Global registration algorithm for low-overlapping-rate point clouds

Abstract

The invention discloses a global registration algorithm for low-overlapping-rate point clouds, and relates to the technical field of point cloud registration, and the algorithm comprises the steps: constructing a multi-scale descriptor under a multi-radius neighborhood for key points in a source point cloud and a target point cloud, screening a rough matching set through employing an evolutionary stabilization strategy, and obtaining a first-order matching set; based on a corresponding propagation mechanism, aggregating each matching element in the first-order matching set into a global candidate set through a corresponding propagation operation; and according to the aggregated global candidate set, generating transformation parameters with the same number, screening an optimal transformation parameter through a quality function, performing coarse alignment on the point clouds, and performing fine optimization through a truncation-ICP (Inductively Coupled Plasma) algorithm. Therefore, the global registration algorithm for the low-overlapping-rate point clouds can effectively cope with the situation of missing part of the point clouds, has relatively high robustness, and improves the registration precision under the conditions of low resolution and low overlapping rate.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud registration, and in particular to a global registration algorithm for point clouds with low overlap rate. Background Art

[0002] The goal of registration is to find the optimal transformation that aligns a pair of point clouds in different coordinate systems using the transformation matrix. This is crucial for many applications in computer vision, such as 3D modeling, robotic navigation and positioning, object recognition and segmentation. However, alignment between point clouds becomes extremely challenging when the data is heavily occluded, has large transformations, or is noisy.

[0003] The iterative closest point (ICP) algorithm is widely used for point cloud registration due to its simplicity. However, it implicitly assumes that the point clouds to be registered completely overlap, which is often not true in practice. The pruned ICP algorithm extends ICP's application to partially overlapping point clouds, but it is prone to falling into local minima and relies on good initialization to achieve satisfactory registration results.

[0004] Among various registration algorithms, methods based on local descriptors are generally effective for achieving coarse point cloud registration. However, local descriptors are very sensitive to point cloud resolution and partial point cloud loss due to occlusion. Their discriminative ability is largely controlled by point cloud quality. In particular, when there is little overlap between point clouds, the established matching relationship becomes more unstable. Therefore, descriptor-based methods are prone to point cloud registration failure when there are significant missing points and little overlap. Summary of the Invention

[0005] The purpose of this invention is to provide a global registration algorithm for low-overlap point clouds, which is suitable for the registration of point clouds with low resolution and low overlap, and effectively solves the problem of point cloud registration accuracy when some point clouds are missing.

[0006] To achieve the above object, the present invention provides a global registration algorithm for low overlap point clouds, comprising the following steps:

[0007] S1. Describe the key points in the source point cloud and the target point cloud from the perspective of geometric and spatial information, construct a multi-scale descriptor under a multi-radius neighborhood, use the chi-square distance to initially obtain a rough matching set, and then use the evolutionary stability strategy to filter the rough matching set to obtain a preliminary matching set;

[0008] S2. Based on the corresponding propagation mechanism, each matching element in the primary matching set is aggregated into a global candidate set through the corresponding propagation operation;

[0009] S3. Generate the same number of transformation parameters based on the aggregated global candidate set, use the quality function to measure the distance error between the target point cloud and the transformed source point cloud, take the set of transformation parameters with the smallest quality function as the optimal transformation, perform coarse alignment on the point cloud, and then perform fine optimization using the truncation-ICP algorithm.

[0010] Preferably, in S1, the method for constructing a multi-scale descriptor includes first calculating the local reference frame LRA of the key point and the normal vector of the point in the local space of the key point; then, radially dividing the local space to divide the space into multiple scales, calculating each attribute value in the local space at multiple scales and generating a sub-histogram at the corresponding radius.

[0011] Preferably, in S1, the multi-scale descriptor includes a distance-weighted angle value attribute Θ, a height attribute h, and angle attributes α, β, and γ, and is expressed as follows:

[0012]

[0013] Where, subscript i represents the index of the neighborhood points around the key point, m is the number of neighborhood point indexes around key point i, R d is the LRA support radius of key point p, p i is the neighborhood point of the key point, L LRA is the reference frame, n is the normal vector, β j and β i The calculation formula is consistent with d j Represents the Euclidean distance between the current key point and the jth neighboring point.

[0014] Preferably, the method for constructing the multi-scale descriptor includes adopting differentiated binning strategies to optimize subhistograms according to the distribution characteristics of different attributes.

[0015] Preferably, in S1, the evolutionary stable strategy calculates the benefits through competition between matching point pairs, excludes abnormal matches with low benefits, and obtains a preliminary matching set, including:

[0016] First, a profit matrix is ​​constructed based on the rigid distance between matching point pairs. Each row and column of the profit matrix points to a matching point pair. The elements of the profit matrix represent the compatibility between the matching relationships pointed to by the corresponding rows and columns. The larger the value, the higher the compatibility between the matching point pairs.

[0017] Secondly, based on the spectrum analysis of the profit matrix, the eigenvector corresponding to the maximum eigenvalue is selected as the main eigenvector;

[0018] Then, the screening threshold is determined based on the components in the main eigenvector, and the rules followed are:

[0019] When the overlap ratio is greater than 0.5, the threshold is set to the median of V1; when the overlap ratio is less than 0.5, the threshold is set to the component with the smallest ranking among the V1 components.

[0020] Preferably, the rigid distance between matching point pairs is defined as:

[0021] d rigid =|d A -d B |;

[0022] in,

[0023] d A =||a i -a l ||;

[0024] d B =||b j -b h ||;

[0025] In the formula, (a i ,b j ) and (a l ,b h ) are two sets of matching point pairs, a i 、a l is the coordinate information of the corresponding point in the source point cloud, b j 、b h is the coordinate information of the corresponding point in the target point cloud.

[0026] Preferably, in S2, for a pairing element (a c ,b h )|a c ∈s d ,b h ∈t d , s d is the source point cloud after voxel downsampling, t d For the target point cloud after voxel downsampling, the steps of global candidate set aggregation include:

[0027] S21, calculate a c Point normal vector and s d The angle θ between the normal vectors of other points a in r , and calculate a c Dot and s d The Euclidean distance of other points a in

[0028]

[0029] Where, Indicates that a is within the neighborhood radius r cThe normal vector of the point, Indicates s d Normal vectors of other points in the neighborhood radius r, θ r represents the normal vector angle calculated under the neighborhood radius r;

[0030] S22, calculate b h With t d Euclidean distance to other points in and in t d Find the closest distance Points:

[0031]

[0032] Where ε1 represents the distance threshold, b is t d Divide b h Other points outside represents the point that satisfies ε1;

[0033] S23, calculate b h The normal vector and The angle between the normal vector and the neighborhood radius:

[0034]

[0035] Where, Indicates that b is within the neighborhood radius r h The normal vector of the point, Indicates t d Other points Normal vector under neighborhood radius r, ф r Represents the normal vector angle calculated under the neighborhood radius r.

[0036] S24. Measure angle θ r With angle ф r Similarities:

[0037]

[0038] Where Δψ is the angle θ under the neighborhood radius r r With angle ф r The average angle difference between them; ε2 is the custom threshold; r=r1,…,r s , represents different neighborhood radii; s is the number of different neighborhood radii;

[0039] S25. Under the constraints of S24, if a point is found that satisfies the two thresholds ε1 and ε2 given by the conditions, that is, in (a c ,b h ) based on which a set of matching point pairs (a, b) with the smallest difference is obtained;

[0040] When the chi-square distance between the corresponding descriptors of a and b satisfies χ 2 (a,b)<ε3, then update the primary matching set to M=M∪(a,b); otherwise, M remains unchanged.

[0041] Therefore, the present invention adopts the above-mentioned global registration algorithm for low overlap point clouds, which has the following technical effects:

[0042] The present invention adopts an evolutionary stability strategy to eliminate abnormal matches in the results, and introduces a corresponding propagation mechanism to obtain the global candidate set of each screened matching point, and then performs global screening. It can effectively deal with the situation where some point clouds are missing, has high robustness, and improves the alignment accuracy at low resolution and low overlap rate.

[0043] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of a global registration algorithm for point clouds with low overlap;

[0045] Figure 2 is a flowchart of descriptor creation in an embodiment of a global registration algorithm for a point cloud with low overlap;

[0046] Figure 3 ESS screening results in an embodiment of a global registration algorithm for a low overlap point cloud, wherein (a) is the matching result before screening at an overlap rate of 0.72, (b) is the matching result after screening at an overlap rate of 0.72, (c) is the matching result before screening at an overlap rate of 0.12, and (d) is the matching result after screening at an overlap rate of 0.12;

[0047] Figure 4 This is a result diagram of the second stage of the global registration algorithm embodiment for low-overlap point clouds, where (a) is the visual result of a correct matching relationship in the initial matching set, (b) is the visual result of the global candidate set after corresponding propagation, and (c) is the visual result of the global candidate set after RANSAC;

[0048] Figure 5 This is a global registration algorithm for low overlap point clouds. In the embodiment, the registration visual results of different data sets at low overlap rates are shown, where (a) is the result of the Chicken_view1_5 data set, (b) is the result of the ArmadilloSide0_60 data set, (c) is the result of the Chef_view1_5 data set, and (d) is the result of the T-rex_view1_5 data set. DETAILED DESCRIPTION

[0049] The present invention can be explained in more detail by the following examples. The purpose of disclosing the present invention is to protect all changes and improvements within the scope of the present invention. The present invention is not limited to the following examples.

[0050] like Figure 1 As shown, the present invention provides a global registration algorithm for low-overlap point clouds, which is mainly divided into three stages: the main work of the first stage is to create descriptors and screen matching relationships based on descriptors to obtain a preliminary matching set; the work of the second stage is to aggregate matching relationships that meet the propagation conditions based on high-compatibility matching relationships, thereby obtaining multiple groups of global candidate sets; the third stage generates hypothetical transformations for multiple groups of candidate sets and selects the optimal transformation matrix.

[0051] The first stage is divided into two parts: the construction of descriptors and the establishment and screening of matching relationships, as follows:

[0052] (1) For key points in the point cloud, they are described from the perspectives of geometric information and spatial information. A multi-scale descriptor is constructed by collecting a set of points within the multi-radius neighborhood of the key point. This embodiment adopts the LDASH descriptor construction method, which is as follows: first, the local reference frame LRA of the key point and the normal vector of the point within the local space of the key point are calculated; then, the local space is radially divided into multiple scales, and each attribute value is calculated in the local space at multiple scales and a subhistogram at the corresponding radius is generated.

[0053] See Figure 2 In this embodiment, five attributes are selected to construct the local feature descriptor: attribute Θ represents the local feature description attribute of the key point - the distance-weighted angular value attribute (DWAV), which improves the robustness of the descriptor to occlusion by weighting the surface change angle values ​​of the neighborhood points around the key point; attribute h is the height attribute, which represents the projection distance of the point in the local space on the LRF; the other three attributes α, β, and γ are angle attributes, which respectively represent the angle information between the vector formed by the key point and the points in the surrounding neighborhood and the normal vector of the point in the neighborhood.

[0054] in,

[0055]

[0056] Where, subscript i represents the index of the neighborhood points around the key point, m is the number of neighborhood point indexes around key point i, R d is the LRA support radius of key point p, p i is the neighborhood point of the key point, L LRA is the reference frame, n is the normal vector, β j and β iThe calculation formula is consistent with d j Represents the Euclidean distance between the current key point and the jth neighboring point.

[0057] Based on the distribution characteristics of different attributes, differentiated binning strategies are adopted to optimize the discriminability and robustness of the histogram. Based on existing prior knowledge, the 3σ principle partitioning strategy is implemented for h, α, β, and Θ, with the regions (-∞, -3σ] and (-∞, μ-3σ], and the regions [+3σ, +∞) and [μ+3σ, +∞) as the first and last bins, respectively, covering outliers with a probability of less than 0.3%. The regions within the intervals (-3σ, +3σ) and (μ-3σ, μ+3σ) are evenly divided to ensure finer binning in areas with higher data distribution density, avoid the problem of "a large amount of data concentrated in a few bins" caused by uniform binning, and improve the histogram's ability to distinguish local geometric details. For the attribute γ, the attribute value range is directly divided into equal widths. In this embodiment, the binning results for each attribute h, α, β, γ, and Θ are [13, 18, 15, 17, 8]. Then, within each radial partition, the frequency or probability density of each attribute value falling into the bin is counted, and each subhistogram is L2 normalized to eliminate the influence of density differences between partitions. The normalized subhistograms are concatenated in ascending order of radius scale to obtain a 355-dimensional local descriptor. By adaptively selecting the binning strategy based on attribute distribution characteristics, the LDASH descriptor significantly enhances the descriptor's robustness to complex interference while preserving local geometric details.

[0058] (2) Perform voxel downsampling on the source point cloud and the target point cloud to obtain the coordinate information of the key points A = {a1, a2, ..., a M}∈R 3×M and B={b1,b2,...,b N}∈R 3×N Then, according to the coordinate information of the key points of the two point clouds, the descriptors S = {x1, x2, ..., x m}∈R D×M and T={y1,y2,…,y N}∈E D×N The chi-square distance metric is used as a similarity metric to perform descriptor-based matching on the two point clouds, and a set of matching relationships is obtained by searching from the target point cloud to the source point cloud.

[0059] Among them, the chi-square distance strengthens the sensitivity to the difference in statistical histogram distribution by normalizing the difference term. Compared with other matching methods, the chi-square distance is more suitable for descriptors based on attribute histograms. The specific calculation formula is as follows:

[0060]

[0061] Where D is the dimension of the descriptor; x jk 、y ik are the kth element value of the descriptor vector of the source point cloud and the target point cloud respectively; ∈ is a small constant (usually 10e-6) to avoid the denominator being zero.

[0062] The chi-square distance can be used to obtain a set of coarse matching sets based on descriptors: U = {(xs, y1), ..., (xk, yq), ..., (xt, yN)}, which can be converted to coordinate points to F = {(as, b1), ..., (ak, bq), ..., (at, bN)}. Since there are a large number of mismatched point pairs in the preliminary matching results, this embodiment adopts the evolutionary stable strategy (ESS) in game theory to calculate the benefits through competition between matching point pairs, thereby eliminating abnormal matches with low benefits and obtaining an accurate preliminary matching set, as follows:

[0063] First, construct the profit matrix D according to the rigid distance between matching point pairs p The coordinate information of the key points of the known source point cloud is A={a1,a2,...,a M}∈R 3×M And the coordinate information of the key points of the target point cloud B={b1,b2,...,b N}∈R 3×N The payoff matrix stores the compatibility between the matching relationships pointed to by the rows and columns, and the compatibility is quantified by the consistency of the rigid distance between the matching point pairs.

[0064] Suppose there are two matching point pairs (a i ,b j ) and (a l ,b h ), let d A =||a i -a l ||, d B =||b j -b h ||, then the rigid distance between matching point pairs is d rigid =|d A -d B |, the quantitative formula of distance consistency is Therefore, the profit matrix D p Expressed as: D p (j,h)=rigid_term (j,h) , where j,h (j≤N,h≤N) respectively point to the matching point pair (a i ,b j ) and matching point pairs (a l ,b h), the larger the value between the matching point pairs in the benefit matrix, the higher the compatibility between the matching point pairs.

[0065] Secondly, based on the principal eigenvector method for determining stable matching relationships, spectral analysis of the payoff matrix is ​​performed to extract the weight distribution of stable matching relationships. The principal eigenvectors obtained through spectral analysis correspond to the distribution of the ESS. The magnitude of the principal eigenvector components reflects the global compatibility of the corresponding matching relationships. Matches with high components are more likely to be correct point pairs, while matches with low components are likely to be outliers. The matches corresponding to high components are selected to form the globally optimal matching set supported by the ESS.

[0066] Specifically, the payoff matrix D p The spectrum of D P =VΛV -1 , Λ is a diagonal matrix containing D p All eigenvalues ​​{λ1,λ2,...λ N}, V is the eigenvector matrix, each column of the eigenvector corresponds to an eigenvalue, and the eigenvector corresponding to the largest eigenvalue is the principal eigenvector. Assuming the principal eigenvector is V1, then D P V1 = λ1V1. Next, the threshold is determined based on the size of each component of the principal eigenvector in V1. The rule follows: when the overlap ratio is greater than 0.5, the threshold is set to the median of V1; when the overlap ratio is less than 0.5, the threshold is set to the lowest-ranked component in V1. After threshold screening, the first-stage preliminary matching set G = {(as, b1), …, (ak, bq)} is obtained.

[0067] By setting the descriptor distance threshold through ESS, the final matching result can be indirectly controlled, a large number of abnormal matching relationships can be removed, and the running time can be greatly saved while ensuring the accuracy of the entire algorithm.

[0068] However, the accuracy of the preliminary matching results obtained in the first stage depends largely on the discriminative power of the descriptors. The stronger the discriminative power of the descriptors, the more accurate the results. In practice, due to various factors such as occlusion, the matching results inevitably contain abnormal matching relationships. To this end, this embodiment designs a second stage to solve for more accurate matching relationships.

[0069] In the second stage, a correspondence propagation mechanism is introduced. Each matching element in the initial matching set is aggregated into a high-precision candidate set through a correspondence propagation operation, providing a good rough alignment for the point cloud. The principle behind the correspondence propagation mechanism is that when propagating and aggregating around correct matches, more correct matches are obtained; while when propagating and aggregating around incorrect matches, no correct matches are produced. In practice, points that simultaneously meet the three conditions are found on the global scale of the two point clouds to form a matching relationship and aggregated into the candidate set.

[0070] Specifically, assuming (a c ,b h )|a c ∈s d ,b h ∈t d is an element in the primary matching set obtained in the first stage, then the primary matching set can be set to M={(a c ,b h )}, the main steps of aggregating candidate sets include:

[0071] Step (1), calculate a c Point normal vector and s d The angle θ between the normal vectors of other points a in r , and calculate a c Dot and s d The Euclidean distance of other points a in

[0072]

[0073] Where, Indicates that a is within the neighborhood radius r c The normal vector of the point, Indicates s d Normal vectors of other points in the neighborhood radius r, θ r Represents the normal vector angle calculated under the neighborhood radius r.

[0074] Step (2), calculate b h With t d Euclidean distance to other points in and in t d Find the closest distance Points:

[0075]

[0076] Where ε1 represents the distance threshold, which is set to one-quarter of the point cloud spacing in this embodiment; b is t d Divide b h Other points outside represents the point that satisfies ε1.

[0077] Step (3), calculate b h The normal vector and The angle between the normal vector and the neighborhood radius:

[0078]

[0079] Where, Indicates that b is within the neighborhood radius r h The normal vector of the point, Indicates t d Other points Normal vector under neighborhood radius r, ф r Represents the normal vector angle calculated under the neighborhood radius r.

[0080] Step (4), measure the angle θ r With angle ф r Similarities:

[0081]

[0082] Where Δψ is the angle θ under the neighborhood radius r r With angle ф r The average angle difference between them; ε2 is a custom threshold, which is set to 6 in this embodiment; r=r1,…,r s , represents different neighborhood radii; s is the number of different neighborhood radii.

[0083] Step (5), under the constraints of step (4), if a point with two thresholds of ε1 and ε2 given by the given conditions is found, that is, in (a c ,b h ) is used to get a set of matching point pairs (a, b) with the smallest difference. When the chi-square distance of the descriptors of a and b satisfies χ 2 (a,b)<ε3, then update the primary matching set to M=M∪(a,b); otherwise, M remains unchanged.

[0084] From the above steps, we can see that the corresponding propagation mechanism is to find the matching relationship that satisfies the rigid constraints, normal vector angle and descriptor similarity threshold in the global to form a global candidate set. If there are q elements in the primary matching set obtained in the first stage, then through the corresponding propagation mechanism, each element performs the above operation to obtain q global candidate sets, which can be expressed as M1, M2, ..., M q In order to further ensure the accuracy of the matching relationship within the candidate set, this embodiment also uses RANSAC to remove mismatches from each candidate set to obtain multiple updated global candidate sets. Figure 4 .

[0085] In this phase, correspondence propagation not only ensures that globally correct matches are aggregated into the matching set, but also ensures that the candidate set does not contain too many false matches that would affect the RANSAC performance. Therefore, setting the threshold during correspondence propagation is crucial. The multiple sets of global matching candidates obtained in this phase are all likely to serve as exact matches for the final transformation. In the third phase, we will introduce methods for selecting the optimal set of matches from the matching candidate set.

[0086] In the third stage, using the q global candidate sets obtained in the second stage, the singular value decomposition (SVD) algorithm is used to generate q hypothetical transformations (R, T), where R is a 3×3 rotation matrix and T is a 3×1 translation vector. A quality function is used to measure the distance error between the target point cloud and the transformed source point cloud. Note that to accommodate partial overlap, this example only uses the first ηN points with the smallest distance to participate in the quality function calculation (η is the overlap ratio, N is the number of point clouds).

[0087] Specifically, let The transformed point cloud Zhong and b k ∈T d The nearest neighbor of b min is the set of the first ηN points with the smallest distance. The quality function is defined as:

[0088] Based on the candidate set to multiple transformations (R, T)1, (R, T)2, ..., (R, T) q , select the transformation that minimizes the quality function as the optimal transformation (R, T) best , used for coarse alignment of point clouds. Then, fine optimization is performed using the truncated-ICP algorithm.

[0089] Example 2

[0090] In order to verify the effectiveness of the global registration algorithm for low-overlap point clouds proposed in this invention, three sets of comparative experiments were designed to illustrate the effect of each improvement.

[0091] Group 1: Three normal vector calculation methods were designed: unweighted, inverse distance weighted, and Gaussian weighted. Three methods, Euclidean distance, chi-squared distance, and KD-tree search, were also designed, resulting in nine combinations to explore the impact of descriptor construction and matching methods on resolution and overlap. The results showed that in point clouds with an overlap ratio greater than approximately 30%, Gaussian weighted chi-squared distance provided strong robustness for descriptor matching at all resolutions. However, when the overlap ratio was minimal, the descriptor based on inverse distance weighting performed significantly better. This was primarily due to the small overlap between the two point clouds in the experimental dataset, resulting in some missing descriptors in the corresponding local areas. Furthermore, due to occlusion, the point density in the overlapping area varied significantly. Therefore, when constructing normal vectors, the sparse density of the collected neighborhood points within multiple radii prevented Gaussian weighting from being fully effective. This demonstrates that the descriptor based on inverse distance weighting offers superior discriminability for point clouds with missing points and uneven density at very low overlap ratios.

[0092] The second group designed a payoff matrix constructed solely from rigid distances, a payoff matrix constructed based on the angle condition between rigid distances and vector addition, and combined four different screening algorithms, considering global compatibility scores based on principal eigenvectors and a replication dynamic equation update strategy for solving matching relationships. Comparing the mismatch screening results of these four algorithms at different overlap rates and resolutions, it was found that the payoff matrix based on rigid distances better helped Ess eliminate mismatches, and that the screening results obtained using the weight calculation method based on principal eigenvectors were more accurate and stable.

[0093] The third group: verify the effect of applying the algorithm of this embodiment to low-overlap point clouds of different datasets. Figure 5 . Explanation of each evaluation index: Use the true transformation matrix to find the overlapping part (Overlap_num) between the source point cloud and the target point cloud and the matching point pairs of the overlapping part as the correct matching information, and calculate the matching point coordinates of the overlapping part of the target point cloud in the source point cloud in the matching relationship without ESS screening in stage one. If the distance between the coordinates of the correct matching information and the coordinates matched by the algorithm is less than the side length of the downsampled grid, the match is correct, otherwise it fails. When ESS screening is not performed, the number of correct matches counted is True_num(before), and Ratio(before) is the ratio of True_num(before) to (Overlap_num). After processing by the algorithm of this embodiment, the corresponding statistical indicators are True_num(after) and Ratio(after), as shown in Table 1.

[0094] Table 1 Effects of the present invention on low-overlap point clouds of different datasets

[0095]

[0096] Therefore, the present invention adopts the above-mentioned global registration algorithm for low-overlap point clouds, which can effectively deal with the situation where some point clouds are missing, has high robustness, and improves the registration accuracy at low resolution and low overlap.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A global registration algorithm for low-overlap point clouds, characterized in that: The following steps are involved: S1. Describe the key points in the source point cloud and the target point cloud from the perspective of geometric and spatial information, construct a multi-scale descriptor under a multi-radius neighborhood, use the chi-square distance to initially obtain a rough matching set, and then use the evolutionary stability strategy to filter the rough matching set to obtain a preliminary matching set; S2. Based on the corresponding propagation mechanism, each matching element in the primary matching set is aggregated into a global candidate set through the corresponding propagation operation; S3. Generate the same number of transformation parameters based on the aggregated global candidate set, use the quality function to measure the distance error between the target point cloud and the transformed source point cloud, take the set of transformation parameters with the smallest quality function as the optimal transformation, perform coarse alignment on the point cloud, and then perform fine optimization using the truncation-ICP algorithm.

2. A global registration algorithm for low overlap point clouds according to claim 1, characterized in that: In S1, the method for constructing a multi-scale descriptor includes first calculating the local reference frame LRA of the key point and the normal vector of the point in the local space of the key point; then, radially dividing the local space into multiple scales, calculating each attribute value in the local space at multiple scales and generating a sub-histogram under the corresponding neighborhood radius.

3. A global registration algorithm for low overlap point clouds according to claim 2, characterized in that: In S1, the multi-scale descriptor includes the distance-weighted angle value attribute Θ, the height attribute h, and the angle attributes α, β, and γ, which are expressed as follows: Where, subscript i represents the index of the neighborhood points around the key point, m is the number of neighborhood point indexes around key point i, R d is the LRA support radius of key point p, p i is the neighborhood point of the key point, L LRA is the reference frame, n is the normal vector, β j and β i The calculation formula is consistent with d j Represents the Euclidean distance between the current key point and the jth neighboring point.

4. A global registration algorithm for low overlap point clouds according to claim 3, characterized in that: The method for constructing multi-scale descriptors includes adopting differentiated binning strategies to optimize sub-histograms based on the distribution characteristics of different attributes.

5. The global registration algorithm for low overlap point clouds according to claim 1, characterized in that: In S1, the evolutionary stable strategy calculates the benefits through competition between matching point pairs, eliminates abnormal matches with low benefits, and obtains the initial matching set, including: First, a profit matrix is ​​constructed based on the rigid distance between matching point pairs. Each row and column of the profit matrix points to a matching point pair. The elements of the profit matrix represent the compatibility between the matching relationships pointed to by the corresponding rows and columns. The larger the value, the higher the compatibility between the matching point pairs. Secondly, based on the spectrum analysis of the profit matrix, the eigenvector corresponding to the maximum eigenvalue is selected as the main eigenvector; Then, the screening threshold is determined based on the components in the main eigenvector, and the rules followed are: When the overlap ratio is greater than 0.5, the threshold is set to the median of V1; when the overlap ratio is less than 0.5, the threshold is set to the component with the smallest ranking among the V1 components.

6. The global registration algorithm for low overlap point clouds according to claim 5, characterized in that: The rigid distance between matching point pairs is defined as: d rigid =|d A -d B |; in, d A =||a i -a l ||; d B =||b j -b h ||; In the formula, (a i ,b j ) and (a l ,b h ) are two sets of matching point pairs, a i 、a l is the coordinate information of the corresponding point in the source point cloud, b j 、b h is the coordinate information of the corresponding point in the target point cloud.

7. The global registration algorithm for low overlap point clouds according to claim 1, characterized in that: In S2, for a pairing element (a c ,b h )|a c ∈s d ,b h ∈t d , s d is the source point cloud after voxel downsampling, t d For the target point cloud after voxel downsampling, the steps of global candidate set aggregation include: S21, calculate a c Point normal vector and s d The angle θ between the normal vectors of other points a in r , and calculate a c Dot and s d The Euclidean distance of other points a in Where, Indicates that a is within the neighborhood radius r c The normal vector of the point, Indicates s d Normal vectors of other points in the neighborhood radius r, θ r represents the normal vector angle calculated under the neighborhood radius r; S22, calculate b h With t d Euclidean distance to other points in and in t d Find the closest distance The points: Where ε1 represents the distance threshold, b is t d Divide b h Other points outside represents the point that satisfies ε1; S23, calculate b h The normal vector and The angle between the normal vector and the neighborhood radius: Where, Indicates that b is within the neighborhood radius r h The normal vector of the point, Indicates t d Other points Normal vector under the neighborhood radius r, φ r Represents the normal vector angle calculated under the neighborhood radius r. S24. Measure angle θ r With angle φ r Similarities: Where Δψ is the angle θ under the neighborhood radius r r With angle φ r The average angle difference between them; ε2 is the custom threshold; r=r1,…,r s , represents different neighborhood radii; s is the number of different neighborhood radii; S25. Under the constraints of S24, if a point is found that satisfies the two thresholds ε1 and ε2 given by the conditions, that is, in (a c ,b h ) based on which a set of matching point pairs (a, b) with the smallest difference is obtained; When the chi-square distance between the descriptors of a and b satisfies χ 2 (a,b)<ε3, then update the primary matching set to M=M∪(a,b); otherwise, M remains unchanged.

Images