CN109767463B - Automatic registration method for three-dimensional point cloud - Google Patents
Automatic registration method for three-dimensional point cloud Download PDFInfo
- Publication number
- CN109767463B CN109767463B CN201910019924.9A CN201910019924A CN109767463B CN 109767463 B CN109767463 B CN 109767463B CN 201910019924 A CN201910019924 A CN 201910019924A CN 109767463 B CN109767463 B CN 109767463B
- Authority
- CN
- China
- Prior art keywords
- point cloud
- point
- points
- feature
- registration
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Image Processing (AREA)
Abstract
The invention provides an automatic registration method of three-dimensional point cloud, aiming at solving the problems of the traditional ICP algorithm, such as calculation efficiency, precision and easy noise interference, the invention firstly establishes KD-tree for source point cloud and target point cloud to accelerate the search of near points; and then, coarse registration is realized by adopting a registration method based on a normal vector and a feature histogram, and a feature extraction part in the coarse registration is improved, so that feature points can be effectively extracted, and a large amount of point cloud information with unobvious features can not be lost. In order to further improve the registration precision, the precise registration provides an improved multi-resolution iteration nearest point algorithm, the algorithm provides that the point cloud resolution is calculated by using the density of the characteristic points, and meanwhile, the key point sampling method is improved. The method improves the precision, speed and noise resistance of the point clouds with different scales and different degrees of noise.
Description
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to an automatic three-dimensional point cloud registration method.
Background
With the progress of scientific technology and the rapid development of three-dimensional information acquisition technology, three-dimensional measurement and three-dimensional reconstruction technology is widely researched and applied in the fields of reverse engineering, cultural relic protection, virtual reality, augmented reality, robot vision, automatic navigation, industrial product appearance design and production, 3D printing and the like. The three-dimensional reconstruction method mainly comprises three major categories of image-based three-dimensional reconstruction, geometric-based three-dimensional reconstruction and scanner-based three-dimensional reconstruction. The invention relates to a three-dimensional reconstruction based on a scanner, which is a technology for rapidly and accurately reconstructing three-dimensional data by rapidly scanning a real object in the real world by using a three-dimensional laser scanning device or other three-dimensional measurement technologies to obtain three-dimensional point cloud data on the surface of the object and representing, analyzing and processing the point cloud data.
In view of the problems of incomplete acquired point clouds, rotation dislocation, translation dislocation and the like caused by the limited visual angle of each scanning of the three-dimensional scanning equipment, the local point clouds need to be registered for the acquired complete point clouds, so that a complete data model of the measured object is acquired. Therefore, point cloud registration is particularly important in the process of reconstructing the three-dimensional object.
The point cloud automatic registration method is generally divided into two steps of coarse registration and fine registration. The rough registration method comprises a mark point method, a gravity center coincidence method, a feature extraction method and the like. The mark point method needs to manually paste mark points, but the manual factors can cause errors and other problems, and the subsequent registration precision is influenced; the gravity center coincidence method is used for superposing two centers of point clouds together and is suitable for the condition of high point cloud superposition degree; feature extraction methods are generally not suitable for point cloud registration where object features are not apparent. Due to the defects of the coarse registration algorithm and the existence of the topological structure, the sparsity, the noise and the like of the point cloud, the accuracy of the coarse registration is difficult to meet the requirements of people, and therefore, the coarse registration result needs to be subjected to fine registration.
At present, the most common fine registration algorithm at home and abroad is an Iterative Closest Point (ICP), but the algorithm has the following defects: when the two point clouds have no better initial positions, the two point clouds are easy to fall into local optimum; in addition, the algorithm processes all points in the point cloud, so that the time for searching the matching points is long, and the efficiency is lower when the point cloud scale is larger. Aiming at the defects of the traditional ICP, a series of optimization algorithms are provided by a plurality of scholars at home and abroad. If a dynamic adjustment factor is introduced for ICP registration, the iteration times of the algorithm are reduced on the premise of not influencing the precision and the convergence direction of the registration algorithm, so that the registration efficiency is improved, but the method still easily leads to local optimization under the condition that the initial pose of the point cloud is not good. For example, aiming at the point cloud registration problem of different scales, an expanded GICP (generalized Gicp) method (namely SGICP) is provided for carrying out point cloud scale estimation, and point clouds of different scales indoors and outdoors can be effectively registered, but under the condition that the point cloud scale difference is large, the SGICP method cannot effectively and accurately search corresponding points, so that the registration cannot be effectively carried out, and in addition, the problem of point cloud registration under the condition of noise interference is not considered. Aiming at the condition of low point cloud coverage rate, the local feature descriptor is established by utilizing local features such as local depth, normal deflection angle, point cloud density and the like of the point cloud, so that the problem of point cloud registration with low point cloud coverage rate is solved, but the problem of point cloud registration in the presence of noise is not considered in the algorithm. Aiming at the problem of noisy point cloud registration, a Sparse Iteration Closest Point (SICP) algorithm is provided, although the algorithm reduces the influence of outliers on point cloud registration, the algorithm cannot process the condition that target point cloud contains outliers, and meanwhile, the algorithm efficiency is low.
In summary, the existing ICP-based improved algorithm has the following disadvantages:
(1) most of the existing feature-based registration algorithms have the problems of high speed but low precision or high precision but low speed. When the key feature points are selected more, the accuracy is relatively high, but the speed is slow due to the increase in the amount of calculation. When less key feature points are selected, the speed is faster but the accuracy is relatively lower.
(2) When the initial pose of the algorithm is not good, the registration is caused to fall into the problem of local optimization.
(3) Under the condition that noise or external point interference exists, the point cloud registration accuracy is not high or the efficiency is low.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides an automatic three-dimensional point cloud registration method, which is used for solving the problems of high calculation efficiency and precision of an ICP algorithm and high possibility of noise interference in the prior art. Therefore, the method determines the resolution ratio of the point cloud by calculating the mean value of the distances between each characteristic point and the adjacent points in the point cloud in the initial registration stage. Secondly, the matching degree is adopted to extract key point pairs, so that the error matching rate is reduced, and the registration precision and robustness are improved.
In order to solve the technical problems, the invention adopts the following technical scheme:
a three-dimensional point cloud automatic registration method is characterized by comprising the following steps:
s1, acquiring a source point cloud P and a target point cloud Q, and establishing a KD-tree based on the source point cloud P and the target point cloud Q to accelerate the search of k neighbor points;
s2, calculating the characteristic degree and the base of each point in the source point cloud P and the target point cloud QExtracting characteristic points in the source point cloud P and the target point cloud Q according to the characteristic degree to respectively obtain a characteristic point set PfAnd Qf；
S3, respectively calculating PfAnd QfThe histogram feature descriptor of each feature point in the image is used for obtaining an initial matching point pair based on the feature descriptor;
s4, taking the initial matching point pairs meeting the distance constraint as candidate matching point pairs, and removing mismatching point pairs in the candidate matching point pairs by using a RANSAC algorithm to obtain final matching point pairs;
s5, based on the final matching point pair, obtaining a rotation matrix R and a translational vector T by adopting a four-element method;
s6, converting the source point cloud P into a coordinate system of the target point cloud Q based on the rotation matrix R and the translation vector T to obtain the source point cloud P1；
S7 based on source point cloud P1The feature of the point in (1) is to point the source point cloud P1The points in (1) are divided into m levels and are based on a characteristic point set PfDense density calculation source point cloud P of middle feature points1The current resolution value is set to be 1;
s8, calculating the source point cloud P under the current resolution1Sampling proportion of each stage and extracting key points;
s9, calculating a source point cloud P1The matching degree of each key point and k adjacent points in the target point cloud Q, the adjacent point with the minimum matching value is taken as the matching point of the key point, and the key point and the matching point are marked as a key matching point pair;
s10, obtaining a rotation matrix R by adopting a four-element method based on key matching point pairs1And translation vector T1；
S11 based on rotation matrix R1And translation vector T1The source point cloud P1Converting the point cloud to a coordinate system of a target point cloud Q to obtain a source point cloud P2；
S12, repeating S8-S11 to make the objective function
Minimum of, in the formula, NP1Representing a source point cloud P1Number of points, qi4Is any point in the target point cloud Q, pi4As source point cloud P with qi4A corresponding point;
s13, if the current resolution is larger than or equal to the resolution upper limit value, or the root mean square error of the objective function meets a preset threshold, ending the iteration; otherwise, 1 is added to the current resolution, and execution returns to S8.
Preferably, in S2, based on the formula
Calculating the feature degree of each point in the source point cloud P and the target point cloud Q;
in the formula, pni1Is any point P in the source point cloud P or the target point cloud Qi1Normal vector of (n), npj1Is pi1Any one k of the neighboring points pj1Normal vector of (1), pni1Is pi1The average value of the included angles between the normal vectors and k adjacent points is the characteristic degree;
if pni1＞ε1A 1 is to pi1As candidate feature point,. epsilon1Extracting a threshold value for the candidate feature point;
and if the feature degree of the candidate feature point is equal to the feature degree of the nearest point with the maximum feature degree in the k nearest points, taking the candidate feature point as the feature point.
Preferably, the first and second electrodes are formed of a metal,
NPthe number of points representing the source point cloud P.
Preferably, in S7, based on the formula
Computing a source point cloud P1Where S represents an adjustment factor, psi2Representing a set of source cloud feature points PfAny one of the characteristic points of the image data,
is shown as a featurePoints psi2K is a neighbor, NP1Representing a source point cloud P1Number of points, NpsRepresenting a set of source cloud feature points PfThe number of feature points in, round is a function of rounding.
Preferably, in S8, based on the formula
And keypointi3,j3＝round(Ri3,j3·countj3) Extracting key points of formula (I) wherein Ri3,j3At j3 sample rate at current resolution i3, keypointi3,j3Is a key point of j3 level under the current resolution i3, countj3Is the total number of points at level j 3.
In summary, the invention provides an automatic three-dimensional point cloud registration method, which aims to solve the problems of the traditional ICP algorithm such as calculation efficiency, accuracy and susceptibility to noise interference, and firstly establishes a KD-tree for a source point cloud and a target point cloud to accelerate the search of a near point; and then, coarse registration is realized by adopting a registration method based on a normal vector and a feature histogram, and a feature extraction part in the coarse registration is improved, so that feature points can be effectively extracted, and a large amount of point cloud information with unobvious features can not be lost. In order to further improve the registration precision, the precise registration provides an improved multi-resolution iteration nearest point algorithm, the algorithm provides that the point cloud resolution is calculated by using the density of the characteristic points, and meanwhile, the key point sampling method is improved. The method improves the precision, speed and noise resistance of the point clouds with different scales and different degrees of noise.
Drawings
FIG. 1 is a flow chart of a three-dimensional point cloud automatic registration method disclosed by the invention;
FIG. 2 is an original view of a Bunny point cloud model before registration;
FIG. 3 is a diagram of the registration result of a classical ICP algorithm to a Bunny point cloud;
FIG. 4 is a diagram of the results of the registration of a Bunny point cloud by the existing improved ICP algorithm;
FIG. 5 is a diagram of the registration result of the prior coarse registration algorithm on the Bunny point cloud;
FIG. 6 is a result diagram of the rough registration of the Bunny point cloud by the automatic three-dimensional point cloud registration method disclosed by the invention;
FIG. 7 is a result diagram of the precise registration of the Bunny point cloud by the automatic three-dimensional point cloud registration method disclosed by the invention;
FIG. 8 is an original view of a noisy Bunny point cloud model;
FIG. 9 is a diagram of the result of a classical ICP algorithm with noisy Bunny point cloud registration;
FIG. 10 is a diagram of the results of a noisy Bunny point cloud registration of the prior art improved ICP algorithm;
FIG. 11 is a diagram of the result of noisy Bunny point cloud registration by the prior coarse registration algorithm;
FIG. 12 is a result diagram of the rough registration of noisy Bunny point clouds by the automatic registration method for three-dimensional point clouds disclosed in the invention;
fig. 13 is a result diagram of the precise registration of the noisy Bunny point cloud by the three-dimensional point cloud automatic registration method disclosed by the invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
As shown in fig. 1, the invention discloses an automatic registration method of three-dimensional point cloud, comprising the following steps:
s1, acquiring a source point cloud P and a target point cloud Q, and establishing a KD-tree based on the source point cloud P and the target point cloud Q to accelerate the search of k neighbor points;
in the invention, a three-dimensional scanner can be used for scanning the surface of an object from different visual angles to obtain at least two visual angle three-dimensional point clouds, the two visual angle three-dimensional point clouds are taken as a source point cloud P and a target point cloud Q respectively, and a KD-tree is established for the source point cloud P and the target point cloud Q
S2, calculating the feature degree of each point in the source point cloud P and the target point cloud Q, extracting the feature points in the source point cloud P and the target point cloud Q based on the feature degree, and respectively obtaining a feature point set PfAnd Qf；
S3, respectively calculating PfAnd QfA histogram feature descriptor for each feature point in the image, based on which the histogram feature descriptor is obtainedInitial matching point pairs;
s4, taking the initial matching point pairs meeting the distance constraint as candidate matching point pairs, and removing mismatching point pairs in the candidate matching point pairs by using a RANSAC algorithm to obtain final matching point pairs;
s5, based on the final matching point pair, obtaining a rotation matrix R and a translational vector T by adopting a four-element method;
s6, converting the source point cloud P into a coordinate system of the target point cloud Q based on the rotation matrix R and the translation vector T to obtain the source point cloud P1；
S7 based on source point cloud P1The feature of the point in (1) is to point the source point cloud P1The points in (1) are divided into m levels and are based on a characteristic point set PfDense density calculation source point cloud P of middle feature points1The current resolution value is set to be 1;
s8, calculating the source point cloud P under the current resolution1Sampling proportion of each stage and extracting key points;
s9, calculating a source point cloud P1The matching degree of each key point and k adjacent points in the target point cloud Q, the adjacent point with the minimum matching value is taken as the matching point of the key point, and the key point and the matching point are marked as a key matching point pair;
source point cloud P1At an arbitrary key point pi5Using KD-tree to find point pi5K neighbor points in the target point cloud Q;
respectively calculate pi5Degree of matching W (p) with its k neighborsi5,qj5)：
qj5Is pi5Any one of k neighbors of (1), wherein p isi5，1、pi5，2、pi5，3、pi5，4Are respectively a point pi5Principal curvature k of1、k2Gaussian curvature K and mean curvature H, qj5，1、qj5，2、qj5，3、qj5，4Are respectively provided withIs qj5Principal curvature k of1、k2Gaussian curvature K, mean curvature H.
When the matching degree is calculated, if a certain point has no curvature information, a KD-tree is used for solving the k adjacent point of the point cloud corresponding to the point cloud, and curvature information is calculated and stored; selecting W (p)i5,qj5) The point with the smallest value is taken as pi5The matching points of (1); from this, the source point cloud P is selected1The matching points of each key point.
S10, obtaining a rotation matrix R by adopting a four-element method based on key matching point pairs1And translation vector T1；
S11 based on rotation matrix R1And translation vector T1The source point cloud P1Converting the point cloud to a coordinate system of a target point cloud Q to obtain a source point cloud P2；
S12, repeating S8-S11 to make the objective function
Minimum of, in the formula, NP1Representing a source point cloud P1Number of points, qi4Is any point in the target point cloud Q, pi4As source point cloud P with qi4A corresponding point;
s13, if the current resolution is larger than or equal to the resolution upper limit value, or the root mean square error of the objective function meets a preset threshold, ending the iteration; otherwise, 1 is added to the current resolution, and execution returns to S8.
In the invention, S1-S6 are coarse registration, S7-S13 are fine registration, and compared with the existing ICP algorithm, the method has the following beneficial technical effects:
the method aims at the requirement of an ICP algorithm on the initial pose of point cloud, carries out rapid registration through a coarse registration algorithm, and then carries out fine registration through an improved multi-resolution iteration closest point registration algorithm, thereby further improving the accuracy of the algorithm.
In order to improve the registration precision and enhance the noise resistance of the algorithm as much as possible, the method adopts Euclidean distance constraint and RANSAC algorithm to remove the outer points in the coarse registration part, improves the precision of the coarse registration and ensures that the method can provide better initial pose for the fine registration of the point cloud containing noise and partial overlapping degree. In the accurate registration part, the problem of setting resolution is firstly improved, the resolution can be self-adaptively valued through the density of the point cloud characteristic points, and the problems of low registration speed caused by overlarge resolution setting and low precision caused by undersize resolution setting are avoided. And secondly, the sampling proportion of the algorithm is improved, the registration precision of the algorithm on the point cloud with unobvious characteristic information is improved, and the applicability of the algorithm is enhanced. The method has the advantages that the feasibility and the effectiveness of the algorithm are proved through real verification, and the method is high in precision and high in speed no matter which scale of point cloud is subjected to registration.
In specific implementation, in S2, the method is based on the formula
Calculating the feature degree of each point in the source point cloud P and the target point cloud Q;
in the formula, pni1Is any point P in the source point cloud P or the target point cloud Qi1Normal vector of (n), npj1Is pi1Any one k of the neighboring points pj1Normal vector of (1), pni1Is pi1The average value of the included angles between the normal vectors and k adjacent points is the characteristic degree;
if pni1＞ε1A 1 is to pi1As candidate feature point,. epsilon1Extracting a threshold value for the candidate feature point;
and if the feature degree of the candidate feature point is equal to the feature degree of the nearest point with the maximum feature degree in the k nearest points, taking the candidate feature point as the feature point.
In the specific implementation process, the first-stage reactor,
NPthe number of points representing the source point cloud P.
In the feature point extraction process, the selection of the threshold value is quite important. If the threshold value is too large, the extracted feature points are relatively few, and key feature information and a large amount of detail of part of point cloud may be lostSection information; if the threshold is chosen too small, a lot of detail information can be retained, but the point cloud registration speed is slow. Threshold value epsilon1The method is inconvenient in practical application because a proper threshold value needs to be selected through multiple experiments aiming at different point cloud data according to an empirical value selected through experiments. The method can adaptively extract the characteristic points according to the change of the normal vector and an adaptive calculation threshold, can well retain key characteristic information of the point cloud, simultaneously removes a great amount of detailed information with small influence, and accelerates the registration speed.
In specific implementation, in S7, the method is based on the formula
Computing a source point cloud P1Where S represents an adjustment factor, psi2Representing a set of source cloud feature points PfAny one of the characteristic points of the image data,
expressed as characteristic points psi2K is a neighbor, NP1Representing a source point cloud P1Number of points, NpsRepresenting a set of source cloud feature points PfThe number of feature points in, round is a function of rounding.
The essence of multi-resolution registration is that the extracted keypoints can be adjusted according to the number of stages and resolution. The number of points extracted at each stage is controlled by the stage number, and the sampling proportion is larger when the stage number is larger; the resolution controls the number of points extracted integrally, the larger the resolution is, the more the points sampled integrally are, the higher the precision is, and the speed is relatively slow. In order to make the point cloud converge quickly, the low-resolution matching point pairs are used for carrying out quick registration, and then the high-resolution matching point pairs are used for improving the registration accuracy, so that the problems of low point cloud registration accuracy and low speed are solved.
The resolution can be reflected by the density degree of the point cloud, wherein the density degree is the mean value of the Euclidean distances between each point in the point cloud and the k adjacent points. However, the calculation amount is large when the density degree of all the points in one point cloud is calculated, so the invention provides that the mean value of the Euclidean distances between the characteristic points and k adjacent points of the characteristic points is used for calculation, and the size of the resolution ratio is dynamically adjusted according to the adjustment factor. The method calculates the resolution ratio of the point cloud according to the density of the point cloud characteristic points, and is more reasonable than the setting by experience. In order to reduce the time complexity without repeatedly extracting the feature points, the resolution of the point cloud is calculated and output when the feature points are extracted by coarse registration.
The larger the average value of the included angle of the normal vector is, the more obvious the characteristic is, and on the contrary, the smaller the average value of the included angle of the normal vector is, the less obvious the characteristic is. However, most of the point cloud data generally obtained has relatively few points with obvious features and relatively many points with no obvious features, so that the registration accuracy is low if only the key points with obvious features are relied on during registration, and the registration may fail. The improved sampling mode provided by the invention not only makes full use of key points with obvious characteristics, but also makes full use of points with unobvious characteristics.
In specific implementation, in S8, the method is based on the formula
And keypointi3,j3＝round(Ri3,j3·countj3) Extracting key points of formula (I) wherein Ri3,j3At j3 sample rate at current resolution i3, keypointi3,j3Is a key point of j3 level under the current resolution i3, countj3Is the total number of points at level j 3.
The sampling mode improves the sampling proportion when the level is lower, makes full use of points with unobvious characteristics, and greatly improves the registration precision.
The following are experiments conducted using the Bunny and Buddha point clouds of Standford 3D Scaning reproducibility. Fig. 2 is an original diagram of the Bunny point cloud model before registration, and fig. 3, 4, 5, 6 and 7 are diagrams of the registration results of the conventional ICP algorithm, the existing improved ICP algorithm, the existing coarse registration algorithm, the coarse registration algorithm of the invention and the fine registration algorithm of the invention on the Bunny point cloud respectively. Fig. 8 is an original diagram of a noisy Bunny point cloud model, and fig. 9, 10, 11, 12 and 13 are diagrams of the registration results of a traditional ICP algorithm, a prior improved ICP algorithm, a prior coarse registration algorithm, a coarse registration algorithm of the invention and a fine registration algorithm of the invention. Table one is registration comparison under different algorithms:
TABLE 1 different Algorithm registration Compare
Table 1 shows the comparison of the registration speed and registration error of the original Bunny point cloud and the noisy Bunny point cloud for various registration algorithms. The method has the advantages that the classic ICP algorithm is most time-consuming, is easily interfered by noise, and is low in precision in noise point cloud registration; the registration speed of the existing improved ICP algorithm is greatly improved, the accuracy of the existing improved ICP algorithm is lower than that of the classical ICP algorithm when the original Bunny point cloud is registered, and the existing improved ICP algorithm is easily interfered by noise; the precision of the existing coarse registration algorithm is higher than that of a classical ICP algorithm and that of an existing improved ICP algorithm, the degree of noise interference is smaller than that of the two algorithms, and the speed of the algorithm is slower; the precision of the rough registration algorithm is slightly lower than that of the existing rough registration algorithm when the original Bunny point cloud registration and the noisy point cloud registration are carried out, but the speed of the rough registration algorithm is nearly 3 times faster than that of the existing rough registration algorithm; the precision of the fine registration algorithm is further improved on the basis of coarse registration, and particularly the noise-carrying point cloud registration is better in stability and higher in precision.
The method comprises the steps of selecting the Buddha point cloud data with different visual angles and different scales for experiments, wherein the table 2 is used for carrying out registration comparison on the Buddha point clouds with different scales through various algorithms, and the table 3 is used for carrying out registration comparison on the Buddha point cloud data (0 degree and 24 degrees) by adding noise with different scales.
TABLE 2 different Scale Buddha Point cloud registration
As can be seen from table 2, no matter which registration algorithm is adopted, the larger the point cloud size is, the more time it takes for registration; regardless of the size of the point cloud data, the overall performance of the algorithm is optimal. The classical ICP algorithm has a very slow speed although the registration accuracy is good; the existing improved ICP algorithm and the existing coarse registration algorithm have almost the same accuracy, but the existing improved ICP algorithm has higher speed, the registration accuracy is further improved by the method through secondary registration, the registration accuracy is the highest in several comparison algorithms, and the overall speed is higher than that of the existing coarse registration algorithm.
TABLE 3 Butdha Point cloud registration for different scale noise
As can be seen from table 3, the accuracy and speed of the various algorithms are affected to different degrees after the noise points are added. The classical ICP algorithm and the existing improved ICP algorithm are affected by noise to the greatest extent, and become less accurate as noise is added. The classical ICP algorithm has higher operation efficiency and lower accuracy after adding noise points, but does not have lower accuracy as the added noise is more, which is caused by the position distribution of the noise points. After the noise points are added, the speed of the existing improved ICP algorithm is not changed greatly, the precision is lowered, but the precision does not increase linearly along with the scale of the noise points. The robustness of the existing coarse registration algorithm and the text coarse registration and fine registration algorithm is good, the precision of the algorithm is basically not influenced by noise, the RANSAC algorithm is added in the existing coarse registration algorithm and the text coarse registration algorithm to further eliminate mismatching points, the robustness of the algorithm is enhanced, and the speed of the text coarse registration algorithm is nearly 3 times faster than that of the existing coarse registration algorithm. The fine registration algorithm further performs registration on the basis of coarse registration, and improves sampling proportion, so that the registration accuracy is higher and the registration stability is better in various algorithms.
The invention provides the automatic three-dimensional point cloud registration method, which aims at the requirement of the ICP algorithm on the initial pose of the point cloud, combines the existing coarse registration algorithm to carry out rapid registration, then carries out accurate registration on the point cloud with better pose, and further improves the accuracy of the algorithm.
In order to improve the registration precision and enhance the noise resistance of the algorithm as much as possible, the method adopts Euclidean distance constraint and RANSAC algorithm to remove the outer points in the coarse registration part, improves the precision of the coarse registration and ensures that the method can provide better initial pose for the fine registration of the point cloud containing noise and partial overlapping degree. In the accurate registration part, the problem of setting resolution is firstly improved, the resolution can be self-adaptively valued through the density of the point cloud characteristic points, and the problems of low registration speed caused by overlarge resolution setting and low precision caused by undersize resolution setting are avoided. And secondly, the sampling proportion of the algorithm is improved, the registration precision of the algorithm on the point cloud with unobvious characteristic information is improved, and the applicability of the algorithm is enhanced. The method has the advantages that the feasibility and the effectiveness of the algorithm are proved through real verification, and the method is high in precision and high in speed no matter which scale of point cloud is subjected to registration.
It should be noted that the above mentioned embodiments are only preferred embodiments of the present invention, and that several variations and modifications may be made by those skilled in the art without departing from the present invention, and the above mentioned variations and modifications should also be considered as falling within the scope of the present invention as claimed.
Claims (2)
1. A three-dimensional point cloud automatic registration method is characterized by comprising the following steps:
s1, acquiring a source point cloud P and a target point cloud Q, and establishing a KD-tree based on the source point cloud P and the target point cloud Q to accelerate the search of k neighbor points;
s2, calculating the feature degree of each point in the source point cloud P and the target point cloud Q, extracting the feature points in the source point cloud P and the target point cloud Q based on the feature degree, and respectively obtaining a feature point set PfAnd Qf；
Based on the formula
Calculating the feature degree of each point in the source point cloud P and the target point cloud Q;
in the formula, npi1In a source point cloud P or a target point cloud QAny one point pi1Normal vector of (n), npj1Is pi1Any one k of the neighboring points pj1Normal vector of (1), pni1Is pi1The average value of the included angles between the normal vectors and k adjacent points is the characteristic degree;
if pni1＞ε1A 1 is to pi1As candidate feature point,. epsilon1A threshold value is extracted for the candidate feature points,
NPpoints representing the source point cloud P;
if the feature degree of the candidate feature point is equal to the feature degree of the nearest point with the maximum feature degree in the k nearest points, taking the candidate feature point as the feature point;
s3, respectively calculating PfAnd QfThe histogram feature descriptor of each feature point in the image is used for obtaining an initial matching point pair based on the feature descriptor;
s4, taking the initial matching point pairs meeting the distance constraint as candidate matching point pairs, and removing mismatching point pairs in the candidate matching point pairs by using a RANSAC algorithm to obtain final matching point pairs;
s5, based on the final matching point pair, obtaining a rotation matrix R and a translational vector T by adopting a four-element method;
s6, converting the source point cloud P into a coordinate system of the target point cloud Q based on the rotation matrix R and the translation vector T to obtain the source point cloud P1；
S7 based on source point cloud P1The feature of the point in (1) is to point the source point cloud P1The points in (1) are divided into m levels and are based on a characteristic point set PfDense density calculation source point cloud P of middle feature points1The current resolution value is set to be 1,
computing a source point cloud P1Where S represents an adjustment factor, psi2Representing a set of source cloud feature points PfAny one of the characteristic points of the image data,
expressed as characteristic points psi2K is a neighbor, NP1Representing a source point cloud P1Number of points, NpsRepresenting a set of source cloud feature points PfThe number of medium feature points, round, is a rounded function;
s8, calculating the source point cloud P under the current resolution1Sampling proportion of each stage and extracting key points;
s9, calculating a source point cloud P1The matching degree of each key point and k adjacent points in the target point cloud Q, the adjacent point with the minimum matching value is taken as the matching point of the key point, and the key point and the matching point are marked as a key matching point pair;
s10, obtaining a rotation matrix R by adopting a four-element method based on key matching point pairs1And translation vector T1；
S11 based on rotation matrix R1And translation vector T1The source point cloud P1Converting the point cloud to a coordinate system of a target point cloud Q to obtain a source point cloud P2；
S12, repeating S8-S11 to make the objective function
Minimum of, in the formula, NP1Representing a source point cloud P1Number of points, qi4Is any point in the target point cloud Q, pi4As source point cloud P with qi4A corresponding point;
s13, if the current resolution is larger than or equal to the resolution upper limit value, or the root mean square error of the objective function meets a preset threshold, ending the iteration; otherwise, 1 is added to the current resolution, and execution returns to S8.
2. The method of automatic registration of three-dimensional point cloud of claim 1, wherein in S8, based on formula
And keypointi3,j3＝round(Ri3,j3·countj3) Extracting key points of formula (I) wherein Ri3,j3At j3 sample rate at current resolution i3, keypointi3,j3Is a key point of j3 level under the current resolution i3, countj3Is the total number of points at level j 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910019924.9A CN109767463B (en) | 2019-01-09 | 2019-01-09 | Automatic registration method for three-dimensional point cloud |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910019924.9A CN109767463B (en) | 2019-01-09 | 2019-01-09 | Automatic registration method for three-dimensional point cloud |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109767463A CN109767463A (en) | 2019-05-17 |
| CN109767463B true CN109767463B (en) | 2021-04-13 |
Family
ID=66453616
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910019924.9A Active CN109767463B (en) | 2019-01-09 | 2019-01-09 | Automatic registration method for three-dimensional point cloud |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109767463B (en) |
Families Citing this family (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110766733B (en) * | 2019-10-28 | 2022-08-12 | 广东三维家信息科技有限公司 | Single-space point cloud registration method and device |
| CN112950706B (en) * | 2019-12-11 | 2023-11-03 | 京东科技信息技术有限公司 | Mobile equipment positioning data processing method, device, equipment and storage medium |
| CN111311576B (en) * | 2020-02-14 | 2023-06-02 | 易思维（杭州）科技有限公司 | Defect detection method based on point cloud information |
| CN111340862B (en) * | 2020-02-18 | 2023-07-07 | 广州智能装备研究院有限公司 | Point cloud registration method and device based on multi-feature fusion and storage medium |
| CN111369602B (en) * | 2020-02-25 | 2023-10-27 | 阿波罗智能技术(北京)有限公司 | Point cloud data processing method and device, electronic equipment and readable storage medium |
| CN112017219B (en) * | 2020-03-17 | 2022-04-19 | 湖北亿咖通科技有限公司 | Laser point cloud registration method |
| CN111445540B (en) * | 2020-03-26 | 2023-04-18 | 重庆理工大学 | Automatic registration method for RGB colored three-dimensional point cloud |
| CN112285735B (en) * | 2020-09-18 | 2023-04-18 | 北京捷象灵越科技有限公司 | System for automatically calibrating angular resolution of single-line laser radar |
| CN112652003B (en) * | 2020-10-29 | 2024-04-12 | 西北工业大学 | Three-dimensional point cloud registration method based on RANSAC measure optimization |
| CN112348864B (en) * | 2020-11-11 | 2022-10-11 | 湖南大学 | Three-dimensional point cloud automatic registration method for laser contour features of fusion line |
| CN112819869A (en) * | 2021-01-22 | 2021-05-18 | 辽宁工程技术大学 | Three-dimensional point cloud registration method based on IHarris-TICP algorithm |
| CN113706587B (en) * | 2021-07-14 | 2022-12-09 | 西安交通大学 | Rapid point cloud registration method, device and equipment based on space grid division |
| CN114549601B (en) * | 2022-02-11 | 2023-03-28 | 中国科学院精密测量科学与技术创新研究院 | Landslide multi-temporal TLS point cloud fine registration method considering point pair reliability |
| CN117289651B (en) * | 2023-11-24 | 2024-04-16 | 南通汤姆瑞斯工业智能科技有限公司 | Numerical control machining method and control system for die manufacturing |
| CN117788537B (en) * | 2024-02-27 | 2024-04-26 | 南京航空航天大学 | Pointnet-based aircraft skin point cloud registration method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106097324A (en) * | 2016-06-07 | 2016-11-09 | 中国农业大学 | A kind of non-rigid 3D shape corresponding point determine method |
| CN106709883A (en) * | 2016-12-20 | 2017-05-24 | 华南理工大学 | Point cloud denoising method based on joint bilateral filtering and sharp feature skeleton extraction |
| CN108470374A (en) * | 2018-04-08 | 2018-08-31 | 中煤航测遥感集团有限公司 | Mass cloud data processing method and processing device |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103236064B (en) * | 2013-05-06 | 2016-01-13 | 东南大学 | A kind of some cloud autoegistration method based on normal vector |
| CN104134216B (en) * | 2014-07-29 | 2016-09-14 | 武汉大学 | The laser point cloud autoegistration method described based on 16 dimensional features and system |
| CN104143210B (en) * | 2014-07-31 | 2017-04-12 | 哈尔滨工程大学 | Multi-scale normal feature point cloud registering method |
| CN106780459A (en) * | 2016-12-12 | 2017-05-31 | 华中科技大学 | A kind of three dimensional point cloud autoegistration method |
| CN107886529B (en) * | 2017-12-06 | 2020-04-10 | 重庆理工大学 | Point cloud registration method for three-dimensional reconstruction |
| CN108133458A (en) * | 2018-01-17 | 2018-06-08 | 视缘(上海)智能科技有限公司 | A kind of method for automatically split-jointing based on target object spatial point cloud feature |
-
2019
- 2019-01-09 CN CN201910019924.9A patent/CN109767463B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106097324A (en) * | 2016-06-07 | 2016-11-09 | 中国农业大学 | A kind of non-rigid 3D shape corresponding point determine method |
| CN106709883A (en) * | 2016-12-20 | 2017-05-24 | 华南理工大学 | Point cloud denoising method based on joint bilateral filtering and sharp feature skeleton extraction |
| CN108470374A (en) * | 2018-04-08 | 2018-08-31 | 中煤航测遥感集团有限公司 | Mass cloud data processing method and processing device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109767463A (en) | 2019-05-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109767463B (en) | Automatic registration method for three-dimensional point cloud | |
| CN108510532B (en) | Optical and SAR image registration method based on deep convolution GAN | |
| CN109887015B (en) | Point cloud automatic registration method based on local curved surface feature histogram | |
| CN107886529B (en) | Point cloud registration method for three-dimensional reconstruction | |
| CN105740899B (en) | A kind of detection of machine vision image characteristic point and match compound optimization method | |
| CN109903319B (en) | Multi-resolution-based fast iteration closest point registration algorithm | |
| CN108171780A (en) | A kind of method that indoor true three-dimension map is built based on laser radar | |
| CN109147040B (en) | Template-based human point cloud hole repairing method | |
| CN111179321B (en) | Point cloud registration method based on template matching | |
| CN105160686B (en) | A kind of low latitude various visual angles Remote Sensing Images Matching Method based on improvement SIFT operators | |
| CN105046694A (en) | Quick point cloud registration method based on curved surface fitting coefficient features | |
| WO2022237225A1 (en) | Online real-time registration method for incomplete three-dimensional scanning point cloud having plane reference | |
| CN105654483A (en) | Three-dimensional point cloud full-automatic registration method | |
| CN111028292A (en) | Sub-pixel level image matching navigation positioning method | |
| CN110910433A (en) | Point cloud matching method based on deep learning | |
| CN103700135B (en) | A kind of three-dimensional model local spherical mediation feature extracting method | |
| CN113450269A (en) | Point cloud key point extraction method based on 3D vision | |
| CN115471682A (en) | Image matching method based on SIFT fusion ResNet50 | |
| CN111161267A (en) | Segmentation method of three-dimensional point cloud model | |
| CN112257721A (en) | Image target region matching method based on Fast ICP | |
| CN113409332B (en) | Building plane segmentation method based on three-dimensional point cloud | |
| CN110942077A (en) | Feature line extraction method based on weight local change degree and L1 median optimization | |
| CN113706588B (en) | Annular forging point cloud registration method based on improved four-point quick robust matching algorithm | |
| CN109345571B (en) | Point cloud registration method based on extended Gaussian image | |
| CN117237428B (en) | Data registration method, device and medium for three-dimensional point cloud |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20230705 Address after: No. 1811, 18th Floor, Building 19, Section 1201, Lushan Avenue, Wan'an Street, Tianfu New District, Chengdu, Sichuan, China (Sichuan) Pilot Free Trade Zone, 610213, China Patentee after: Sichuan Jiulai Technology Co.,Ltd. Address before: No. 69 lijiatuo Chongqing District of Banan City Road 400054 red Patentee before: Chongqing University of Technology |
|
| TR01 | Transfer of patent right |