A coarse matching algorithm of multi-source point clouds for road fine reconstruction

Abstract

The present invention discloses a multi-source point cloud rough matching algorithm for fine road reconstruction, which is used for fine modeling and analysis of road surfaces. It uses the similarity of concavity and convexity of roughly overlapping areas of point clouds to quickly calculate the approximate ratio between point clouds, making up for the defect that 4PCS algorithms cannot match point clouds of unknown scales; in view of the low matching efficiency of 4PCS algorithms for point clouds with large data volumes, and the problem that the uniform downsampling used cannot guarantee the feature distribution and repetition rate of simplified point clouds, the original point cloud is simplified using a classic three-dimensional feature point extraction algorithm, and the sampling rate, calculation efficiency and repetition rate of the feature point cloud are compared to find the best data downsampling method; in view of the fact that 4PCS algorithms cannot effectively match point clouds with certain planar distribution characteristics, Nc‑5PCS designs a non-coplanar 5-point basis and a corresponding feature description method, and completes the matching of features of the same name based on the hypersphere indexing technology, effectively solving the matching failure caused by the feature degradation of the coplanar 4-point basis in the 4PCS algorithms.

Description

Technical Field

[0001] The invention relates to the technical field of three-dimensional point cloud processing, and in particular to a multi-source point cloud coarse matching algorithm for fine road reconstruction. Background Art

[0002] Establishing a detailed road model is the basis for efficient and accurate road quality evaluation, which helps to improve the accuracy of relevant feature extraction and classification, and assists in formulating reasonable road maintenance plans. Traditional digital highway data collection vehicles are expensive and have certain requirements for the passability of roads. They are unable to collect and reconstruct road point clouds that are severely damaged, especially unable to handle road damage surveys after severe disasters such as earthquakes. In recent years, with the continuous development of drone-related technologies, more and more scholars have begun to pay attention to the efficiency and flexibility of drone point cloud acquisition, using drones equipped with corresponding sensors to obtain road point clouds at low altitudes to complete road modeling and quality evaluation.

[0003] The use of drones can flexibly and efficiently obtain road point cloud data. Among them, although the drone LiDAR point cloud has high accuracy, the point cloud density is low and the distribution is uneven; the drone photogrammetry point cloud density is relatively high, but the accuracy is poor. In addition, the drone point cloud acquisition angle is single and the data acquisition distance is far. Therefore, the road models of the above two point clouds cannot fully reflect the detailed characteristics of the road surface. Fusion of close-range photogrammetry point cloud and drone point cloud can enrich the road surface details and obtain a more refined road surface model. At present, there are few applications of fusing multi-source point clouds for fine road modeling. In addition, during the autonomous operation of drones, ground personnel can simultaneously collect images of local areas of interest. This combination of data acquisition methods will not reduce the overall operation efficiency. It combines the flexibility and efficiency of data acquisition and provides better point cloud data for subsequent fine modeling.

[0004] The 4PCS method uses the affine invariant of the intersection of four coplanar points to obtain the coplanar four-point features between point clouds to complete point cloud matching. 4PCS is developed based on the LCP (Largest Common Pointsets) and RANSAC matching strategies. Since the algorithm does not need to calculate the local features of the point cloud, it has high robustness to overlap and noise, and is currently one of the best point cloud rough matching algorithms.

[0005] In 2008, Aiger and Mitra proposed the 4PCS algorithm based on the two-dimensional four-point affine invariant, which further reduced the complexity of point cloud matching to O(n2). Since the local features of the point cloud do not need to be calculated, the algorithm is highly robust to noise. In 2010, Li Min used surface variational estimation to extract multi-scale features of point clouds to improve the matching efficiency and accuracy of 4PCS, but the essence of the improvement is the preprocessing of point clouds, and the improvement of 4PCS is limited. In 2011, Yao et al. proposed the NPCS (N-Points Approximate Congruent Sets) algorithm based on point clusters, which matches the features of points in the clusters and the distance between point pairs as high-dimensional features. Since the feature discrimination of point clusters is higher than that of the planar four-point basis in 4PCS, the algorithm has a high matching accuracy and can complete point cloud matching with low overlap, but the algorithm is complex to calculate and has a high memory usage, and cannot effectively process large point clouds. In 2012, Lam et al. established segmentation curves based on DoN feature points to perform surface segmentation and merging of point clouds, and then used 4PCS to determine the matching relationship of segmented areas to complete point cloud matching. The segmentation and merging of shapes in the algorithm are complex and cannot handle point clouds with large data volumes. Subsequently, Lam and Greenspan improved the previous work by merging segmented areas in an unsupervised manner, combining voxel and shape hashing to improve retrieval efficiency, and using clustering methods to obtain transformation parameters. This method is sensitive to point cloud loss caused by occlusion, and the matching accuracy is directly related to the segmentation accuracy. In 2013, Qian Pengpeng et al. added curvature similarity restrictions when extracting equidistant point pairs from 4PCS, reducing the number of sets of congruent coplanar 4 points and improving the matching efficiency of the algorithm. In 2013, Theiler et al. proposed the K-4PCS (Keypoint-based 4PCS) algorithm, which performs 4PCS matching based on the 3D DoG feature point cloud of the original data, which increases the calculation efficiency of 4PCS by 7 times, and the matching accuracy is better than the original algorithm. In 2014, Theiler et al. extended the K-4PCS algorithm and used 3D DoG and 3D-Harris to downsample large point clouds. Theiler pointed out that the 3D DOG+4PCS method is more suitable for the registration of smooth surfaces and LiDAR point clouds with high contrast, and analyzed the influence of the texture properties and geometric symmetry of the point cloud on the matching algorithm. After that, Theiler et al. used MLESAC as the matching strategy in K-4PCS to improve the accuracy of the results; used the minimum distance prior value to reduce the error caused by the scanning position (Near-field Bias); clustered the matching results, and combined with parallel optimization to improve the efficiency of the algorithm.In 2014, Krishnan et al. proposed a congruent pyramid base to solve the problem that four congruent coplanar points cannot be effectively extracted from point clouds on complex terrain. The added fifth point is only used as a test of the set of candidate four points and is not used to solve the transformation matrix. The matching model of the algorithm is complex and the computational efficiency is low. In 2014, Mellado et al. proposed the Super 4PCS algorithm. When selecting four congruent points, a diagonal angle restriction is added to reduce the cost of confirmation calculation. The matching efficiency is greatly improved through the intelligent indexing technology after point cloud rasterization. The final computational complexity is reduced from O(n2+k) to O(n+m+k). At present, Super-4PCS is the most efficient algorithm among the improved 4PCS algorithms and has the best robustness to low overlap and noise. Mohamad proposed the Super G4PCS (Super Generalized 4PCS) algorithm, which uses the intelligent indexing technology in Super-4PCS to improve the computational efficiency of the General-4PCS algorithm. Although Super G4PCS can effectively reduce the number of congruent sets, it has a large computational overhead when searching for congruent 4 points. The theoretical computational complexity increases from O(n+m+c) of Super-4PCS to O(n+k1m+c), where k1 is the number of times each congruent point circularly traverses the corresponding non-coplanar points. In 2017, Huang et al. proposed the V4PCS (Volumetric 4PCS) algorithm based on Super-4PCS. The algorithm uses a tetrahedral basis to improve the distinguishability of features and establishes a link index table for the edges of the tetrahedron. When searching for congruent sets, it removes the computational overhead of intersection retrieval and angle retrieval. The algorithm efficiency is nearly 20 times higher than that of Super-4PCS. However, due to the large overhead of the algorithm in calculating and storing point pairs, it cannot handle the matching of point clouds with large data volumes.

[0006] The proportional features of the four coplanar points in 4PCS can theoretically be used to solve the similarity transformation relationship of the point cloud, but in actual calculation, the number of intermediate points stored and the computational overhead increase exponentially, resulting in algorithm matching failure. In 2013, Corsini et al. used VSA (Variational Shape Approximation algorithm) to segment the point cloud to ensure that the four coplanar points are distributed in different segmentation planes, thereby reducing the number of intermediate points and the number of congruent sets, and completed the similarity transformation relationship between the photogrammetric point cloud and the LiDAR point cloud. However, the algorithm is complex and still cannot effectively handle the registration of large amounts of point clouds. In addition, the surface segmentation effect directly affects the subsequent matching accuracy. In 2017, Zhang et al. proposed the G-super 4PCS (Generalized Super4-points Congruent Sets) algorithm. The algorithm uses the proportional solution method based on multi-scale Spin Image feature statistics proposed by Lin et al. to complete the matching between the photogrammetric point cloud and the TLS point cloud in the mountain area. However, the algorithm has a large computational overhead and the resulting proportional error is large, requiring secondary optimization of the proportion based on the point pair error fitting parabola. The algorithm has low computational efficiency and cannot effectively process large number of point clouds.

[0007] Among many point cloud matching algorithms, 4PCS algorithms do not need to calculate the local descriptors of point clouds; they are highly robust to noise and overlap; the algorithm only uses the geometric information of the point cloud, so it has a wide range of applications and is an excellent point cloud rough matching algorithm. However, this type of algorithm still cannot effectively match heterogeneous point clouds. The bottlenecks of the algorithm are mainly: ① Although the 4PCS algorithm reduces the computational complexity of the RANSAC matching algorithm to O(n2), in subsequent algorithm improvements, Super-4PCS further reduces the computational complexity to linear, but this type of algorithm still has difficulty processing point clouds with large data volumes. ②4PCS is mainly used to solve the rigid body transformation relationship between point clouds, and the algorithm cannot handle similar transformation relationships with unknown scales. ③When there is a certain distribution of planar features in the point cloud, the coplanar four-point features of the 4PCS algorithm are severely degraded, resulting in matching failure.

[0008] In order to illustrate the problem of 4PCS algorithms in matching heterogeneous point clouds, the present invention conducts Super-4PCS matching experiments by using LiDAR point clouds with unknown scales and partial planar features and close-range photogrammetry point clouds. The experimental results are shown in the attached figure. Figure 1 As shown in the attached Figure 1 It can be seen that Super-4PCS has the problems described above. Summary of the invention

[0009] In view of the above-mentioned problems, the present invention takes road fine modeling as the application background and proposes a multi-source point cloud coarse matching algorithm for road fine reconstruction. It proposes a point cloud matching algorithm for non-coplanar five-point congruent sets (Non-coplanar 5Points Congruent Sets for Robust Pairwise Surface Registration, Nc-5PCS), and combines the sample data of UAV LiDAR point cloud and close-range photogrammetry point cloud to illustrate the relevant principles and the intermediate processing process.

[0010] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0011] A multi-source point cloud coarse matching algorithm for road fine reconstruction, characterized by comprising the following steps:

[0012] Step 1: Obtain the source point cloud P and target Q of the target road;

[0013] Step 2: Based on the similarity of the degree of concavity and convexity in the roughly overlapping area of ​​the point clouds, the approximate ratio between the source point cloud P and the target point cloud Q is obtained, and the similar changes between the point clouds are converted into rigid body changes;

[0014] Step 3: Simplify the point clouds P and Q by using the 3D-Harris feature point extraction algorithm to obtain the feature source point cloud P and the feature target point cloud Q;

[0015] Step 4: Based on the approximate ratio obtained in step 2, a non-coplanar 5-point basis B is constructed in the feature source point cloud P, and a feature description of the non-coplanar 5-point basis is established;

[0016] Step 5: Using hypersphere rasterization and hash retrieval to search for an approximate congruent 5-point basis set {B'} in the feature target point cloud Q according to the feature description of the non-coplanar 5-point basis;

[0017] Step 6: Traverse B and {B'} to match and obtain the congruent non-coplanar 5-point basis and the corresponding transformation matrix H;

[0018] Step 7: Perform consistency measurement based on LCP, calculate the optimal solution under the consistency measurement, and use the obtained optimal solution as the optimal transformation matrix T;

[0019] Step 8: Based on the obtained optimal transformation matrix T, complete the road multi-source point cloud matching.

[0020] Furthermore, step 2 includes the following specific steps:

[0021] Step 21: Determine the rough overlapping area based on the GPS location information, assume that the plane equation is Ax+By+Cz+D=0 (C≠0), and convert the plane equation into z=ax+by+c, where a=-A / C, b=-B / C, and c=-D / C; randomly select 3 points from the point cloud to determine whether the three points are collinear. If not, calculate the plane equations corresponding to the 3 points, otherwise reselect 3 points;

[0022] Step 22: Calculate the distance di from each point (xi, yi, zi) in the point cloud to the obtained plane:

[0023]

[0024] Among them, a, b, c are the parameters of the plane equation z;

[0025] Step 23: Calculate the standard deviation σ of the obtained distances {di}:

[0026]

[0027] in, Take the threshold τ = 2σ;

[0028] Step 24: Determine the distance d of each point i >τ, the point is an elevation outlier and is stored as an outlier. Other points are used as inliers for plane fitting to form an outlier set {inliers} and an outlier set {outliers}, and the number of points Noutliers and Ninliers are counted respectively.

[0029] Step 25: Repeat steps 21-24 for N iterations, select the point set with the largest number of inliers for plane fitting, and perform SVD decomposition on the covariance matrix composed of all inliers to finally obtain the plane equation z=a0x+b0y+c0, and:

[0030]

[0031] Where a0, b0 and c0 are the parameters of the plane equation, p is the set correct probability, generally p = 99%, and w is the interior point probability;

[0032] Step 26: When solving the distance, only the outliers are traversed, and finally the extreme value of the concavity of the overlapping area dis=Max({dis i})for:

[0033]

[0034] Step 27: Calculate the approximate ratio between the point clouds, and the calculation formula is:

[0035]

[0036] Among them, d Q-average With d P-average are the mean values ​​of the rough overlapping areas of point cloud Q and point cloud P.

[0037] Furthermore, the steps of constructing the non-coplanar 5-point basis in step 4 include:

[0038] Step 41: Based on the hash storage of the plane inliers and outliers during the plane fitting process, the subset of outliers that are far from the plane {outliers, d outliers ≥d Localmax ×60%}, as the candidate set of non-coplanar points {p5}, for randomly selecting one non-coplanar point; the interior point set {inliers} is the candidate set of 4 coplanar points;

[0039] Step 42: Assume that the approximate overlap between the feature source point cloud P and the feature target point cloud Q is ω. If there is no estimated overlap, traverse {1, 0.75, 0.5, 0.25, ...} from large to small according to the value of ω, and estimate the side length constraint d of the four points on the bottom surface according to the overlap ω. ω :

[0040] d ω =dia p ×ω (6)

[0041] Among them, dia p is the diameter of point cloud p;

[0042] Step 43: According to the edge length constraint, set the edge length to d0 = d ω ×A, where A is a random value in the range of (0.7, 0.95). In the interior point set, two points p1 and p3 are randomly selected so that:

[0043] d0-σ≤||p1-p3||≤d0+σ (7)

[0044] Among them, σ is the error threshold of the distance;

[0045] Step 44: Randomly find the third point p2 in the inner point set and set the side length to d1 = d ω ×A, so that p2 satisfies:

[0046] d1-σ≤||p2-p3||≤d1+σ (8)

[0047] And satisfy: That is, while satisfying the side length restriction, the obtained side should be as vertical as possible;

[0048] Step 45: Randomly find the fourth point p4, so that ||p4-p2||≥d1×0.5 and satisfy:

[0049] π / 2≤∠P2P4P1≤2π / 3 (9);

[0050] Step 46: Randomly select a point in the candidate set {p5} as a non-coplanar point to form a non-coplanar five-point basis B with p1, p2, p3, and p4.

[0051] Furthermore, the characteristic description of the non-coplanar 5-point basis in step 4 is:

[0052] {d1, d2, d3, r1, r2, θ1, θ2}, where {d1, d2, r1, r2, θ1} are coplanar features in the Super-4PCS algorithm, d3 is the distance between the non-coplanar point f and the intersection point e of the bottom diagonal line, and d3 = ||fe||2, θ2 is the fitting vector of point e and The angle of In order to simplify the calculation cost of the normal vector of the bottom surface {a, b, c, d}, the normal vector average of the four points {a, b, c, d} of the bottom surface is Denoted as:

[0053]

[0054] in, Represent the normal vectors of the bottom surfaces a, b, c, and d respectively.

[0055] Furthermore, the step of searching for the congruent non-coplanar 5-point basis described in step 5 includes:

[0056] Step 51: input the target point cloud Q and the feature invariants {d1, d2, d3, r1, r2, θ1, θ2} corresponding to the non-coplanar 5-point basis B, wherein the point cloud model Q has established the in-plane point set {inliers} and the outlier point set {outliers} in step 24;

[0057] Step 52: Use the hypersphere rasterization in the Super-4PCS algorithm to calculate the intersection of the three-dimensional sphere and the point. Under a certain distance error ε, extract the corresponding equidistant point pair set {q i ,q j};

[0058] Step 53: Based on r1 and r2, the intersection points {e1} and {e2} of the point pairs are calculated, and a hash index is established based on the positions of the intersection points and the corresponding line segment directions, and the intermediate points are stored using a three-dimensional grid;

[0059] Step 54: Use the location hash to retrieve {q i ,q j}The middle point of the midpoint pair {e 1i , e 2j} is exactly in the same hash square, and the middle point pair is used as the intersection candidate set;

[0060] Step 55: Use the hash table to search for angles on the Gaussian spherical grid, take the middle point pair with the direction angle equal to θ1 as the intersection point, and establish the approximate congruent bottom surface set Bottom = {Bt1, Bt2, Bt3, ...} of the pentahedral basis B and the corresponding diagonal intersection point set Intersection = {It1, It2, It3, ...};

[0061] Step 56: Create an approximate vector for each intersection point in the set Intersection And the corresponding Gaussian sphere hash value;

[0062] Step 57: Based on each intersection position {It i} and the distance d3, using the hypersphere rasterization described in step 52, search for an alternative set of non-coplanar points f in the set of outer points {outliers} {f i};

[0063] Step 58: Based on {f i}'s normal vector Establish Gaussian spherical index and extract The subset of non-coplanar points with an angle of θ2 {f' i}, and finally obtain a set of 5 approximately identical non-coplanar points {B'}.

[0064] Furthermore, the steps of calculating the change matrix in step 6 are:

[0065] Step 61: Assume that B = {p1, p2, p3, p4, p5}, and the approximately congruent basis set in Q {B i Any basis in '} is B'={q1, q2, q3, q4, q5}, and the resulting transformation relationship is H, including rotation and scaling L and translation t, that is, the similarity transformation relationship is expressed as:

[0066] H(B′)=LB′+t (11);

[0067] Step 62: According to the scaling of step 27, similar changes between point clouds are converted into rigid body transformation relationship L:

[0068] L(q i )+t=p i (12);

[0069] Step 63: Subtract equation (6) from equation (7) to obtain:

[0070] L(q i -q j )+t=(p i -p j ) (13);

[0071] Among them, {(q i -q j )} = [q1-q2, q1-q3, q1-q4, q1-q5], that is, the edges in B are linearly independent of each other; {(p i -p j )} = [p1-p2, p1-p3, p1-p4, p1-p5], i.e., the corresponding linearly independent edges in B';

[0072] Step 64: According to formula (8), the matrix L can be obtained as follows:

[0073] L = {(p i -p j )}×{(q i -q j )} -1 (14);

[0074] Among them, L is a positive definite square matrix;

[0075] Step 65: Decompose using SVD SVD(L) = U∑V T , U and V represent rotation, ∑ represents scale, and the rotation matrix R = UV T = L, after L transformation {q i} and the center point of {p i} can be obtained by subtracting the center point of the translation vector t.

[0076] Furthermore, the steps for obtaining the optimal transformation matrix T in step 7 are:

[0077] Step 71: Assume that a point p in the source point cloud P i , the normal vector is The corresponding point in the target point cloud Q is q i , the normal vector is Approximately congruent base B i 'The corresponding transformation matrix is ​​H i , use formula (10) to perform consistency constraints and obtain LCP:

[0078]

[0079] Step 72: According to the obtained LCP, the transformation matrix H of the point cloud matching is obtained. iPerform consistency measurement and calculate the optimal solution under the consistency measurement. The optimal solution obtained is the optimal transformation matrix T.

[0080] The beneficial effects of the present invention are:

[0081] The present invention proposes a multi-source point cloud rough matching algorithm for fine road reconstruction. The data used are road point clouds obtained by UAV Lidar and road surface close-range photogrammetry point clouds. The algorithm of the present invention is an improved algorithm of Super-4PCS. Based on the matching strategy of RANSAC+LCP, the algorithm uses a similar intelligent indexing technology in Super-4PCS to complete the rapid retrieval and matching of congruent 5-point sets. The present invention completes the matching of heterogeneous point clouds with scale differences by designing a non-coplanar 5-point basis and corresponding feature descriptions, and combining the similarity of road surface roughness. Compared with the prior art, it has the following advantages:

[0082] First, the present invention obtains the approximate proportion between point clouds based on the similarity of the degree of concavity and convexity in the roughly overlapping area of ​​the point clouds, and converts the similar changes between the point clouds into rigid body changes, thus avoiding the matching failure caused by the quadratic increase in the number of congruent point pairs and congruent bases in the similarity transformation. The required scale error is small, and can be continuously optimized in the subsequent ICP precise matching, and finally approaches the real scale of the point cloud, making up for the defect that the 4PCS type algorithm cannot match point clouds of unknown scales.

[0083] Second, when matching point clouds with large amounts of data, Super-4PCS uses uniform sampling and cannot guarantee the repetition rate and feature distribution of the sampling points, affecting the subsequent matching accuracy. The present invention uses 3D-Harris to extract feature point clouds and complete point cloud sampling, laying a solid foundation for subsequent matching accuracy.

[0084] Third, when Super-4PCS matches point clouds with certain planar distribution characteristics, the obvious degradation of planar features will lead to matching failure. The present invention proposes a high-discrimination non-coplanar five-point basis and a corresponding feature description method, combined with intelligent indexing (hypersphere rasterization and hash retrieval) to complete the matching of related point clouds, effectively improving the matching success rate.

[0085] Fourth, in order to further improve the matching effect of the algorithm, the present invention adds a normal vector angle restriction of the common point pair when determining the optimal transformation matrix, thereby enhancing the consistency evaluation of the common point statistics and improving the accuracy of the matching algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is the matching experiment result diagram of Super-4PCS algorithm;

[0087] Figure 2It is the algorithm flow chart of Nc-5PCS;

[0088] Figure 3 (a)-(b) are the extraction effect diagrams of 3D-Harris feature point cloud, where (a) is the UAV Lidar point cloud and (b) is the close-range photogrammetry point cloud;

[0089] Figure 4 This is a diagram of the selection process of the non-coplanar 5-point basis in the source point cloud;

[0090] Figure 5a It is a characteristic description diagram of a non-coplanar 5-point basis; Figure 5b It is the distribution map of the normal vector slope component (Dip) of the close-range photogrammetry point cloud;

[0091] Figure 6 This is a search flow chart for non-coplanar 5-point basis;

[0092] Figure 7 (a)-(c) are the matching results of LiDAR point cloud and close-range photogrammetry point cloud, where (a) is the UAV LiDAR point cloud, (b) is the close-range photogrammetry point cloud, and (c) is the matching effect of the two point clouds;

[0093] Figure 8 Search for equidistant point pairs;

[0094] Fig. 9 is the error corresponding basis in the set of approximately congruent four points;

[0095] Fig.10 (a)-(b) is the search for four congruent points. DETAILED DESCRIPTION

[0096] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0097] Aiming at the difficult problems of low processing efficiency of 4Points Congruent Sets for Robust Pairwise Surface Registration (4PCS) for point clouds with large data volumes and inability to effectively match point clouds with certain planar distribution characteristics and point clouds with unknown scales, the present invention proposes a non-coplanar 5Points Congruent Sets for Robust Pairwise Surface Registration (Nc-5PCS) algorithm.

[0098] Attached Figure 2The computational complexity of the Nc-5PCS algorithm flow and key parts is shown in Table 1.

[0099] Table 1 Pseudocode of Nc-5PCS algorithm

[0100]

[0101]

[0102] The present invention uses the UAV Lidar point cloud and the close-range photogrammetry point cloud of the road as example data of the road multi-source point cloud, and applies the matching algorithm proposed in the present invention to perform road multi-source point cloud matching, and the algorithm is applicable to all types of road point cloud data.

[0103] 1. Problems with 4PCS algorithms

[0104] 4PCS algorithms can handle point cloud matching of similarity transformations in principle based on affine invariance. However, when the distance restriction of point pairs is removed to solve the similarity transformation relationship, the number of intermediate points that need to be calculated and stored changes from O(n) to O(n). 2 ), n is the number of point clouds; the time complexity of inserting and retrieving intermediate points based on range tree changes from O(k1logk1+k2) to O(n 2 logn+k'2), k1 is the number of equidistant point pairs to be found (usually k1 is much smaller than n), k2 and k'2 are the number of intersection points (equal to the number of congruent candidate four-point sets); the number of congruent candidate four-point sets k2 is much smaller than k'2, resulting in a large filtering calculation overhead; although Super-4PCS reduces the complexity of matching to linear, the intelligent index it uses mainly optimizes the computational efficiency of the equidistant point pair extraction part and cannot solve similarity transformations. Therefore, 4PCS algorithms cannot effectively solve point cloud matching with unknown scales.

[0105] At present, 4PCS algorithms are mainly used to solve rigid body transformation relationships. However, whether it is the optimization method of selecting point pairs in different segmentation areas of the point cloud to reduce the number of intermediate points, or the method of calculating the approximate scale based on the Spin Image statistics of the entire point cloud, the segmentation of the point cloud and the calculation of local features are very expensive, which seriously affects the efficiency of the entire algorithm and cannot effectively process a large number of point clouds.

[0106] 2. Scale solution between multi-source point clouds in Nc-5PCS algorithm

[0107] (2) Scale (i.e. approximate proportion) calculation method

[0108] Aiming at the fact that road point clouds have a certain area planar distribution feature, the present invention proposes a simple and efficient approximate proportion calculation method, which converts the subsequent similarity transformation relationship into a rigid body transformation relationship for solution. The obtained approximate proportion can be iteratively refined by the ICP algorithm to finally obtain the correct scaling coefficient.

[0109] The approximate ratio calculation method obtains the approximate ratio between point clouds by solving the local concavity and convexity similarity after plane fitting. Figure 3 As shown, since point clouds P and Q have a certain area of ​​planar distribution, the average convexity and concavity (the convexity and concavity is expressed as the depth of potholes in the road surface) can be obtained by fitting the plane. The approximate proportion between point clouds is solved by the similarity of local convexity and concavity in the overlapping area of ​​the road surface between multi-source point clouds after plane fitting. Since the UAV LiDAR point cloud data has GPS location information, and the photos obtained by close-range photogrammetry have GPS location information, although the above location information is too poor to participate in point cloud matching, it can determine the approximate position of the point cloud in the area, and then determine the rough overlapping area. Therefore, after GPS determines the rough overlapping area, the RANSAC method is first used to find the point with a small height difference (coordinate Z value) in the area as the internal point, and the least squares method is used to solve the road surface plane equation Ax+By+Cz+D=0 (C≠0). The specific approximate proportion solution method is:

[0110] S1: Determine the rough overlapping area based on GPS location information, assume that the plane equation is Ax+By+Cz+D=0 (C≠0), and convert the plane equation into z=ax+by+c, where a=-A / C, b=-B / C, c=-D / C; randomly select 3 points from the point cloud to determine whether the three points are collinear. If not, calculate the plane equation corresponding to the three points, otherwise reselect 3 points;

[0111] A, B, C, D and a, b, c are all parameters of the plane equation;

[0112] S2: Calculate each point (x i ,y i , z i ) to the distance d from the resulting plane i :

[0113]

[0114] S3: Calculate the distance {d i The standard deviation σ of}:

[0115]

[0116] in, Take the threshold τ = 2σ;

[0117] S4: Determine the distance d of each point i When >τ, the point is an elevation outlier and is stored as an outlier. Other points are used as inliers for plane fitting. The inliers set {inliers-Q} and the outliers set {outliers-Q} are established, and the number of points N is counted respectively. outliers With N in1iers , in the subsequent loop traversal, it serves as a constraint on the number of traversals;

[0118] S5: Repeat S1-S4 for N times, select the point set with the largest number of inliers for plane fitting, and solve the least squares problem by performing SVD decomposition on the covariance matrix composed of all inliers. Finally, the plane equation z=a0x+b0y+c0 is obtained, and the calculation formula for the number of iterations N is:

[0119]

[0120] Where p is the preset correct probability, usually p = 99%, and w is the interior point probability;

[0121] S6: To improve the efficiency of the algorithm, only the outliers are traversed when solving the distance, and finally the average value of the rough overlapping area concavity dis = Average ({dis i}):

[0122]

[0123] S6: Calculate the scale between point clouds. The calculation formula is:

[0124]

[0125] Among them, d Q-average and d P-average are the mean values ​​of the rough overlapping areas of point cloud Q and point cloud P;

[0126] 3. Simplification between multi-source point clouds in Nc-5PCS algorithm

[0127] The 3D Harris algorithm is simple and computationally efficient. Since it only uses the geometric properties of the point cloud, it is suitable for extracting feature points from multi-source point clouds. Therefore, the present invention uses 3D-Harris as a feature point extraction algorithm for multi-source point clouds to simplify the point clouds (since 3D-Harris is a classic algorithm for extracting features from three-dimensional point clouds, it will not be described in detail). Figure 3 Shown is the extraction effect of 3D-Harris feature point cloud of UAV Lidar point cloud and close-range photogrammetry point cloud.

[0128] 4. Construction of non-coplanar 5-point basis in Nc-5PCS algorithm

[0129] The basis used for point cloud matching is the corresponding features between point clouds and is the basis for point cloud matching. Generally speaking, design and construction are performed in the source / scene point cloud P, congruent bases are searched in the model / target point cloud Q, and the transformation matrix of the point cloud is calculated based on the corresponding point positions.

[0130] The present invention constructs a three-dimensionally distributed pentahedral basis by introducing a non-coplanar fifth point. Figure 4 As shown in the figure, the non-coplanar fifth point can significantly enhance the geometric features of the basis, greatly reducing the number of congruent bases in the model point cloud, thereby reducing the confirmation calculation overhead of the congruent set. In particular, when the point cloud has a planar distribution feature of a certain area, the advantages of the non-coplanar basis are more obvious. Figure 4 The leftmost picture in the figure is the distance distribution of the source point cloud after plane fitting. The four small pictures in the middle are, from top to bottom, the outer points of the fitting plane, the inner points of the fitting plane, the selection of non-coplanar points in the subset with larger depth values ​​among the outer points, and the selection of four bottom points in the combination of the inner points. The rightmost picture shows the five non-coplanar points finally selected.

[0131] Based on the hash storage of the plane inliers and outliers during the plane fitting process, the subset of outliers that are far from the plane {outliers, d outliers ≥d Localmax ×60%}, as the candidate set of non-coplanar points {p5}, randomly select non-coplanar points, where d outliers is the distance from the outer point to the fitting plane after the point cloud plane is fitted, d Localmax is the farthest distance from the outer point to the fitting plane; select 4 coplanar points from the inner point set {inliers} to form a non-coplanar 5-point basis, such as Figure 4 As shown, the specific steps of constructing the above non-coplanar 5-point basis are:

[0132] S1: Assume that the approximate overlap between the feature source point cloud P and the feature target point cloud Q is ω. Since the UAV LiDAR point cloud data has GPS location information, the approximate overlap can be estimated based on the positioning information. However, when there is no auxiliary information for overlap estimation, it is necessary to traverse the assumed approximate overlap ω from large to small {1, 0.75, 0.5, 0.25, ...}, and estimate the side length constraints d of the four points on the bottom surface according to the overlap ω. ω :

[0133] d ω =dia p ×ω (6)

[0134] Among them, dia p is the diameter of the point cloud p (i.e. the side length of the point cloud outer volume);

[0135] S2: According to the calculated edge length constraint, set the edge length to d0 = d ω ×A, A is a random value in the range of (0.7, 0.95) to avoid the situation where the four bottom points that meet the angle requirements cannot be found due to the wider side length; in the set of interior points, two points p1 and p3 are randomly selected so that:

[0136] d0-σ≤||p1-p3||≤d0+σ (7)

[0137] Among them, σ is the error threshold of the distance;

[0138] S3: Randomly find the third point p2 in the inner point set and set the side length to d1 = d ω ×A, so that p2 satisfies:

[0139] d1-σ≤||p2-p3||≤d1+σ (8)

[0140] And satisfy That is, while satisfying the side length restriction, the obtained side should be as vertical as possible;

[0141] S4: Randomly search for the fourth point p4 in the interior point set, so that ||p4-p2||≥d1×0.5 and satisfy:

[0142] π / 2≤∠P2P4P1≤2π / 3 (9)

[0143] S5: In the subset O of the external point set, randomly select a point as a non-coplanar point to form a non-coplanar 5-point basis B with p1, p2, p3, and p4.

[0144] Compared with the 4PCS type search for coplanar four points, the non-common surface five-point basis has obvious advantages in constructing the four points on the bottom surface: ① Since the four points on the bottom surface in the non-coplanar five points are all selected from the inner points of the plane fitting, the computational overhead of judging the coplanarity of the four points is saved. ② Since the side length and angle when searching for the four points on the bottom surface are restricted, the judgment and calculation of the concave and convex quadrilaterals on the bottom surface are also saved, and the appearance of narrow and long quadrilaterals is avoided.

[0145] 5. Characterization of non-coplanar 5-point basis in Nc-5PCS algorithm

[0146] After establishing the non-coplanar 5-point basis B, it is necessary to describe and match the features of the non-coplanar 5-point basis B. It is necessary to construct the invariant features of B, that is, to establish the invariant based on which the corresponding congruent basis is found. The constructed features should be simple and easy to store and calculate; they should also fully describe the geometric properties of B in order to reduce the number of approximately congruent bases, thereby reducing the calculation overhead of confirming the best matching basis.

[0147] By referring to the Super-4PCS algorithm, the plane features are simplified into diagonal length, angle and affine ratio, and the relevant features are efficiently stored and retrieved based on intelligent indexing. The present invention still uses the intelligent indexing method in Super-4PCS and expands the plane features into pentahedral features by adding a small number of feature elements. Figure 5a As shown in the figure, based on the coplanar features {d1, d2, r1, r2, θ1} of the original Super-4PCS, the distance d3 between the non-coplanar point f and the intersection point e of the bottom diagonal is added, d3 = ||fe||2, and the fitting vector of point e is added. and The angle θ2, Right now Figure 5a The characteristics of the non-coplanar 5-point basis shown are {d1, d2, d3, r1, r2, θ1, θ2}. Since the 4 points of the bottom surface are all selected from the interior points of the plane, as shown in Figure 5a Most of the normal vectors of the inner points shown are perpendicular to the fitting plane. In order to simplify the normal vector calculation cost of the bottom surface {a, b, c, d}, the sum of the normal vectors of the four points a, b, c, d on the bottom surface is taken as the mean normal vector of the plane, that is:

[0148]

[0149] Compared with other feature descriptions, this feature description records fewer feature elements and can effectively reduce the search and calculation overhead of approximate congruent bases; the approximate solution method for the normal vector in this feature description is simple to calculate and efficient.

[0150] 6. Searching for non-coplanar 5-point basis in Nc-5PCS algorithm (i.e. matching)

[0151] After establishing the feature description of the non-coplanar 5-point basis B in the source point cloud P, the corresponding congruent 5-point basis set is searched in the model point cloud Q based on this feature. A smart indexing method similar to that in Super-4PCS is used to complete the traversal of the congruent set. Figure 6 A schematic diagram of the search process is shown below, which includes the following steps:

[0152] S1: Input the target point cloud Q and the feature invariants {d1, d2, d3, r1, r2, θ1, θ2} corresponding to the non-coplanar 5-point basis B, where the point cloud model Q has established the in-plane point set {inliers-Q} and the outer point set {outliers-Q} in the approximate scale calculation;

[0153] S2: The hypersphere rasterization in the Super-4PCS algorithm is used to complete the intersection calculation of the three-dimensional sphere and the point. When searching for congruent non-coplanar bases, the vectors corresponding to the point pairs are mapped to points on the Gaussian sphere, and the angle matching is converted into the intersection of the spherical circle and the spherical rasterization to quickly complete the angle retrieval. During the retrieval, a hash index is established based on the coordinates of the middle point of the point pair and the normalized direction vector to quickly complete the retrieval. Under a certain distance error ε, the corresponding equidistant point pair set {q ... i ,q j}, the hypersphere gridding process is as follows Figure 8-10 As shown;

[0154] in, Figure 8 The diagonal lengths of the coplanar four-point basis {p1, p2, p3, p4} are d1 and d2. Based on this distance invariant, find equidistant point pairs in Q; Fig. 9 The four-point basis {p1, p2, p3, p4} in the point cloud P and {q1, q2, q3, q4} in the point cloud Q are corresponding congruent bases. With the intersection e as the center, all the rotation lines {(q'3, q'4)} of (q3, q4) can form a redundant four-point basis with (q1, q2) that satisfies the affine invariance. Fig.10 (a) is the included angle α of the diagonal of the four-point basis, Fig.10 (b) Find the approximately congruent four points based on the position Hash of the intermediate proportional points e1 and e2, and find the congruent coplanar four points based on the angle invariant α and the Gaussian sphere index, d is the mapping point of the diagonal vector on the Gaussian sphere;

[0155] S3: Based on r1 and r2, calculate the point pair (for example Figure 5a The middle point / intersection point set {e1} and {e2} of {a,c},{d,b}) in the image are obtained, and a hash index is established based on the position of the intersection point and the corresponding line segment direction, and the middle points are stored using a three-dimensional grid;

[0156] S4: Use the location hash to get the middle point in the same grid location {e 1i ,e 2j}, and use it as the candidate set of intersection points. Then, use the hash table to search for angles on the Gaussian spherical grid, and use the middle point pairs with a direction angle equal to θ1 as the intersection points. So far, the approximate congruent bottom surface set Bottom = {Bt1, Bt2, Bt3, ...} and the corresponding diagonal intersection point set Intersection = {It1, It2, It3, ...} of the pentahedral basis B are established;

[0157] S5: Create an approximate vector for each intersection point in the set And the corresponding Gaussian sphere hash value;

[0158] S6: Based on each intersection position {It i} and the distance d3, use the hypersphere rasterization in step S2 to search for an alternative set of non-coplanar points f in the outer point set {outliers-Q} {f i};

[0159] S7: Based on {f i} Establish a Gaussian spherical index and extract the The subset of non-coplanar points with an angle of θ2 {f' i}, and finally obtain a set of 5 approximately identical non-coplanar points {B'}.

[0160] 7. Solution of transformation matrix between multi-source point clouds in Nc-5PCS algorithm

[0161] When an approximate congruent basis {B i '}, the transformation matrix H can be calculated based on the corresponding relationship of the vertices. i , and select the optimal solution according to the consistency metric to complete the model matching. At present, the method of solving the transformation matrix has been disclosed in the prior art, so the present invention only introduces the solution steps and does not elaborate on the formula derivation.

[0162] (1) Solving the transformation matrix

[0163] Assume B = {p1, p2, p3, p4, p5}, the approximately congruent basis set in Q {B i Any basis in '} is B'={q1, q2, q3, q4, q5}, and the resulting transformation relationship is H, including rotation and scaling L and translation t, that is, the transformation relationship is expressed as:

[0164] H(B′)=LB′+t (11)

[0165] Since the scaling between point clouds has been solved before, the point clouds are in a rigid body transformation relationship:

[0166] L(q i )+t=p i (12)

[0167] Subtracting (6) from (7) yields:

[0168] L(q i -q j )+t=(p i -p j ) (13)

[0169] Among them, {(q i -q j)} = [q1-q2, q1-q3, q1-q4, q1-q5], that is, the edges in B are linearly independent of each other; {(p i -p j )} = [p1-p2, p1-p3, p1-p4, p1-p5], i.e., the corresponding linearly independent edges in B';

[0170] The rigid body change matrix L can be obtained by formula (8):

[0171] L = {(p i -p j )}×{(q i -q j )} -1 (14)

[0172] Among them, the matrix L is a positive definite square matrix;

[0173] Decompose using SVD SVD(L) = UEV T , U and V represent rotation, ∑ represents scale (it is the unit matrix under rigid body transformation), so the rotation matrix R = UV T = L, after L transformation {q i} and the center point of {p i} can be obtained by subtracting the center point of the translation vector t.

[0174] The matching (searching) algorithm of non-coplanar 5-point basis selects approximately congruent basis {B i '} and RANSAC iterations need to compare the obtained transformation matrices, so it is necessary to select the optimal solution under the consistency metric as the final matching result. At present, 4PCS algorithms generally use the number of consistent points that meet a certain distance error threshold δ for measurement, that is, the maximum number of common points (LCP).

[0175] In order to improve the matching accuracy of the algorithm, the present invention uses the distance error constraint δ and the normal vector angle constraint Assume that a point p in the source point cloud P i , the normal vector is The corresponding point in the model point cloud Q is q i , the normal vector is Approximately congruent base B i 'The corresponding transformation matrix is ​​H i , use the following formula to perform consistency constraints and obtain LCP:

[0176]

[0177] Finally, the transformation matrix parameter T that meets the above convergence conditions is obtained, and the road multi-source point cloud matching is completed. The matching results are shown in the attached Figure 7 shown.

[0178] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A multi-source point cloud coarse matching algorithm for road fine reconstruction, characterized in that: The following steps are involved: Step 1: Obtain the source point cloud P and target Q of the target road; Step 2: Based on the similarity of the degree of concavity and convexity in the roughly overlapping area of ​​the point clouds, the approximate ratio between the source point cloud P and the target point cloud Q is obtained, and the similar changes between the point clouds are converted into rigid body changes; Step 3: Simplify the point clouds P and Q by using the 3D-Harris feature point extraction algorithm to obtain the feature source point cloud P and the feature target point cloud Q; Step 4: Based on the approximate ratio obtained in step 2, a non-coplanar 5-point basis B is constructed in the feature source point cloud P, and a feature description of the non-coplanar 5-point basis is established; Step 5: Using hypersphere rasterization and hash retrieval to search for an approximate congruent 5-point basis set {B'} in the feature target point cloud Q according to the feature description of the non-coplanar 5-point basis; Step 6: Traverse B and {B'} to match and obtain the congruent non-coplanar 5-point basis and the corresponding transformation matrix H; Step 7: Perform consistency measurement based on LCP, calculate the optimal solution under the consistency measurement, and use the obtained optimal solution as the optimal transformation matrix T; Step 8: According to the obtained optimal transformation matrix T, complete the road multi-source point cloud matching; The characteristic description of the non-coplanar 5-point basis in step 4 is: {d1, d2, d3, r1, r2, θ1, θ2}, where {d1, d2, r1, r2, θ1} are coplanar features in the Super-4PCS algorithm, d3 is the distance between the non-coplanar point f and the intersection point e of the bottom diagonal line, and d3 = ||fe||2, θ2 is the fitting vector of point e and The angle of In order to simplify the calculation cost of the normal vector of the bottom surface {a, b, c, d}, the normal vector average of the four points {a, b, c, d} of the bottom surface is Denoted as: in, Represent the normal vectors of the bottom surfaces a, b, c, and d respectively.

2. The multi-source point cloud coarse matching algorithm for road fine reconstruction according to claim 1 is characterized in that: Step 2 includes the following specific steps: Step 21: Determine the rough overlapping area based on the GPS location information, assume that the plane equation is Ax+By+Cz+D=0 (C≠0), and convert the plane equation into z=ax+by+c, where a=-A / C, b=-B / C, and c=-D / C; randomly select 3 points from the point cloud to determine whether the three points are collinear. If not, calculate the plane equations corresponding to the 3 points, otherwise reselect 3 points; Step 22: Calculate each point (x i ,y i ,z i ) to the distance d from the resulting plane i : Among them, a, b, c are the parameters of the plane equation z; Step 23: Calculate the distance {d i The standard deviation σ of}: in, Take the threshold τ = 2σ; Step 24: Determine the distance d of each point i When >τ, the point is an elevation outlier and is stored as an outlier. Other points are used as inliers for plane fitting to form an outlier set {inliers} and an outlier set {outliers}, and the number of points N is counted separately. outliers With N inliers ; Step 25: Repeat steps 21-24 for N iterations, select the point set with the largest number of inliers for plane fitting, and perform SVD decomposition on the covariance matrix composed of all inliers to finally obtain the plane equation z=a0x+b0y+c0, and: Where a0, b0 and c0 are the parameters of the plane equation, p is the set correct probability, p = 99%, and w is the interior point probability; Step 26: When solving the distance, only the outliers are traversed, and finally the extreme value of the concavity of the overlapping area dis=Max({dis i })for: Step 27: Calculate the approximate ratio between point clouds, and the calculation formula is: Among them, d Q-average With d P-average are the mean values ​​of the rough overlapping areas of point cloud Q and point cloud P.

3. The multi-source point cloud coarse matching algorithm for road fine reconstruction according to claim 2 is characterized in that: The steps of constructing the non-coplanar 5-point basis in step 4 include: Step 41: Based on the hash storage of the plane inliers and outliers during the plane fitting process, the subset of outliers that are far from the plane {outliers, d outliers ≥d Localmax ×60%}, as the candidate set of non-coplanar points {p5}, for randomly selecting one non-coplanar point; the interior point set {inliers} is the candidate set of 4 coplanar points; Step 42: Assume that the approximate overlap between the feature source point cloud P and the feature target point cloud Q is ω. If there is no estimated overlap, traverse {1, 0.75, 0.5, 0.25, ...} from large to small according to the value of ω, and estimate the side length constraint d of the four points on the bottom surface according to the overlap ω. ω : d ω =day p ×ω (6) Among them, dia p is the diameter of point cloud p; Step 43: According to the edge length constraint, set the edge length to d0 = d ω ×A, where A is a random value in the range of (0.7, 0.95). In the interior point set, two points p1 and p3 are randomly selected so that: d0-σ≤||p1-p3||≤d0+σ (7) Among them, σ is the error threshold of the distance; Step 44: Randomly find the third point p2 in the inner point set and set the side length to d1 = d ω ×A, so that p2 satisfies: d1-σ≤||p2-p3||≤d1+σ (8) And satisfy: That is, while satisfying the side length restriction, the obtained side should be as vertical as possible; Step 45: Randomly find the fourth point p4, so that ||p4-p2||≥d1×0.5 and satisfy: π / 2≤P2P4P1≤2π / 3 (9); Step 46: Randomly select a point in the candidate set {p5} as a non-coplanar point to form a non-coplanar five-point basis B with p1, p2, p3, and p4.

4. The multi-source point cloud coarse matching algorithm for road fine reconstruction according to claim 3 is characterized in that: The steps of searching for the congruent non-coplanar 5-point basis described in step 5 include: Step 51: input the target point cloud Q and the feature invariants {d1, d2, d3, r1, r2, θ1, θ2} corresponding to the non-coplanar 5-point basis B, wherein the point cloud model Q has established the in-plane point set {inliers} and the outlier point set {outliers} in step 24; Step 52: Use the hypersphere rasterization in the Super-4PCS algorithm to calculate the intersection of the three-dimensional sphere and the point. Under a certain distance error ε, extract the corresponding equidistant point pair set {q i ,q j }; Step 53: Based on r1 and r2, the intersection points {e1} and {e2} of the point pairs are calculated, and a hash index is established based on the positions of the intersection points and the corresponding line segment directions, and the intermediate points are stored using a three-dimensional grid; Step 54: Use the location hash to retrieve {q i ,q j }The middle point of the midpoint pair {e 1i ,e 2j } is exactly in the same hash square, and the middle point pair is used as the intersection candidate set; Step 55: Use the hash table to search for angles on the Gaussian spherical grid, take the middle point pair with the direction angle equal to θ1 as the intersection point, and establish the approximate congruent bottom surface set Bottom = {Bt1, Bt2, Bt3, ...} of the pentahedral basis B and the corresponding diagonal intersection point set Intersection = {It1, It2, It3, ...}; Step 56: Create an approximate vector for each intersection point in the set Intersection And the corresponding Gaussian sphere hash value; Step 57: Based on each intersection position {It i } and the distance d3, using the hypersphere rasterization described in step 52, search for an alternative set of non-coplanar points f in the set of outer points {outliers} {f i }; Step 58: Based on {f i }'s normal vector Establish Gaussian spherical index and extract The subset of non-coplanar points with an angle of θ2 {f' i }, and finally obtain a set of 5 approximately identical non-coplanar points {B'}.

5. The multi-source point cloud coarse matching algorithm for road fine reconstruction according to claim 4 is characterized in that: The steps to calculate the change matrix in step 6 are: Step 61: Assume that B = {p1, p2, p3, p4, p5}, and the approximately congruent basis set in Q {B i Any basis in '} is B'={q1, q2, q3, q4, q5}, and the resulting transformation relationship is H, including rotation and scaling L and translation t, that is, the similarity transformation relationship is expressed as: H(B′)=LB′+t (11); Step 62: According to the scaling of step 27, similar changes between point clouds are converted into rigid body transformation relationship L: L(q i )+t=p i (12); Step 63: Subtract equation (6) from equation (7) to obtain: L(q i -q j )+t=(p i -p j ) (13); Among them, {(q i -q j )} = [q1-q2, q1-q3, q1-q4, q1-q5], that is, the edges in B are linearly independent of each other; {(p i -p j )} = [p1-p2, p1-p3, p1-p4, p1-p5], i.e., the corresponding linearly independent edges in B'; Step 64: According to formula (8), the matrix L can be obtained as follows: L={(p i -p j )}×{(q i -q j )} -1 (14); Among them, L is a positive definite square matrix; Step 65: Decompose using SVD SVD(L) = U∑V T , U and V represent rotation, ∑ represents scale, and the rotation matrix R = U∑ T = L, after L transformation {q i The center point of} and {p i } can be obtained by subtracting the center point of the translation vector t.

6. The multi-source point cloud coarse matching algorithm for road fine reconstruction according to claim 5 is characterized in that: The steps to obtain the optimal transformation matrix T in step 7 are: Step 71: Assume that a point p in the source point cloud P i , the normal vector is The corresponding point in the target point cloud Q is q i , the normal vector is Approximately congruent base B i 'The corresponding transformation matrix is ​​H i , use formula (10) to perform consistency constraints and obtain LCP: Step 72: According to the obtained LCP, the transformation matrix H of the point cloud matching is obtained. i Perform consistency measurement and calculate the optimal solution under the consistency measurement. The optimal solution obtained is the optimal transformation matrix T.

Images