presence of outliers. It can only be used when the number of points in each view is relatively small,and where time is not critical.
Algebraic surface model(S)
l This method estimate the motion between sufraces represented as a polynomial model.It's based on obtaining a function of the distance between the polynomial model and the points.
l No need of points or curve correspondences.It's faster than others.A normal vector at each point is required.Good results need less than 15% of non-overlapping region.
Line-based algorithm
l This alg. proposed to use (straight or curve)lines to find pairs of correspondences.
l It is applied to large and structured environments in which planar regions and straight lines can be easily found.
l It obtains good results and the main drawback is the difficulty the segment the straight segme-nts as well as the supervisor.
l Improvement:Matching bitangent curves
l The main idea is that all bitangent points are co-incident in the dual space.
l Advantage:The range image is previously trans-formed into the dual space before the search for possible matches starts,which decreases the computing time and increases the robustness.
Genetic algorithm
l Defining a vector that contains the n index of correspondences between both range images. Also a fitness function is required to measure the quality of each potential solution.
l In most genetic approaches,the results obtained are quite good but the computing time is expen-sive,specially in the presence of a large number of points.
Principal curvature
l This approach proposed that use of the differe-ntial of points characterized by the principal curvature (max & min)in the matching process.
l The result is considered to be a good estimation. The main problem is that it is not robust.As the correspondence is not validated,it might be a false matching.
Fine registration methods
l The term fine registration use an initial guess to iterate and converge to a more accurate solutio-n,in which a distance function is minimized.
ICP
l The goal is to obtain an accurate solution by minimizing the distance between point-corresp-ondences,known as closest point.
l ICP obtains good results even in the presence of Gaussian noise.But it can not cope with non-overlapping regions because the outliers are never removed.The convergence is not guaran-teed.
l Modifications:
l Nearest Neighbor Problem is applied to facilita-te the search of closest points.This method dec-reases the computing time drastically.
l Muti-resolution Scheme ICP:It's for fast registra-tion,whose main idea is to solve the first few iterations using down sampled points and to pr-ogressively increase the resolution.
l ICP using invariant features(ICPIF):The matchi-ng process uses a weighted feature distance. Different invariant features were proposed,such as curvature,moment,spherical harmonics,amo-ng which the spherical harmonics provided the best convergence rate.
l Color and curvature features:As color is used as a matching constraint,symmetric objects can be well registered.
l RICP method making use of the Least Median of Squares:Aiming at finding a registration with-out outliers.It's robust!
l Picky ICP:It is based on outlier thresholding.The main difference with respects to the previous methods is that at every iteration only the pairs of correspondences with the smallest distances are used in the motion compution.
Method of Chen
l This is an alternative to the ICP algorithm,which was based on minimizing the distance between points and planes.
l To improve the precision of the algorithm,the Normal Space Sampling is proposed.The main idea is to select more points in the region where the normal is different from the other parts of the surface.
l This method is more robust to local minimum, and in general,results are better.Despite no robust method are applied ,this method is less sensible in the presence of non-overlapping regions.
l Chen's approach usually requires less iterations than the ICP approach.
Matching signed distance fields
l The method is a robust one so the outliers are removed,and all the views of a given object are registered at the same time(muti-view).
l Advantage:It registers all the views at the same time,which implies a more robust solution and avoids the error propagation problem present in pair-wise registration methods.And the final result is directly an integrated model.
l Drawback:It cannot be used in real time.
Genetic algorithm
l The goal of this method is to find a chromoso-me composed of the 6 parameters of the motion that aligns a pair of range images accurately.
l A similar method related to the original one has a more robust fitness function and the initial guess is not required.
l Overall,genetic alg. could avoid local minima especially when the initial motion is not provid-ed or it is given with low precision.It also works well in the presence of noise and ouliers given by non-overlapping regions.
l The main problem is the time required to converge.
内容来自用户分享和网络整理，不保证内容的准确性，如有侵权内容，可联系管理员处理 
相关文章推荐
- 当前深度图像配准方法介绍
- ITK图像配准介绍
- 图像配准的方法
- opencv 图像的深度和通道介绍
- 图像分割值OTSU方法介绍
- 当前最先进人脸识别方法简单介绍
- 图像处理中的阴影消除方法－一篇2007年PR上的文章介绍
- 简单介绍图像分割的方法
- Request.RawUrl获取当前请求重写（UrlRewrite）Url的方法介绍
- 深度学习（Deep Learning）的基本思想和方法，入门介绍
- [caffe]深度学习之CNN检测object detection方法摘要介绍
- PHP获取当前url的具体方法介绍
- CImage类的介绍与使用 【图像打开,另存为,转为灰度图像(利用cimage方法实现)】
- PHP获取当前url的具体方法介绍
- Javascript图像处理―虚拟边缘介绍及使用方法
- 图像配准的方法
- PHP获取当前url的具体方法介绍
- 基于特征的图像配准方法
- 图像配准的方法
- opencv中图像深度与通道介绍