A point cloud feature extraction method suitable for multi-source cross-scale data accurate registration
By jointly constructing a feature description matrix using global and local poses, the problems of inaccurate structure and large data connection errors in the 3D reconstruction of multi-source, multi-scale data are solved, and high-precision multi-source point cloud data matching and fusion are achieved.
Patent Information
- Application Number
- CN202210875510.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-25
- Publication Date
- 2025-11-25
- Estimated Expiration
- 2042-07-25
AI Technical Summary
In fully automated 3D reconstruction based on optical images, the separate processing of global and local data from multi-source, multi-scale data leads to problems such as inaccurate structure and large data connection errors. Existing methods are prone to large neighboring point errors in the extraction and matching stages of corresponding feature points.
A feature description matrix is constructed by jointly using global pose and local pose. By combining global pose features and local features, the feature matrix of multi-source point cloud data is calculated, including distance and angle features. A feature histogram is then constructed to achieve high-precision matching.
It improves the accuracy of multi-source point cloud data matching and fusion, reduces errors within the neighborhood range, and achieves high-precision 3D reconstruction of large-scale, multi-block, and multi-source data.
Smart Images
Figure CN115294344B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of three-dimensional reconstruction and feature matching in surveying and computer vision, and particularly relates to a point cloud feature extraction method suitable for accurate registration of multi-source cross-scale data. BACKGROUND
[0002] In the field of full-automatic three-dimensional reconstruction based on optical images, when facing large-scale, cross-scale, multi-block and multi-source data, there is a problem that the overall and local data are processed separately, resulting in inaccurate structure and poor data connection. At present, this problem is mainly solved by selecting a part with reliable position accuracy as the registration subject, and other data as the registration object, then using an algorithm to extract the homonymic feature points of the registration subject and the registration object, and realizing the matching and connection of the registration subject and the registration object through the matching of the homonymic feature points. In the extraction and matching stage of the homonymic feature points, the existing method is used to calculate the point features of the high-density point cloud and the feature similarity, which faces the problem that the features of adjacent points are extremely similar, and then the adjacent points within a certain range are likely to become matching points when the data is matched and connected, resulting in large data connection error. Solving this error and optimizing the data connection become the core problem of multi-source point cloud data registration.
[0003] Multi-source data includes airborne laser scanning data, handheld laser scanning data, unmanned aerial vehicle optical images and mobile device optical images. Since each data source has its own advantages and disadvantages in describing the target, fusion processing of multi-source data can achieve complementary advantages. Multi-source point cloud data registration is an important research direction, and there is a certain research basis in surveying and computer vision, and it is expanded into two categories of implementation methods: (1) registration using laser scanning data geometric information; (2) registration using laser scanning data intensity information.
[0004] The present application belongs to the registration method based on geometric information, and the core content of the application is to provide a high-precision point cloud geometric feature descriptor. By using the descriptor provided by the present application, through unified analysis and fusion calculation of different scale data, the overall and local registration problems in high-precision automatic three-dimensional reconstruction can be solved, so as to realize high-precision three-dimensional reconstruction of large-scale, cross-scale, multi-block and multi-source data. SUMMARY
[0005] The present application provides a point cloud feature extraction method suitable for accurate registration of multi-source cross-scale data, aiming at the deficiencies of the prior art. By jointly constructing a feature description matrix from global pose and local pose, high-precision point feature description of laser scanning point cloud and photogrammetry matching point cloud is realized, and then high-precision matching of various data sources is realized.
[0006] In order to achieve the above object, the technical scheme provided by the application is a point cloud feature extraction method suitable for multi-source cross-scale data accurate registration, comprising the following steps:
[0007] Step 1, taking the global pose description reference point of the point cloud as the reference, the distance feature and the angle feature of all points of the point cloud are calculated to obtain the global pose feature;
[0008] Step 2, taking all points of the point cloud as the origin, a local coordinate system is constructed in batches;
[0009] Step 3, the feature point set is constructed in the local coordinate system, and the local feature of all points in the feature point set is calculated;
[0010] Step 4, the feature statistics calculation is performed on all point clouds, the feature histogram is obtained, and the feature matrix is constructed.
[0011] Moreover, the calculation of the global pose description reference point in step 1 is different for different data.
[0012] The global pose description reference point of the undirected single data is the point cloud centroid, and the calculation formula is:
[0013] C = mean (X, Y, Z) (1)
[0014] In the formula, C is the global pose description reference point, (X, Y, Z) is the point cloud coordinate, and mean () represents the mean operation.
[0015] The global pose description reference point of the undirected multi-data source is the coordinate mean of the centroid of each single data point cloud, and the calculation formula is:
[0016]
[0017] In the formula, C m is the global pose description reference point, C ix is the x coordinate of the centroid of the i th point cloud data, C iy is the y coordinate of the centroid of the i th point cloud data, C iz is the z coordinate of the centroid of the i th point cloud data, and mean () represents the mean operation.
[0018] The global pose description reference point of the directed data is the coordinate mean of multiple image control points, and the calculation method is the same as formula (2).
[0019] The global pose feature is a three-tuple {D0, A0, A1} composed of 1 distance feature and 2 angle features, and the query point P Q The global pose feature matrix is obtained by calculating the feature three-tuple of each neighborhood point, and the expression form of the feature matrix is:
[0020]
[0021] where D 0i represents a distance feature, which is the length of the line segment P i P G ; A 0i represents an angle feature, which is the angle between the normal vector N i of P i P i ; A G is the angle between the normal vector N 1i of P Q P G and the line segment P i P G ; P G is the global pose description reference point; P Q is the query point, i.e., the point for which the feature is currently being calculated; and P i is the feature point obtained within the query radius of the current query point.
[0022] Furthermore, in step 2, the current point is taken as the query point, the normal vector of the query point is taken as the U-axis, the normal vector of the plane formed by the global reference point and the query normal vector is taken as the V-axis, and the direction perpendicular to both the U-axis and the V-axis is taken as the W-axis. The calculation method of the local coordinate system is as follows:
[0023] U = norm(N q ) (4)
[0024] V = U x norm(N j ) (5)
[0025] W = U x V (6)
[0026] where N j is the direction vector of the global pose description reference point P G pointing to the query point P Q , and N q is the normal vector of P Q .
[0027] Furthermore, in step 3, the local pose feature is a five-tuple {L, B, C, D, E} composed of one distance feature and four angle features, which is calculated for each neighborhood point P Q to obtain a global pose feature matrix, and the expression form of the feature matrix is as follows:
[0028]
[0029] where L i represents a distance feature, which is the length of the line segment P i P Q ; B irepresents an angle feature, P i represents the normal vector N i of the angle with the local coordinate axis U; C i represents an angle feature, P i represents the normal vector N i of the angle with the local coordinate axis V; D i represents an angle feature, P i represents the normal vector N i of the angle with the local coordinate axis W; E i represents an angle feature, P i represents the normal vector N i of the angle with the line segment P i P Q ; P Q is a query point, i.e. a point for which the feature needs to be calculated; U, V, W are local coordinate systems constructed based on P Q ; P i is a feature point obtained within a query radius of the current query point; N i is the normal vector of P i .
[0030] Moreover, the step 4 is used for calculating the feature statistics of all point clouds, obtaining a feature histogram, setting a bin value as N when constructing the histogram, and obtaining n1 feature elements a1 dimensions of the global pose feature and n2 feature elements b1 dimensions of the local pose feature, then horizontally concatenating the global feature and the local feature to obtain an (a1+b1) dimensional feature matrix, thereby completing the feature extraction of the multi-source cross-scale point cloud data.
[0031] Compared with the prior art, the present application has the following advantages:
[0032] 1) The global reference point is introduced and the global pose is calculated, the effective dimension of the point cloud feature is increased, the accuracy of the point feature is improved, and the accuracy of the multi-source point cloud data matching and fusion is improved.
[0033] 2) Under the constraint of the global pose, the local uncertainty caused by calculating the point feature only through the local pose of the neighborhood point is reduced, and the error caused by the approximation of the neighborhood point feature in the neighborhood range during the matching and connection of the multi-source data is greatly reduced. BRIEF DESCRIPTION OF DRAWINGS
[0034] Figure 1 is a flowchart of the embodiment of the present application.
[0035] Figure 2 is a schematic diagram of the global pose calculation method.
[0036] Figure 3 is a schematic diagram of the local coordinate system construction method.
[0037] Figure 4 A local pose calculation method.
[0038] Figure 5 Feature matrix visualization results for an arbitrary point. DETAILED DESCRIPTION
[0039] The application provides a point cloud feature extraction method suitable for multi-source cross-scale data accurate registration.
[0040] As Figure 1 shown, the flow of the embodiment of the application includes the following steps:
[0041] Step 1, taking the global pose description reference point as the benchmark, the distance feature and the angle feature are calculated for all points of the point cloud, and the global pose feature is obtained.
[0042] Different ways are adopted for the calculation of the global pose description reference point for different data:
[0043] The global pose description reference point of the undirected single data is the point cloud centroid, and the calculation formula is:
[0044] C = mean (X, Y, Z) (1)
[0045] In the formula, C is the global pose description reference point, (X, Y, Z) is the point cloud coordinate, and mean () represents the mean operation.
[0046] The global pose description reference point of the undirected multi-data source is the coordinate mean of the centroid of each single data point cloud, and the calculation formula is:
[0047]
[0048] In the formula, C m is the global pose description reference point, C ix is the x coordinate of the centroid of the i th point cloud data, C iy is the y coordinate of the centroid of the i th point cloud data, and C iz is the z coordinate of the centroid of the i th point cloud data, and mean () represents the mean operation.
[0049] The global pose description reference point of the directed data is the coordinate mean of multiple image control points, and the calculation method is the same as formula (2).
[0050] The global pose feature is a three-tuple {D0, A0, A1} composed of 1 distance feature and 2 angle features. Figure 2 As G shown, P Q is the global pose description reference point; P iis the feature point acquired by the current query point within the query radius. P i is the distance feature D 0i of the line segment P i is the length of the line segment P G , the angle feature A 0i is the angle between the normal vector N i of the line segment P i and the line segment P i , the angle feature A G is the angle between the normal vector N 1i of the line segment P Q and the line segment P G , and the angle feature A i is the angle between the normal vector N G of the line segment P Q and the line segment P G . The feature triplets of each neighborhood point are calculated, and the global pose feature matrix is obtained. The expression form of the feature matrix is:
[0051]
[0052] Step 2: Batch construction of local coordinate system with all points of the point cloud as the origin.
[0053] The purpose of constructing the local coordinate system is to realize the local feature calculation of the points. As shown in Figure 3 , the current point is taken as the query point, the normal vector of the query point is taken as the U-axis, the normal vector of the plane formed by the global reference point and the query normal vector is taken as the V-axis, and the direction perpendicular to the U-axis and the V-axis is taken as the W-axis.
[0054] The global pose description reference point P G is defined as the direction vector N Q pointing to the query point P j , and the normal vector N Q of P q , and the calculation method of the local coordinate system is as follows:
[0055] U=norm(N q ) (4)
[0056] V=U×norm(N j ) (5)
[0057] W=U×V (6)
[0058] Step 3: Construction of feature point set in the local coordinate system, and calculation of local features for all points in the feature point set.
[0059] The local pose feature is a five-tuple {L, B, C, D, E} composed of one distance feature and four angle features. As shown in Figure 4 , P Q is the query point, i.e., the point for which the feature needs to be calculated; U, V, and W are based on P QThe local coordinate system is constructed, N i is the normal vector of P i . The distance feature L of P i is the length of the line segment P i P Q . The angle feature B is the angle between the normal vector N i of P i and the local coordinate axis U. The angle feature C is the angle between the normal vector N i of P i and the local coordinate axis V. The angle feature D is the angle between the normal vector N i of P i and the local coordinate axis W. The angle feature E is the angle between the normal vector N i of P i and the line segment P i P Q . The global pose feature matrix is obtained by calculating the five-tuple of features of each neighborhood point of P Q . The expression form of the feature matrix is:
[0060]
[0061] Step 4: Feature statistical calculation is performed on all point clouds to obtain a feature histogram and construct a feature matrix.
[0062] Feature statistical calculation is performed on all point clouds to obtain a feature histogram. The bin value is set to 11 when constructing the histogram. The global pose feature obtained in this way is 3 feature elements of 33 dimensions, and the local pose feature is 5 feature elements of 55 dimensions. Then, the global feature and the local feature are horizontally concatenated to obtain an 88-dimensional feature matrix, thereby completing the feature extraction of multi-source cross-scale point cloud data. Usually, the bin value is set to 11 when calculating point cloud features, but the feature dimension can also be changed by adjusting the histogram bin value without changing the number of features.
[0063] Figure 5 The matrix feature visualization result of any point in the photogrammetric matching point cloud when the bin value is set to 11. The feature similarity of two points can be judged by comparing the histogram similarity of the two points. By matching the histogram to find points with high feature similarity, the matching connection points of two point cloud data can be detected, and the splicing and registration of point cloud data can be carried out.
[0064] The specific embodiments described in this paper are only illustrative of the spirit of the application. Those skilled in the art to which the present application belongs can make various modifications or supplements to the described specific embodiments or replace them with similar ways, but will not deviate from the spirit of the application or exceed the scope defined by the appended claims.
Claims
1. A point cloud feature extraction method suitable for multi-source cross-scale data accurate registration, characterized in that, Comprising the following steps: Step 1, taking the global pose description reference point of the point cloud as the benchmark, calculate the distance feature and angle feature for all points of the point cloud, and obtain the global pose feature; Different ways are adopted for the calculation of the global pose description reference point for different data; The global pose description reference point of the undirected single data is the point cloud centroid, and the calculation formula is: （1） where C is a global pose description reference point, is a point cloud coordinate, denotes a mean operation; The global pose description reference point of the undirected multi-data source is the coordinate mean of the point cloud centroid of each single data, and the calculation formula is: （2） wherein C m is a global pose description reference point, is the x-coordinate of the centroid of the point cloud data, i is the y-coordinate of the centroid of the point cloud data, is the z-coordinate of the centroid of the point cloud data, i is the x-coordinate of the centroid of the point cloud data, is the y-coordinate of the centroid of the point cloud data, i is the z-coordinate of the centroid of the point cloud data, denotes a mean operation; The global pose description reference point of the oriented data is the coordinate mean of multiple image control points, and the calculation method is the same as formula (2); Step 2, batch construct local coordinate system with all points of the point cloud as the origin; Taking the current point as the query point, taking the normal vector of the query point as the U-axis, taking the normal vector of the plane formed by the global reference point and the query normal vector as the V-axis, and taking the direction perpendicular to the U-axis and V-axis as the W-axis, then the calculation method of the local coordinate system is as follows: （4） （5） （6） wherein is a global pose description reference point points to the query point is a direction vector, is a normal vector to the plane Step 3, construct a feature point set in the local coordinate system, and calculate the local feature for all points in the feature point set; Step 4, perform feature statistical calculation on all point clouds to obtain a feature histogram and construct a feature matrix.
2. The point cloud feature extraction method for multi-source cross-scale data precise registration according to claim 1, characterized in that: The global pose feature in step 1 is a triplet composed of 1 distance feature and 2 angle features , for the query point The global pose feature matrix is obtained by calculating the feature triplet of each neighborhood point, and the expression form of the feature matrix is: （3） wherein denotes the distance feature, the length of the line segment ; denotes the angle feature, the angle between the normal vector of the line segment and the line segment ; denotes the angle feature, the angle between the line segment and the line segment ; is the global pose description reference point; is the query point, i.e. the point for which the feature is currently to be calculated; is the feature point acquired by the current query point within the query radius.
3. The point cloud feature extraction method for multi-source cross-scale data precise registration of claim 1, wherein: The local pose feature in step 3 is a five-tuple composed of one distance feature and four angle features To The global pose feature matrix is obtained by calculating the feature five-tuple of each neighborhood point, and the expression form of the feature matrix is: （7） wherein denotes a distance feature, being the length of the line segment ; denotes an angle feature, being the angle between the normal vector of the plane and the local coordinate axis U; denotes an angle feature, being the angle between the normal vector of the plane and the local coordinate axis V; denotes an angle feature, being the angle between the normal vector of the plane and the local coordinate axis W; denotes an angle feature, being the angle between the normal vector of the plane and the line segment ; is the query point, i.e. the point for which the feature is currently computed; U, V, W are based on a local coordinate system constructed; is a feature point obtained within a query radius of the current query point; is a normal vector of the 4. The point cloud feature extraction method for multi-source cross-scale data precise registration of claim 3, wherein: In step 4, feature statistics calculation is performed on all point clouds to obtain a feature histogram, and the bin value is set to N when constructing the histogram, and the global pose feature obtained therefrom is feature elements dimensional, and the local pose feature is feature elements dimensional, and then the global feature and the local feature are horizontally concatenated to obtain a dimensional feature matrix, thereby completing the feature extraction of multi-source cross-scale point cloud data.