Point cloud registration method and system based on the angle between the closest point pair vector and the centroid vector

Abstract

The present invention discloses a point cloud registration method and system based on the angle between the closest point pair vector and the centroid vector, relating to the field of computer vision technology. The method comprises: obtaining a target point cloud and a source point cloud as raw data; bidirectionally filtering the target point cloud and the source point cloud based on the minimum distance between the midpoints of the target point cloud and the source point cloud to obtain a first filtered point pair set and a first filtered point cloud; obtaining a target point-source point vector set based on the first filtered point pair set; obtaining a centroid vector based on the raw data or the first filtered point cloud; obtaining an angle parameter based on the target point-source point vector set and the centroid vector; filtering the first filtered point pair set based on the angle parameter to obtain a second filtered point pair set and a second filtered point cloud; obtaining a covariance matrix based on the second filtered point cloud; obtaining a mean square error value based on the covariance matrix and the second filtered point pair set; and determining and outputting a final registered source point cloud based on the mean square error value. This improves the accuracy and stability of point cloud registration.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more particularly to a point cloud registration method and system based on the angle between a closest point pair vector and a centroid vector. Background Art

[0002] With the continuous advancement of computer vision, 3D imaging, and sensor technologies, point cloud technology is increasingly being used across various industries. Point cloud data, a collection of 3D spatial data points acquired by sensors (such as laser scanners and depth cameras), is widely used in fields such as architecture, robotics, autonomous driving, and virtual reality (VR). Since data acquired from multiple sensors or multiple viewpoints is unlikely to be completely consistent, point cloud registration techniques are needed to optimize the alignment of different point cloud data. In autonomous driving, data generated by multiple sensors (such as LiDAR and stereo cameras) must be accurately registered to construct an accurate environmental model. In the construction industry, point cloud data from multiple viewpoints can help create precise building models. In VR, 3D scenes captured from point cloud data require precise registration to achieve a seamless immersive experience. Common techniques in point cloud registration include feature-based registration, iterative optimization-based registration, probability distribution-based registration, and deep learning-based registration.

[0003] However, traditional methods (feature-based, iterative optimization, and probability distribution) are highly reliable in structured scenarios, but are limited by feature sensitivity, initial pose dependence, and computational complexity, and suffer from problems such as inaccurate point pair matching and large registration errors. Deep learning methods use data to achieve end-to-end optimization, which can overcome some traditional bottlenecks, but their performance is limited by the quality of training data and insufficient generalization capabilities. They require a large amount of training data and high-performance computing resources, have poor interpretability, and the training process is very time-consuming.

[0004] Therefore, how to improve the accuracy and stability of point cloud registration is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a point cloud registration method and system based on the angle between the closest point pair vector and the centroid vector, which improves the accuracy and stability of point cloud registration.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] Point cloud registration methods based on the angle between the closest point pair vector and the centroid vector include:

[0008] Obtain target point cloud and source point cloud as raw data;

[0009] Bidirectional screening based on the minimum distance between the midpoints of the target point cloud and the source point cloud obtains a first screening point pair set and a first screening point cloud;

[0010] Obtaining a target point-source point vector set based on the first screening point pair set;

[0011] Obtaining a centroid vector based on the original data or the first filtered point cloud;

[0012] Obtaining an angle parameter based on the target point-source point vector set and the centroid vector;

[0013] Filtering the first screening point pair set based on the angle parameter to obtain a second screening point pair set and a second screening point cloud;

[0014] Obtaining a covariance matrix based on the second filtered point cloud;

[0015] Obtaining a mean square error value based on the covariance matrix and the second screening point pair set;

[0016] The final registration source point cloud is determined and output based on the mean square error value.

[0017] Preferably, the method for obtaining the first screening point pair set and the first screening point cloud is:

[0018] Obtaining a target point with the smallest distance in the target point cloud based on the current source point in the source point cloud as the first selected point;

[0019] Based on the first selected point, a source point with the smallest distance in the source point cloud is obtained as a second selected point;

[0020] Determining whether the second selected point is consistent with the current source point;

[0021] If they are consistent, it indicates that the match is successful, and the current source point is used as the first screening source point, the first selected point is used as the first screening target point, and the first screening source point and the first screening target point form the first screening point pair;

[0022] Otherwise, it indicates that the matching is unsuccessful;

[0023] Repeat the above process to search and filter all source points in the source point cloud to obtain all the first filtered point pairs to form the first filtered point pair set;

[0024] All first screening source points obtained based on the first screening point pair set are sequentially composed of a first screening source point cloud;

[0025] All first screening target points obtained based on the first screening point pair set are sequentially composed of a first screening target point cloud;

[0026] The first screening point cloud is composed based on the first screening source point cloud and the first screening target point cloud.

[0027] Preferably, the method for obtaining the centroid vector is specifically as follows:

[0028] Obtaining the centroid of local k-nearest neighbor points based on all the first filtered source points in the first filtered source point cloud to obtain the centroid of the first source point cloud;

[0029] Obtaining the centroid of local k-nearest neighbor points based on all the first screened target points in the first screened target point cloud to obtain the centroid of the first target point cloud;

[0030] A first centroid vector is determined based on the difference between the first target point cloud centroid and the first source point cloud centroid;

[0031] Based on the first centroid vector as the centroid vector.

[0032] Preferably, the method for obtaining the centroid vector is specifically as follows:

[0033] Taking the mean of all source points in the source point cloud to obtain a second source point cloud centroid;

[0034] Taking an average of the target points in the target point cloud to obtain a second target point cloud centroid;

[0035] A second centroid vector is determined based on the difference between the second target point cloud centroid and the second source point cloud centroid;

[0036] Based on the second centroid vector as the centroid vector.

[0037] Preferably, the method for obtaining the angle parameter is:

[0038] Acquire a cosine value of a vector angle based on the target point-source point vector in the target point-source point vector set and the first centroid vector as a first angle parameter;

[0039] Acquire a cosine value of a vector angle based on the target point-source point vector in the target point-source point vector set and the second centroid vector as a second angle parameter;

[0040] The angle parameter includes the first angle parameter or the second angle parameter.

[0041] Preferably, the second screening point pair set and the second screening point cloud are obtained by:

[0042] Get the angle threshold;

[0043] Obtaining a cosine value of the angle threshold as a threshold parameter;

[0044] Determining whether the angle parameter is greater than or equal to the threshold parameter;

[0045] If yes, retain the first screening point pair corresponding to the angle parameter as the second screening point pair;

[0046] Otherwise, the first screening point corresponding to the angle parameter is not retained;

[0047] Circularly determining the angle parameters corresponding to all the first screening point pairs to obtain all the second screening point pairs as the second screening point pair set;

[0048] forming a second screening target point cloud based on all second screening target points in the second screening point pair set;

[0049] forming a second screening source point cloud based on all second screening source points in the second screening point pair set;

[0050] The second screening point cloud is composed based on the second screening target point cloud and the second screening source point cloud.

[0051] Preferably, the covariance matrix is ​​obtained by:

[0052] Based on the second filtered source point cloud and the second filtered target point cloud, the second filtered source point cloud and the second filtered target point cloud are transferred to the de-centroided coordinate system, and the centroid of the second filtered source point cloud and the centroid of the second filtered target point cloud are correspondingly obtained;

[0053] Subtracting the second screening source points in the second screening source point cloud from the centroid of the second screening source point cloud, thereby obtaining a plurality of centroid-removed second screening source points;

[0054] Subtracting the second screening target points in the second screening target point cloud from the centroid of the second screening target point cloud, thereby obtaining a plurality of second screening target points without centroids;

[0055] Based on the second screening source point removed from the centroid and the second screening target point removed from the centroid, the closest point pairs are formed according to the corresponding relationship, and all the closest point pairs are obtained;

[0056] The covariance matrix is ​​constructed based on all the closest point pairs.

[0057] Preferably, the method for obtaining the mean square error value is:

[0058] Performing singular value decomposition based on the covariance matrix to obtain a first orthogonal matrix and a second orthogonal matrix;

[0059] Obtain a rotation matrix based on the first orthogonal matrix and the second orthogonal matrix;

[0060] Obtaining a translation vector based on the rotation matrix, the second filtered source point cloud centroid, and the second filtered target point cloud centroid;

[0061] The mean square error value is obtained based on the translation vector, the rotation matrix and the second set of screening point pairs.

[0062] Preferably, judging and outputting the final registration source point cloud based on the mean square error value specifically includes:

[0063] Determine whether the mean square error value is less than a preset value;

[0064] If yes, transform all source points in the source point cloud based on the rotation matrix and the translation vector to obtain a new source point cloud and output it as the final registered source point cloud;

[0065] Otherwise, all source points in the source point cloud are transformed based on the rotation matrix and the translation vector to obtain a new source point cloud as the first process source point cloud;

[0066] The source point cloud of the first process is used to replace the source point cloud, and the process of solving the mean square error value is continued and judged until the end condition is met, and the new source point cloud outputted last time is used as the final registration source point cloud.

[0067] The point cloud registration system based on the angle between the nearest point pair vector and the centroid vector includes: a data acquisition module, a first screening module, a vector acquisition module, an angle acquisition module, a second screening module, an error value acquisition module and a registration output module;

[0068] The data acquisition module is used to acquire the target point cloud and the source point cloud as original data;

[0069] The first screening module is configured to obtain a first screening point pair set and a first screening point cloud based on a minimum distance between the target point cloud and the source point cloud;

[0070] The vector acquisition module is configured to obtain a target point-source point vector set based on the first screening point pair set; and obtain a centroid vector based on the original data or the first screening point cloud;

[0071] The angle acquisition module is used to obtain the angle parameter based on the target point-source point vector set and the centroid vector;

[0072] The second screening module is configured to screen the first screening point pair set based on the angle parameter to obtain a second screening point pair set and a second screening point cloud;

[0073] The error value acquisition module is configured to obtain a covariance matrix based on the second filtered point cloud; and obtain a mean square error value based on the covariance matrix and the second filtered point pair set;

[0074] The registration output module is used to determine and output the final registration source point cloud based on the mean square error value.

[0075] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a point cloud registration method and system based on the angle between the closest point pair vector and the centroid vector, which has the following beneficial effects:

[0076] 1. The present invention enhances the geometric consistency in the matching process by introducing the angle constraint between the centroid vector and the nearest point vector of the source point cloud and the target point cloud, avoids the registration error caused by ignoring the geometric structure, and ensures that the registration result can still maintain a high degree of accuracy even in the presence of noisy or irregular point clouds; the iterative process is guided by the angle between the centroid vector and the nearest point vector, thereby improving the quality of precise registration and avoiding the problem of non-convergence or convergence to a local optimal solution caused by incorrect matching point pairs.

[0077] 2. The present invention only relies on the relationship between the geometric information of the point cloud and the centroid vector for matching, has physical interpretability, the calculation process is intuitive and efficient, and can avoid the time and resource overhead of the training and inference stages of the deep learning model. This method can still provide efficient and accurate point cloud registration under resource-constrained conditions, which has significant advantages.

[0078] 3. This method uses the target point cloud centroid-source point cloud centroid vector to reflect the global motion trend of the entire point cloud, while the target point-source point vector represents the motion trend of the local matching point pair. Only when the angle between the two is small enough can the matching of the point pair be considered geometrically reasonable; otherwise, the point pair may be caused by noise, occlusion, or mismatch and should be eliminated. The directional consistency constraint ensures that the overall trend of the matching point pair in space is consistent with the main transformation direction of the point cloud, thereby effectively eliminating abnormal point pairs caused by noise or mismatch, which is more in line with the characteristics of rigid transformation, ensuring the geometric consistency of the point cloud during the registration process, and effectively reducing the transformation error caused by mismatch. Due to the reduction of mismatched points, the registration quality can be further improved when the point cloud after coarse registration is fine-registered, making the final result more stable and accurate.

[0079] 4. After filtering the matching point pairs for directional consistency and bidirectional verification, the present invention retains the more reliable matching point pairs that are more consistent with the global geometric relationship. When these matching points are used to calculate the transformation in each iteration, the estimated rotation and translation parameters will be closer to the actual global transformation, which helps to avoid falling into local optimality, thereby improving the convergence speed and final accuracy of the ICP algorithm.

[0080] 5. The present invention combines directional consistency checking with bidirectional matching verification to further ensure the reliability of matching pairs. Bidirectional verification requires that the closest point pair found from the target to the source can also find a corresponding match from the source to the target. This bidirectional constraint can eliminate errors that may exist in unidirectional closest point matching and improve the stability and accuracy of the overall alignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0082] Figure 1 Flowchart of the point cloud registration method based on the angle between the closest point pair vector and the centroid vector provided by the present invention.

[0083] Figure 2 This is a flow chart of the method for obtaining the covariance matrix provided by the present invention.

[0084] Figure 3 This is an overall schematic diagram of the registration results provided by the present invention.

[0085] Figure 4 This is a schematic diagram of the head registration result provided by the present invention.

[0086] Figure 5 This is a schematic diagram of the tail of the registration result provided by the present invention.

[0087] Figure 6 Schematic diagram of the structure of the point cloud registration system based on the angle between the closest point pair vector and the centroid vector provided by the present invention. DETAILED DESCRIPTION

[0088] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0089] Example 1

[0090] like Figure 1 As shown, an embodiment of the present invention discloses a point cloud registration method based on the angle between the closest point pair vector and the centroid vector, including:

[0091] Obtain target point cloud and source point cloud as raw data;

[0092] Bidirectional screening based on the minimum distance between the midpoints of the target point cloud and the source point cloud obtains a first screening point pair set and a first screening point cloud;

[0093] Obtain a target point-source point vector set based on the first filtered point pair set;

[0094] Obtaining a centroid vector based on the original data or the first filtered point cloud;

[0095] Obtain the angle parameter based on the target point-source point vector set and the centroid vector;

[0096] The first screening point pair set is screened based on the angle parameter to obtain a second screening point pair set and a second screening point cloud;

[0097] Obtaining a covariance matrix based on the second filtered point cloud;

[0098] Obtain the mean square error value based on the covariance matrix and the second screening point pair set;

[0099] The final registration source point cloud is determined and output based on the mean square error value.

[0100] Example 2

[0101] Point cloud data is a collection of three-dimensional spatial data points acquired by sensors (such as laser scanners and depth cameras). It is widely used in architecture, robotics, autonomous driving, virtual reality (VR), and other fields. Point cloud registration, as one of the core technologies for processing this data, is responsible for aligning and fusing multiple point cloud datasets into a unified 3D model. Because data acquired by multiple sensors or from multiple viewpoints is unlikely to be completely consistent, point cloud registration is required to optimize the alignment of different point cloud data.

[0102] Existing point cloud registration methods usually rely on simple nearest neighbor matching to select corresponding points, but this method is easily affected by noise, outliers and coarse registration errors, resulting in inaccurate matching and thus affecting the calculation accuracy of the transformation matrix.

[0103] Traditional point cloud registration methods primarily rely on the shortest point-to-point distance for matching, ignoring the geometric structure of the point cloud, particularly the relative orientation between points. This can lead to suboptimal registration results when the point cloud contains complex geometries or is subject to some noise. Traditional methods rely on a simple iterative process for fine registration after coarse registration. However, due to the insufficient precision of the point pair matching used, the iterative process can lead to local optimal solutions, resulting in poor fine registration results.

[0104] Due to the low registration accuracy of existing methods, the final transformation matrix is ​​often accompanied by large registration errors, especially when the point cloud density is uneven or the point cloud distribution is asymmetric, the error may be more significant.

[0105] In view of the above problems of the existing methods, an embodiment of the present invention discloses a point cloud registration method based on the angle between the closest point pair vector and the centroid vector, including:

[0106] Get the target point cloud and source point cloud as raw data.

[0107] Preferably, in this embodiment, the target point cloud and the source point cloud are coarsely registered, wherein the target point cloud is a known reference data set, and the source point cloud is the point cloud data to be registered; the point cloud data to be registered includes: building source point cloud data, robot source point cloud data, autonomous driving source point cloud data or virtual reality (VR) source point cloud data.

[0108] Based on the minimum distance between the midpoints of the target point cloud and the source point cloud, a first filtered point pair set and a first filtered point cloud are obtained by bidirectional screening.

[0109] Preferably, the method for obtaining the first screening point pair set and the first screening point cloud is:

[0110] Based on the current source point in the source point cloud, the target point with the smallest distance in the target point cloud is obtained as the first selected point;

[0111] Based on the first selected point, a source point with the smallest distance in the source point cloud is obtained as the second selected point;

[0112] Determine whether the second selected point is consistent with the current source point;

[0113] If they match, it indicates that the match is successful, and the current source point is used as the first screening source point, the first selected point is used as the first screening target point, and the first screening source point and the first screening target point form the first screening point pair;

[0114] Otherwise, it indicates that the matching is unsuccessful;

[0115] Repeat the above process to search and filter all source points in the source point cloud to obtain all first filtered point pairs to form a first filtered point pair set;

[0116] All first screening source points obtained based on the first screening point pair set are sequentially composed of a first screening source point cloud;

[0117] All first screening target points obtained based on the first screening point pair set are sequentially composed of a first screening target point cloud;

[0118] A first screening point cloud is composed based on the first screening source point cloud and the first screening target point cloud.

[0119] Preferably, based on the source point cloud P={p1,p2,...,p n The current source point p1 in the target point cloud Q={q1,q2,...,q n} to obtain the target point with the smallest distance, which is used as the first selected point q5; where n represents the number of source points or target points, that is, the number of source points and target points is the same;

[0120] ;

[0121] q j Represents all target points in the target point cloud Q;

[0122] Based on the first selected point q5, the source point with the smallest distance in the source point cloud P is obtained as the second selected point;

[0123] If the second selected point is p1, it means that p1 and q5 are a pair of points that have passed the bidirectional verification, and p1 is used as the first screening source point p x1 , take q5 as the first screening target point q y1 , forming the first screening point pair (p x1 ,q y1 );

[0124] ;

[0125] p i Represents all source points in the source point cloud P;

[0126] Repeat the above process to obtain all the first screening point pairs to form the first screening point pair set;

[0127] All the first screening source points obtained based on the first screening point pair set are sequentially composed of the first screening source point cloud {p x1 ,p x2 ,...,p xa};

[0128] All the first screening target points obtained based on the first screening point pair set are sequentially composed of the first screening target point cloud {q y1 ,q y2 ,...,q ya}, xa=ya≤n, xa represents the number of the first screening source points, and ya represents the number of the first screening target points.

[0129] Preferably, the first screening source point p xi The target point obtained by finding the nearest point in the target point cloud Q is q yi , the first screening target point q yi The source point obtained by finding the nearest point in the source point cloud P is p xi , and at this time p x1 Corresponding to q y1 , p x2 Corresponding to q y2 ...finally xa Corresponding to q yaHowever, in actual situations, due to the different viewpoints of the point cloud obtained by the acquisition equipment, the positions of the points on the surface of the object are different and the number of points is also different. In actual applications, there may be multiple source points corresponding to one target point. (p x1 ,q y1 ) and (p x2 ,q y2 ) are two closest points, but at this time it is possible that q y1 With q y2 is the same target point, i.e. q y1 and q y2 Belong to the same point in the target point cloud Q.

[0130] Preferably, screening based on directional consistency between point pairs is combined with bidirectional verification, which ensures the consistency of the matching point pairs.

[0131] A target point-source point vector set is obtained based on the first filtered point pair set.

[0132] Preferably, a target point-source point vector is obtained based on each first screening point pair in the first screening point pair set:

[0133] v i = q yi -p xi i∈[1,a];

[0134] v i represents the i-th target point-source point vector, q yi represents the i-th first screening target point, p xi represents the i-th first screening source point;

[0135] A target point-source point vector set is obtained based on all target point-source point vectors.

[0136] The centroid vector is obtained based on the original data or the first filtered point cloud.

[0137] Preferably, the method for obtaining the centroid vector is as follows:

[0138] Based on all the first filtered source points p in the first filtered source point cloud xi Get the centroid of the local k-nearest neighbor points and get the centroid C1 of the first source point cloud:

[0139] ;

[0140] Among them, k represents the acquisition of the first screening source point p xi The total number of nearest points nearby, j represents p xi Nearest point index, Indicates the first screening source point p xi Local k-nearest neighbor points in the first filtered source point cloud;

[0141] Based on all the first filtered target points q in the first filtered target point cloud yi Get the centroid of the local k-nearest neighbor points and obtain the centroid C2 of the first target point cloud:

[0142] ;

[0143] in, Indicates the first screening target point q yi Local k-nearest neighbor points in the first filtered target point cloud;

[0144] The difference between the first target point cloud centroid and the first source point cloud centroid is used as the first centroid vector u1:

[0145] u1=C2-C1;

[0146] Based on the first centroid vector u1 as the centroid vector u t .

[0147] Preferably, the centroid of the k nearest neighbor points is calculated in the local neighborhood of each matching point in the respective point cloud, and the line vector connecting these two local centroids is used for subsequent calculations, which is biased towards local feature analysis and consistency judgment based on local structure rather than global vector direction screening.

[0148] Preferably, in another preferred embodiment, a global centroid line vector is used to replace a local centroid line vector, and a centroid vector acquisition method is specifically as follows:

[0149] Based on the source point cloud P={p1,p2,...,p n} to obtain the centroid of the second source point cloud :

[0150] ;

[0151] Among them, p i represents the i-th source point, i∈[1,n];

[0152] Based on the target point cloud Q={q1,q2,...,q n} to obtain the second target point cloud centroid :

[0153] ;

[0154] Among them, q j represents the j-th target point, j∈[1,n];

[0155] Based on the second target point cloud centroid and the second source point cloud centroid The difference is taken as the second centroid vector u2:

[0156] ;

[0157] Based on the second centroid vector u2 as the centroid vector u t , that is, the centroid vector u t Including the first centroid vector u1 or the second centroid vector u2, t=[1,2].

[0158] The angle parameter is obtained based on the target point-source point vector set and the centroid vector.

[0159] Preferably, the method for obtaining the angle parameter is:

[0160] Based on the target point-source point vector v in the target point-source point vector set i Get the cosine value of the vector angle with the first centroid vector u1 as the first angle parameter :

[0161] ;

[0162] Based on the target point-source point vector v in the target point-source point vector set i The cosine value of the vector angle with the second centroid vector u2 is obtained as the second angle parameter :

[0163] ;

[0164] Angle parameters Including the first angle parameter Or the second angle parameter .

[0165] The first screening point pair set is screened based on the angle parameter to obtain a second screening point pair set and a second screening point cloud.

[0166] Preferably, the second screening point pair set and the second screening point cloud acquisition method are:

[0167] Get the angle threshold;

[0168] Get the cosine value of the angle threshold as the threshold parameter;

[0169] Determine whether the angle parameter is greater than or equal to the threshold parameter;

[0170] If so, retain the first screening point pair corresponding to the angle parameter as the second screening point pair;

[0171] Otherwise, the first screening point corresponding to the angle parameter is not retained;

[0172] Circularly determine the angle parameters corresponding to all first screening point pairs, and obtain all second screening point pairs as the second screening point pair set;

[0173] forming a second screening target point cloud based on all second screening target points in the second screening point pair set;

[0174] forming a second screening source point cloud based on all second screening source points in the second screening point pair set;

[0175] The second screening point cloud is composed based on the second screening target point cloud and the second screening source point cloud.

[0176] Preferably, in this embodiment, the angle threshold is 30°; the cosine value based on the angle threshold is used as the threshold parameter , filter to meet ≥ The first screening point pair is used as the second screening point pair; a second screening point pair set is formed based on all the second screening point pairs; a second screening target point cloud is formed based on all the second screening target points in the second screening point pair set {q w1 ,q w2 ,...,q wb}; Based on all the second screening source points in the second screening point pair set, the second screening source point cloud {p h1 ,p h2 ,...,p hb}, and wb=hb≤xa=ya;q wb represents the bth second screening target point, p hb Indicates the bth second screening source point.

[0177] Preferably, the screening based on the consistency of the direction between point pairs is combined with bidirectional verification. The bidirectional verification ensures the consistency of the matching point pairs; the angle screening further optimizes the quality of the matching point pairs. The combination of the two can effectively reduce mismatches, improve the accuracy, stability and robustness of the registration, and thus improve the final point cloud alignment effect.

[0178] A covariance matrix is ​​obtained based on the second filtered point cloud.

[0179] Preferably, Figure 2 As shown, the method to obtain the covariance matrix is:

[0180] Based on the second filtered source point cloud and the second filtered target point cloud, the second filtered source point cloud centroid and the second filtered target point cloud centroid are obtained;

[0181] Subtracting the second screening source points in the second screening source point cloud from the centroid of the second screening source point cloud respectively to obtain a plurality of centroid-removed second screening source points;

[0182] Subtracting the second screening target points in the second screening target point cloud from the centroid of the second screening target point cloud, thereby obtaining a plurality of second screening target points without centroids;

[0183] Based on the second screening source point and the second screening target point without the centroid, the closest point pairs are formed according to the corresponding relationship to obtain all the closest point pairs;

[0184] The covariance matrix is ​​constructed based on all the closest point pairs.

[0185] Preferably, the second screening source point cloud centroid is :

[0186] ;

[0187] p hi Represents the i-th second screening source point.

[0188] Preferably, the second screening target point cloud centroid is :

[0189] ;

[0190] q wi Represents the i-th second screening target point.

[0191] Preferably, the second screening source point of the i-th decentering ;

[0192] The second target point for the i-th centroid removal =q wi - .

[0193] Preferably, the covariance matrix H is constructed based on all the closest point pairs consisting of the corresponding relationship between the second screening source points without the centroid and the second screening target points without the centroid:

[0194] ;

[0195] Wherein, H is a 3×3 matrix, which represents the correlation between the second screening source point cloud and the second screening target point cloud.

[0196] The mean square error value is obtained based on the covariance matrix and the second set of filtered point pairs.

[0197] Preferably, the method for obtaining the mean square error value is:

[0198] Perform singular value decomposition based on the covariance matrix H:

[0199] ;

[0200] Get the first orthogonal matrix U and the second orthogonal matrix V, both U and V are 3×3 matrices; represents a diagonal matrix (3×3) containing singular values.

[0201] The rotation matrix R is obtained based on the first orthogonal matrix U and the second orthogonal matrix V:

[0202] R=VU T ;

[0203] Based on the rotation matrix R, the second filter source point cloud centroid and secondly filter the target point cloud centroid Get the translation vector L:

[0204] ;

[0205] A mean square error value is obtained based on the translation vector, the rotation matrix, and the second set of filtered point pairs.

[0206] Preferably, a mean square error (MSE) is obtained based on the translation vector L, the rotation matrix R, and all second screening point pairs in the second screening point pair set:

[0207] .

[0208] The final registration source point cloud is determined and output based on the mean square error value.

[0209] Preferably, judging and outputting the final registration source point cloud based on the mean square error value specifically includes:

[0210] Determine whether the mean square error value is less than the preset value;

[0211] If so, all source points in the source point cloud are transformed based on the rotation matrix and translation vector to obtain a new source point cloud and output it as the final registration source point cloud;

[0212] Otherwise, all source points in the source point cloud are transformed based on the rotation matrix and the translation vector to obtain a new source point cloud as the first process source point cloud;

[0213] The source point cloud of the first process is used to replace the source point cloud, and the above process of solving the mean square error value is continued and judged until the end condition is met, and the new source point cloud outputted last time is used as the final registration source point cloud.

[0214] Preferably, all source points in the source point cloud are transformed based on the rotation matrix and translation vector to obtain a new source point cloud P new :

[0215] P new ={P new,i = +L, i=1,2,...,n};

[0216] Among them, P new,i Represents the i-th new source point obtained by transformation.

[0217] Preferably, the translation vector L and the rotation matrix R together form a transformation matrix, and when the final registration source point cloud is output, the corresponding transformation matrix is ​​output as the final transformation matrix.

[0218] Preferably, the termination conditions in this embodiment include: the mean square error value output by the current iteration is less than a preset value, the difference between the mean square error value output by the current iteration and the mean square error value output by the previous iteration is less than a preset threshold, or the preset number of iterations is reached.

[0219] Preferably, if the mean square error value is greater than or equal to the preset value, the output condition is not met, and a new source point cloud is obtained as the first process source point cloud to replace the source point cloud P and the target point cloud Q to continue the above two point cloud pair screening, construct the covariance matrix, calculate the mean square error value and judge whether the output condition is met. If not, the new source point cloud obtained based on the current rotation matrix and translation vector transformation is used as the second process source point cloud to replace the first process source point cloud and the target point cloud Q to continue the above process, and so on, until the output condition is met, and the new source point cloud output for the last time is used as the final alignment source point cloud, and the corresponding final transformation matrix and target point cloud are output at the same time.

[0220] Preferably, Figure 3-Figure 5 As shown, the Stanford University rabbit point cloud data to be registered, i.e., the source point cloud, is processed based on the method of the present invention, wherein red is the source point cloud and green is the target point cloud. From the registration results, it can be seen that the red and green are closely integrated in the ears, back, legs and other positions, with small misalignment, which further proves that the method of the present invention improves the accuracy and stability of point cloud registration.

[0221] Example 3

[0222] like Figure 6 As shown, the point cloud registration system based on the angle between the nearest point pair vector and the centroid vector includes: a data acquisition module, a first screening module, a vector acquisition module, an angle acquisition module, a second screening module, an error value acquisition module and a registration output module;

[0223] A data acquisition module is used to obtain the target point cloud and the source point cloud as raw data;

[0224] A first screening module is configured to obtain a first screening point pair set and a first screening point cloud based on a minimum distance between a target point cloud and a source point cloud;

[0225] A vector acquisition module is used to obtain a target point-source point vector set based on the first filtered point pair set; and obtain a centroid vector based on the original data or the first filtered point cloud;

[0226] An angle acquisition module, used to obtain angle parameters based on the target point-source point vector set and the centroid vector;

[0227] A second screening module is used to screen the first screening point pair set based on the angle parameter to obtain a second screening point pair set and a second screening point cloud;

[0228] An error value acquisition module is used to obtain a covariance matrix based on the second filtered point cloud; and obtain a mean square error value based on the covariance matrix and the second filtered point pair set;

[0229] The registration output module is used to judge and output the final registration source point cloud based on the mean square error value.

[0230] Preferably, the functional implementation of each functional module in this embodiment corresponds to the above method one by one, and will not be described in detail here.

[0231] Example 4

[0232] Based on the same inventive concept, the present invention further provides a computer device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0233] Memory for storing computer programs;

[0234] The processor, when used to execute the program stored in the memory, can implement the point cloud registration method based on the angle between the closest point pair vector and the centroid vector as in Example 1 or 2.

[0235] The electronic device may include: a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may call logic instructions in the memory to execute the point cloud registration method based on the angle between the closest point pair vector and the centroid vector in embodiment 1 or 2.

[0236] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0237] Through the above technical solutions, it can be seen that the present invention discloses a point cloud registration method and system based on the angle between the closest point pair vector and the centroid vector, which has the following beneficial effects:

[0238] 1. The present invention enhances the geometric consistency in the matching process by introducing the angle constraint between the centroid vector and the nearest point vector of the source point cloud and the target point cloud, avoids the registration error caused by ignoring the geometric structure, and ensures that the registration result can still maintain a high degree of accuracy even in the presence of noisy or irregular point clouds; the iterative process is guided by the angle between the centroid vector and the nearest point vector, thereby improving the quality of precise registration and avoiding the problem of non-convergence or convergence to a local optimal solution caused by incorrect matching point pairs.

[0239] 2. The present invention only relies on the relationship between the geometric information of the point cloud and the centroid vector for matching, has physical interpretability, the calculation process is intuitive and efficient, and can avoid the time and resource overhead of the training and inference stages of the deep learning model. This method can still provide efficient and accurate point cloud registration under resource-constrained conditions, which has significant advantages.

[0240] 3. This method uses the target point cloud centroid-source point cloud centroid vector to reflect the global motion trend of the entire point cloud, while the target point-source point vector represents the motion trend of the local matching point pair. Only when the angle between the two is small enough can the matching of the point pair be considered geometrically reasonable; otherwise, the point pair may be caused by noise, occlusion, or mismatch and should be eliminated. The directional consistency constraint ensures that the overall trend of the matching point pair in space is consistent with the main transformation direction of the point cloud, thereby effectively eliminating abnormal point pairs caused by noise or mismatch, which is more in line with the characteristics of rigid transformation, ensuring the geometric consistency of the point cloud during the registration process, and effectively reducing the transformation error caused by mismatch. Due to the reduction of mismatched points, the registration quality can be further improved when the point cloud after coarse registration is fine-registered, making the final result more stable and accurate.

[0241] 4. After filtering the matching point pairs for directional consistency and bidirectional verification, the present invention retains the more reliable matching point pairs that are more consistent with the global geometric relationship. When these matching points are used to calculate the transformation in each iteration, the estimated rotation and translation parameters will be closer to the actual global transformation, which helps to avoid falling into local optimality, thereby improving the convergence speed and final accuracy of the ICP algorithm.

[0242] 5. The present invention combines directional consistency checking with bidirectional matching verification to further ensure the reliability of matching pairs. Bidirectional verification requires that the closest point pair found from the target to the source can also find a corresponding match from the source to the target. This bidirectional constraint can eliminate errors that may exist in unidirectional closest point matching and improve the stability and accuracy of the overall alignment.

[0243] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0244] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A point cloud registration method based on the angle between the closest point pair vector and the centroid vector, characterized in that: include: Obtain target point cloud and source point cloud as raw data; Bidirectional screening based on the minimum distance between the midpoints of the target point cloud and the source point cloud obtains a first screening point pair set and a first screening point cloud; The method for obtaining the first screening point pair set and the first screening point cloud is: Obtaining a target point with the smallest distance in the target point cloud based on the current source point in the source point cloud as the first selected point; Based on the first selected point, a source point with the smallest distance in the source point cloud is obtained as a second selected point; Determining whether the second selected point is consistent with the current source point; If they are consistent, it indicates that the match is successful, and the current source point is used as the first screening source point, the first selected point is used as the first screening target point, and the first screening source point and the first screening target point form the first screening point pair; Otherwise, it indicates that the matching is unsuccessful; Repeat the above process to search and filter all source points in the source point cloud to obtain all the first filtered point pairs to form the first filtered point pair set; All first screening source points obtained based on the first screening point pair set are sequentially composed of a first screening source point cloud; All first screening target points obtained based on the first screening point pair set are sequentially composed of a first screening target point cloud; forming the first screening point cloud based on the first screening source point cloud and the first screening target point cloud; Obtaining a target point-source point vector set based on the first screening point pair set; Obtaining a centroid vector based on the original data or the first filtered point cloud; Obtaining an angle parameter based on the target point-source point vector set and the centroid vector; Filtering the first screening point pair set based on the angle parameter to obtain a second screening point pair set and a second screening point cloud; Obtaining a covariance matrix based on the second filtered point cloud; Obtaining a mean square error value based on the covariance matrix and the second screening point pair set; The final registration source point cloud is determined and output based on the mean square error value.

2. The point cloud registration method based on the angle between the closest point pair vector and the centroid vector according to claim 1, characterized in that: The method for obtaining the centroid vector is specifically as follows: Obtaining the centroid of local k-nearest neighbor points based on all the first filtered source points in the first filtered source point cloud to obtain the centroid of the first source point cloud; Obtaining the centroid of local k-nearest neighbor points based on all the first screened target points in the first screened target point cloud to obtain the centroid of the first target point cloud; A first centroid vector is determined based on the difference between the first target point cloud centroid and the first source point cloud centroid; Based on the first centroid vector as the centroid vector.

3. The point cloud registration method based on the angle between the closest point pair vector and the centroid vector according to claim 1, characterized in that: The method for obtaining the centroid vector is specifically as follows: Taking the mean of all source points in the source point cloud to obtain a second source point cloud centroid; Taking an average of the target points in the target point cloud to obtain a second target point cloud centroid; A second centroid vector is determined based on the difference between the second target point cloud centroid and the second source point cloud centroid; Based on the second centroid vector as the centroid vector.

4. The point cloud registration method based on the angle between the closest point pair vector and the centroid vector according to claim 2 or 3, characterized in that: The method for obtaining the angle parameter is: Acquire a cosine value of a vector angle based on the target point-source point vector in the target point-source point vector set and the first centroid vector as a first angle parameter; Acquire a cosine value of a vector angle based on the target point-source point vector in the target point-source point vector set and the second centroid vector as a second angle parameter; The angle parameter includes the first angle parameter or the second angle parameter.

5. The point cloud registration method based on the angle between the closest point pair vector and the centroid vector according to claim 4, characterized in that: The method for obtaining the second screening point pair set and the second screening point cloud is: Get the angle threshold; Obtaining a cosine value of the angle threshold as a threshold parameter; Determining whether the angle parameter is greater than or equal to the threshold parameter; If yes, retain the first screening point pair corresponding to the angle parameter as the second screening point pair; Otherwise, the first screening point corresponding to the angle parameter is not retained; Circularly determining the angle parameters corresponding to all the first screening point pairs to obtain all the second screening point pairs as the second screening point pair set; Composing a second screening target point cloud in sequence based on all second screening target points in the second screening point pair set; Composing a second screening source point cloud in sequence based on all second screening source points in the second screening point pair set; The second screening point cloud is composed based on the second screening target point cloud and the second screening source point cloud.

6. The point cloud registration method based on the angle between the closest point pair vector and the centroid vector according to claim 5, characterized in that: The method for obtaining the covariance matrix is: Based on the second filtered source point cloud and the second filtered target point cloud, the second filtered source point cloud and the second filtered target point cloud are transferred to the de-centroided coordinate system, and the centroid of the second filtered source point cloud and the centroid of the second filtered target point cloud are correspondingly obtained; Subtracting the second screening source points in the second screening source point cloud from the centroid of the second screening source point cloud, thereby obtaining a plurality of centroid-removed second screening source points; Subtracting the second screening target points in the second screening target point cloud from the centroid of the second screening target point cloud, thereby obtaining a plurality of second screening target points without centroids; Based on the second screening source point removed from the centroid and the second screening target point removed from the centroid, the closest point pairs are formed according to the corresponding relationship, and all the closest point pairs are obtained; The covariance matrix is ​​constructed based on all the closest point pairs.

7. The point cloud registration method based on the angle between the closest point pair vector and the centroid vector according to claim 6, characterized in that: The method for obtaining the mean square error value is: Performing singular value decomposition based on the covariance matrix to obtain a first orthogonal matrix and a second orthogonal matrix; Obtain a rotation matrix based on the first orthogonal matrix and the second orthogonal matrix; Obtaining a translation vector based on the rotation matrix, the second filtered source point cloud centroid, and the second filtered target point cloud centroid; The mean square error value is obtained based on the translation vector, the rotation matrix and the second set of screening point pairs.

8. The point cloud registration method based on the angle between the closest point pair vector and the centroid vector according to claim 7, characterized in that: Determining and outputting the final registration source point cloud based on the mean square error value specifically includes: Determine whether the mean square error value is less than a preset value; If yes, transform all source points in the source point cloud based on the rotation matrix and the translation vector to obtain a new source point cloud and output it as the final registered source point cloud; Otherwise, all source points in the source point cloud are transformed based on the rotation matrix and the translation vector to obtain a new source point cloud as the first process source point cloud; The source point cloud of the first process is used to replace the source point cloud, and the process of solving the mean square error value is continued and judged until the end condition is met, and the new source point cloud outputted last time is used as the final registration source point cloud.

9. A point cloud registration system based on the angle between the closest point pair vector and the centroid vector, applied to the point cloud registration method based on the angle between the closest point pair vector and the centroid vector as claimed in any one of claims 1 to 8, characterized in that: include: Data acquisition module, first screening module, vector acquisition module, angle acquisition module, second screening module, error value acquisition module and registration output module; The data acquisition module is used to acquire the target point cloud and the source point cloud as original data; The first screening module is configured to obtain a first screening point pair set and a first screening point cloud by bidirectional screening based on the minimum distance between the midpoints of the target point cloud and the source point cloud; The vector acquisition module is configured to obtain a target point-source point vector set based on the first screening point pair set; and obtain a centroid vector based on the original data or the first screening point cloud; The angle acquisition module is used to obtain the angle parameter based on the target point-source point vector set and the centroid vector; The second screening module is configured to screen the first screening point pair set based on the angle parameter to obtain a second screening point pair set and a second screening point cloud; The error value acquisition module is used to obtain a covariance matrix based on the second screening point cloud; Obtaining a mean square error value based on the covariance matrix and the second screening point pair set; The registration output module is used to determine and output the final registration source point cloud based on the mean square error value.

Images