A vector surface feature matching method based on entity alignment

Abstract

The application relates to a vector surface element matching method based on entity alignment, which comprises the following steps: constructing a training sample set containing a plurality of surface element candidate matching pairs (a i ,b j ), and calculating the similarity indexes of the respective candidate matching pairs (a i ,b j ); constructing a matching relation recognition model by using an OCSVM classifier, inputting the training sample set into the matching relation recognition model for learning; obtaining a surface element pair matching result with a matching relation of 1:1; using a road network, a Vornoi graph and the 1:1 matching result as a segmentation boundary to align data with overall deviation in a closed domain; obtaining candidate matching pairs of to-be-matched surface elements with various matching relations, and recognizing the candidate matching pairs of the respective to-be-matched surface elements by using the matching relation recognition model; the method can learn the matching features of heterogeneous surface elements under the condition of a small amount of positive example samples, effectively recognize the matching relation of the same-named surface elements with position deviation, and provides technical support for digital topographic map quality inspection, incremental updating and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vector data matching production, and particularly relates to a vector surface feature matching method based on entity alignment. BACKGROUND

[0002] Topographic feature data is a collection of data describing the shape and geographical features of the earth's surface, and has wide research in the fields of infrastructure planning, engineering design, water resource management, natural disaster management and risk assessment. The data is mainly derived from ground surveying, digitizing of existing topographic maps, laser radar and aerial photography, including different types of features expressed in point-line-surface form such as landform feature points, roads, rivers and buildings. Surface features are one of the basic material features of a city, and account for a large proportion in spatial databases. For example, building base, water surface and slope surface all need to express their spatial position and range characteristics through surface features. However, due to different production methods and collection principles, different source data sets have different focuses in terms of data accuracy, richness and present situation. Surface feature matching can establish the connection between the same entities in multi-source data, and is a key technology for topographic feature data integration and updating, and promotes the construction of urban informationization.

[0003] After nearly three decades of development, research on vector surface feature matching has made significant progress and fruitful research results. It can be mainly divided into two categories: index weighting and learning-based matching method. The index weighting method uses the characteristics of vector surface features, such as geometric descriptors, semantic features and spatial relationships, to individually or combinedly weight the similarity of surface features. The learning-based matching method regards spatial entity matching as a kind of supervised classification problem, so as to weaken the subjective influence of expert weights, and to evaluate the matching of surface features. At the same time, in order to further improve the matching effect, some hierarchical matching rules, clustering combination, probability relaxation and other methods are also used to optimize the feature description rules to better realize the recognition of many-to-many relationship.

[0004] Although the above matching strategies have achieved good matching results, the weight and threshold setting of the index usually needs to rely on the knowledge and experience of experts, and cannot effectively adapt to different data situations, which may lead to unreliable decisions in the results. In addition, although the current artificial intelligence methods can objectively obtain improved results, they need enough positive and negative samples for labeling and training. However, the matching and non-matching relationship is diverse, and when manually labeling samples, the spatial proximity relationship between a building and the surrounding buildings needs to be observed as the basis for judgment, and the standard deviation of the feature distribution of the negative samples is much larger than that of the positive samples. It is a time-consuming work to reasonably select samples for the commonly used binary classification machine learning methods. SUMMARY

[0005] The present application aims at the technical problems existing in the prior art, and provides a vector surface feature matching method based on entity alignment, which can learn the matching features of heterogeneous surface features under the condition of a small number of positive samples, effectively identify the matching relationship of the same-named surface features with position deviation, and provide technical support for digital topographic map quality inspection and incremental updating.

[0006] According to a first aspect of the present application, a vector surface feature matching method based on entity alignment is provided, comprising the following steps: step 1, constructing a training sample set containing a plurality of candidate matching pairs (a i , b j ) of the surface features, and calculating the similarity indexes of each candidate matching pair (a i , b j ); i and j represent the serial numbers;

[0007] Step 2, constructing a matching relationship identification model by using an OCSVM classifier, inputting the training sample set into the matching relationship identification model for learning, inputting the surface features to be matched into the trained matching relationship identification model, and screening the surface feature pair matching result with a 1:1 matching relationship from the output result of the matching relationship identification model;

[0008] Step 3, using the road network, the Voronoi diagram and the 1:1 matching result as the segmentation boundary together, segmenting the experimental area into a plurality of closed domains, obtaining the control polygons in the closed domains according to the importance of the polygons, aligning the data with overall deviation in the closed domains according to the control polygons, and aligning the data with overall deviation in the closed domains according to the offset direction and distance of the control polygons;

[0009] Step 4, obtaining the candidate matching pairs of the surface features to be matched in various matching relationships, and identifying the candidate matching pairs of each surface feature to be matched by using the matching relationship identification model.

[0010] On the basis of the above technical solution, the present application can also be improved as follows.

[0011] Optionally, the step 1 comprises:

[0012] After the reference of the data in the heterogeneous surface feature set A and the heterogeneous surface feature set B is unified, the topological intersection and topological proximity analysis are performed on any surface feature a i in the heterogeneous surface feature set A and any surface feature b j in the heterogeneous surface feature set B, so as to obtain each surface feature candidate matching pair (a i , b j );

[0013] A plurality of surface feature candidate matching pairs (a i , bj ) constructing a training sample set;

[0014] the step 2 inputs the training sample set into the matching relationship identification model to learn the similarity features of each of the face element candidate matching pairs (a i , b j ) in the training sample set.

[0015] Optionally, the process of performing the topological intersection and topological proximity analysis in the step 1 comprises:

[0016] superimposing the heterogeneous face element data to obtain the topological intersection relationship between the target face element and the reference face element; using the Delaunay triangulation of the face element centroid to obtain the topological proximity relationship of the face element, and obtaining the face element candidate matching pairs with a matching relationship of 1:1.

[0017] Optionally, the similarity index comprises at least one of position similarity, direction similarity, area similarity, shape similarity, context similarity, and topological similarity.

[0018] The topological similarity index calculation formula is:

[0019]

[0020] wherein, TP I is a topological strength used to describe the relationship between the face element to be matched and its topologically adjacent face element, nI represents the number of adjacent face elements, x h represents a label value indicating whether to match, if the adjacent match, x h = 1, if the adjacent does not match, x h = -1.

[0021] TP S is a topological depth, nS represents the number of adjacent polygons, l i-part represents the closed domain edge length of the adjacent element in the Voronoi diagram, l all represents the total closed domain edge length of the face element to be matched in the Voronoi diagram, decision-v i represents the decision function value of the i-th neighbor matching pair.

[0022] TP represents the matching support degree of the adjacent face element to the face element to be matched.

[0023] sim tp (a, b) is the topological similarity of the face elements a and b.

[0024] Optionally, the process of training the matching relationship identification model in the step 2 comprises:

[0025] Step 201: Obtain the candidate matching relationship for each building, then identify the buildings with matching relationships and mark their similarity feature vectors as positive samples, while marking the candidate matching pairs without matching relationships as negative samples;

[0026] Step 202: Divide the training sample set into a strict training sample set and a general training sample set according to the area overlap. Use the positional similarity, orientational similarity, area similarity and shape similarity of the strict sample set to train a strict recognition model. Use the positional similarity, orientational similarity, area similarity, shape similarity, topological similarity and contextual similarity of the general training sample set to obtain a general recognition model.

[0027] Step 203: Select the Gaussian kernel function to construct the OCSVM classification model as the matching relationship recognition model.

[0028] Optionally, the process of predicting matching relationships based on the trained matching relationship recognition model in step 2 includes: strict matching relationship prediction and general matching relationship prediction.

[0029] The strict matching relationship prediction includes: selecting samples to be matched from reference data A and target data B to obtain candidate matching pairs; predicting the position, orientation, area and shape similarity of the candidate matching pairs using a trained strict classification model, and outputting prediction results 1 and -1 to represent matching and non-matching, respectively.

[0030] The general matching relationship prediction includes: traversing the unmatched feature sets A1 and B1 in the initial matching results, assuming the set of buildings to be matched is a = {a1, a2, ..., a...} n b = {b1, b2, ..., b} n}, exhaustively search for b and a certain element a to be matched. i The candidate matching situation is analyzed, the similarity feature value of the elements to be matched is calculated, the final prediction result is obtained by using a general matching relationship recognition model, the matching relationship is determined, and the matched polygon elements are updated to the original data to provide context information for the remaining matching; this process is iterated until all elements are traversed to obtain the matching result of the polygon element pair with a matching relationship of 1:1.

[0031] Optionally, step 3 includes:

[0032] Step 301: Obtain the surface feature matching results for each road network closed domain, and use a hierarchical clustering algorithm to analyze the matched dataset A. m and B m Perform cluster analysis and obtain cluster result A. mc ={a c1 a c2 , ..., a cj} and Bmc = {b1, b2, …, b c1}; c2 , …, b ci};

[0033] where matched data sets A m = {a1, a2, …, a j} and B m = {b1, b2, … b i}, and unmatched data sets A n = {a1, a2, …, a j} and B n = {b1, b2, … b i};

[0034] Step 302, using the respective cluster centers and directions of A mc and B mc , a boundary is generated in combination with a Voronoi diagram;

[0035] Step 303, the inner boundary is combined with a road network to segment the unmatched surface elements into multiple groups, and a control polygon of the group is obtained by using shape similarity, shape complexity and matching degree indexes; and data with overall deviation in the closed domain is aligned according to the offset direction and distance of the control polygon.

[0036] Optionally, in step 3,

[0037] The calculation formula of shape similarity is:

[0038]

[0039] where e (·) (·) is an accumulated value of the corner of the surface element, max is a maximum value function, and the measurement value range is [0, 1], and the greater the value is, the more similar it is;

[0040] The calculation formula of shape complexity is:

[0041]

[0042] where the shape complexity is calculated by the ratio of the perimeter (P EAR ) of the equal-area rectangle (EAR) of the original polygon to the perimeter (P R ) of the original polygon. Where P MABR is the perimeter of the minimum area circumscribed rectangle, A R is the area of the original polygon, and A MABR is the area of the minimum area circumscribed rectangle;

[0043] The calculation formula of the matching degree index is:

[0044]

[0045] wherein f(x h，k ) is a normalized value d f of the decision function value of the surface element h and the surface element k, d f ∈[-1, 1], which is the matching degree of the two polygons.

[0046] The calculation formula of the important polygon is as follows:

[0047]

[0048] wherein S represents the shape similarity, C represents the average shape complexity, M is the decision function matrix, and l mn represents the polygon importance.

[0049] Optionally, the process of generating the inner boundary in step 3 comprises:

[0050] Step 301, skeletonizing the surface element blank area to obtain a Voronoi-like diagram;

[0051] Step 302, performing hierarchical clustering on the matched surface elements in each region of the road network to obtain a clustering center and a direction;

[0052] Step 303, taking the clustering center as the starting point, fusing the Voronoi-like diagram in which the matched surface elements are located, and performing boundary merging along the Voronoi-like diagram according to the clustering direction;

[0053] Step 304, when the Voronoi-like diagram in which the matched surface elements are located cannot be fused, determining whether it is adjacent to other classes, if yes, continuing to fuse, if no, fusing along the direction until it is closed with the outer boundary or an inner ring is generated.

[0054] Optionally, step 4 comprises:

[0055] Step 401, obtaining all candidate matching sets of the un-matched surface elements after alignment, and performing exhaustive combination on the un-matched surface elements under the constraint of the adjacent relationship;

[0056] Step 402, using the Concave Hull merging method to aggregate the building groups in the non-empty matching and the non-1:1 matching relationship into a single element, and converting into a 1:1 matching relationship;

[0057] Step 403, repeating the prediction matching relationship processing method of step 2 on the merged data, and using the strict recognition model and the general recognition model to recognize the 1:N, M:1 and M:N matching relationships of the un-matched surface elements;

[0058] Step 404, the matching relationship of 1:1 obtained in step 2 and the matching relationship of 1:N, M:1 and M:N of the surface elements obtained in step 403 are counted, and the surface elements that have not been matched through step 2 and step 403 are identified as 1:0 or 0:1 matching relationship.

[0059] The vector surface element matching method based on entity alignment provided by the application has the beneficial effects including: the application introduces single-class classifier and designs hierarchical matching strategy of simple and complex surface elements, strict and general classification models. First, the classification model is trained by using a small amount of positive samples, and then hierarchical matching is performed from the matching type and the angle of building data, which takes into account the bias of positive and negative sample feature expression based on learning and the generalization problem of feature learning; (2) the application introduces surface entity alignment strategy to solve the problem that the heterogenous surface elements with position deviation are easily wrong and missed in the general method matching, not only considers simple feature comparison, but also proposes comparison of structure in closed domain, reduces the position deviation by controlling the polygon alignment surface element group, and solves the problem that the traditional geometric feature description and local neighborhood information support cannot comprehensively describe the surface elements with large feature deviation. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The flowchart of the vector surface element matching method based on entity alignment provided by the application;

[0061] Figure 2 The schematic diagram of the heterogenous surface element data candidate matching pair acquisition example provided by the embodiment of the application;

[0062] Figure 3 The schematic diagram of the topological similarity calculation of the vector surface element provided by the embodiment of the application;

[0063] Figure 4 The schematic diagram of the different matching relationship of the intention recognition provided by the embodiment of the application;

[0064] Figure 5 The schematic diagram of the constraint boundary acquisition in the entity alignment strategy provided by the embodiment of the application;

[0065] Figure 6 The schematic diagram of the implementation of the entity alignment step provided by the embodiment of the application;

[0066] Figure 7 The schematic diagram of the M:N candidate matching pair acquisition and surface element merging provided by the embodiment of the application. DETAILED DESCRIPTION

[0067] The principles and characteristics of the application are described below in combination with the drawings, and the examples are only used to explain the application and are not used to limit the scope of the application.

[0068] Figure 1 A flow chart of a vector surface feature matching method based on entity alignment provided by the present application is shown in Figure 1 The matching method comprises the following steps:

[0069] Step 1, a training sample set containing a plurality of surface feature candidate matching pairs (a i , b j ) is constructed, and the similarity indexes of each candidate matching pair (a i , b j ) are calculated; i and j represent the serial numbers.

[0070] Step 2, a matching relationship identification model is constructed by using an OCSVM classifier, the training sample set is input into the matching relationship identification model for learning, and the surface feature pairs with a matching relationship of 1:1 are screened from the matching relationship identification model to obtain a matching result.

[0071] Step 3, the experimental area is segmented into a plurality of closed domains by using a road network, a Vornoi diagram and the 1:1 matching result as a segmentation boundary, the control polygons in the closed domains are obtained according to the importance of the polygons, the data with overall deviation in the closed domains are aligned according to the control polygons, and the data with overall deviation in the closed domains are aligned according to the offset direction and distance of the control polygons.

[0072] Step 4, the candidate matching pairs of the surface features to be matched in various matching relationships are obtained, and the matching relationship identification model is used to identify the candidate matching pairs of the surface features to be matched.

[0073] The vector surface feature matching method based on entity alignment provided by the present application can learn the matching features of heterogeneous surface features under the condition of a small number of positive samples, effectively identify the matching relationship of the same-named surface features with position deviation, and provide technical support for digital topographic map quality inspection and incremental updating.

[0074] Embodiment 1

[0075] The embodiment 1 provided by the present application is an embodiment of the vector surface feature matching method based on entity alignment provided by the present application, and in combination with Figure 1 It can be known that the embodiment of the matching method comprises the following steps:

[0076] Step 1, a training sample set containing a plurality of surface feature candidate matching pairs (a i , b j ) is constructed, and the similarity indexes of each candidate matching pair (a i , b j ) are calculated; i and j represent the serial numbers.

[0077] In one possible embodiment, step 1 includes:

[0078] After unifying the baseline of the data in heterogeneous surface feature set A and heterogeneous surface feature set B, for any surface feature a in heterogeneous surface feature set A i and any surface element b in the heterogeneous surface element set B j Perform topological intersection and topological proximity analysis to obtain candidate matching pairs for each face feature (a i b j ).

[0079] In specific implementation, two heterogeneous surface feature datasets A = {a1, a2, ... a...} are unified. m} and dataset B = {b1, b2, ..., b} n The baseline for this includes the unification of data format, coordinate system, and feature type. The data format is converted to the ArcGIS-supported "*.shp" and the easily transferable and program-recognizable "GeoJSON" format, and then each feature is defined as a... i ∈A and b j ∈B is the basic unit for performing topological intersection and topological proximity analysis on two sets of surface feature sets.

[0080] Select multiple candidate matching pairs of face features (a i b j The training sample set is constructed.

[0081] In one possible embodiment, the process of performing topological intersection and topological proximity analysis in step 1 includes:

[0082] By overlaying heterogeneous surface feature data, the topological intersection relationship between the target surface feature and the reference surface feature is obtained; the topological proximity relationship of the surface feature is obtained using the Delaunay triangulation of the surface feature mass center, and candidate matching pairs of surface features with a matching relationship of 1:1 are obtained.

[0083] Specifically, the steps for candidate surface feature matching are as follows: First, the heterogeneous surface feature data are overlaid to obtain b from the target dataset A and the reference dataset B. j The set of intersecting surface features A1∈A is b j The candidate set to be matched is then used, and the topological neighbor set T1 of set A1 is obtained as a supplementary candidate set using the Delaunay triangulation of the centroid. Then the face feature b j The candidate matching set is {A1, T1}. Similarly, the polygon feature a in the target dataset... i The candidate matching set is also obtained using this method.

[0084] like Figure 2Fig. 1 is a schematic diagram of an example of obtaining a candidate matching pair of heterogeneous surface element data according to an embodiment of the present application, referring to Figure 2 In the specific embodiment, a detailed example of obtaining a candidate matching pair is shown, where data set A = {a1, a2, a3, a4, a5} and data set B = {b1, b2, b3, b4, b5, b6, b7}. Figure 2 (a) is a schematic diagram of obtaining an example of a neighboring candidate matching pair, where the dashed line is a Delaunay triangulation of the centroid of data set B, and it can be seen that the neighboring surface elements of b4 are {b2, b3, b5, b6, b7}, and the intersection relationship of Figure 2 (b) can obtain the topological intersection candidate set of b4 as {a4, a5}; in the same way, the neighboring surface elements of a5 are {a1, a4}, and the topological intersection candidate set of a5 is {b4}; when b4 is matched with a5, the candidate matching set corresponding to b4 is {a1, a4, a5}, and the candidate matching set corresponding to a5 is {b2, b3, b4, b5, b6, b7}.

[0085] In a possible embodiment, the similarity index includes at least one of position similarity, direction similarity, area similarity, shape similarity, context similarity, and topological similarity.

[0086] Let the polygons to be matched in two heterogeneous surface element data sets A and B be a and b, and the calculation formulas of the six similarity features are as follows:

[0087] (1) The calculation formula of the position similarity index is as follows:

[0088]

[0089] Where (x1, y1) and (x2, y2) are the centroid coordinates of the two surface elements a and b to be matched, and U is the maximum distance of the boundary point set of the polygons to be matched a and b; the measurement value domain is [0, 1], and the greater the value, the more similar.

[0090] (2) The calculation formula of the direction similarity index is as follows:

[0091] sim dir (a, b) = |cos(|θ b -θ g |)|

[0092] Where θ (·) is the direction of the long side of the minimum circumscribed rectangle of the surface element, and the measurement value domain is [0, 1], and the greater the value, the more similar.

[0093] (3) The calculation formula of the area similarity index is as follows:

[0094]

[0095] where, Area(·) is the area of the surface element, a∩b is the intersection area of surface elements a and b, min is the minimum function; the measurement value domain is [0, 1], and the greater the value, the more similar.

[0096] (4) The shape similarity index calculation formula is as follows:

[0097]

[0098] where, e (·) (·) is the cumulative value of the corner of the surface element, max is the maximum function, the measurement value domain is [0, 1], and the greater the value, the more similar.

[0099] (5) The context similarity index calculation formula is as follows:

[0100]

[0101] where, (a i , b j ) represents the surface element pair to be matched, (a h , b k ) represents the neighborhood matched surface element pair. d(b i , b h ) and d(g j , g k ) are the Euclidean distances between surface elements; N i and N j are the spatial neighbor target sets of a i and b j , respectively; n is the number of neighbors; represents the angle between the centroid connecting line of a i , a h and the x positive axis, represents the angle between the centroid connecting line of b j , b k and the x positive axis; r dis and r dir are the relative distance and relative direction of the to-be-matched pair (a i , a h ) and the neighborhood matched pair (b j , b k ), respectively.

[0102] Figure 3is a schematic diagram of a calculation example of topological similarity provided by the embodiment of the present application, the adjacent relationship of surface elements is obtained through the topological connection relationship of the Voronoi diagram, the matching pairs of the data set A and the data set B in the figure are {a1;b1}, {a2;b2}, {a3;b3}, {a4;b4}, {a5;b5}, {a6;b6} and {a7;b7}; taking the topological similarity calculation of the matching pair {a3;b3} as an example, the topological neighbors of a3 are {a1, a4, a7, a8}, wherein the matched surface elements are {a1, a4, a7}, the topological neighbors of b3 are {b1, b4, b7}, wherein the matched surface elements are {b1, b4, b7}, and the matched surface elements in the adjacent relationship are {a1;b1}, {a4;b4} and {a7;b7}, the (decision function value, a i The length ratio of the adjacent edge, b j The length ratio of the adjacent edge) are (0.78, 0.18, 0.24), (0.97, 0.44, 0.26) and (0.72, 0.38, 0.50), then the TP I of a3 is 0.75, the TP S of b3 is 0.84, the TP I of b3 is 1, the TP S of b3 is 0.80, and the topological similarity sim tp (a, b) of the matching pair {a3;b3} is 0.95.

[0103] Step 2, a matching relationship identification model is constructed by using the OCSVM classifier, the training sample set is input into the matching relationship identification model for learning, and the surface element pair matching result of the matching relationship 1:1 is obtained by screening from the output result of the matching relationship identification model.

[0104] Figure 4 It is shown that the matching relationship of the intention recognition provided by the embodiment of the present application includes 1:1 and M:N, that is, 1:1, 1:N (N>1), M:1 (M>1) and M:N (N>1 & M>1) matching relationship, such as {a1;b1}, {a2;b2}, {a3;b3} and the like are 1:1 matching pairs, {a8;b9, b 10} is a 1:N matching pair, {a9, a 10 ;b 11} is a M:1 matching pair, {a6, a7;b6, b7, b8} is a M:N matching pair, a 11 and b 12 are determined as empty matching because there is no corresponding matching surface element.

[0105] In a possible implementation manner, in step 2, the training sample set is input into the matching relationship identification model to learn the similarity features of each candidate matching pair of surface elements in the training sample set i , b j .

[0106] Step 2 includes two parts: training the matching relationship identification model and predicting the matching relationship based on the trained matching relationship identification model. Specifically, in a possible implementation manner, the process of training the matching relationship identification model in step 2 includes:

[0107] Step 201, marking sample data: obtaining candidate matching relationships of each building, then identifying buildings with matching relationships, and marking the similarity feature vectors thereof as positive samples (y = 1), while marking the candidate matching pairs without matching relationships as negative samples (y = -1). Wherein, the training data has only a small number of positive samples, and the verification data includes a certain number of positive and negative samples.

[0108] Step 202, screening training sample data: dividing the training sample set into a strict training sample set (area overlap degree > 0.5) and a general training sample set (without overlap degree restriction) according to the area overlap degree, training the strict identification model by using the position similarity, direction similarity, area similarity and shape similarity of the strict sample set, and obtaining the general identification model by using the position similarity, direction similarity, area similarity, shape similarity, topology similarity and context similarity of the general training sample set, wherein the topology similarity and context similarity of the general training sample set are calculated based on the identification results of the strict identification model as adjacent information.

[0109] In a specific implementation, the area overlap degree calculation formula is as follows:

[0110]

[0111] Wherein, Area(·) is the area of the surface element, a∩b is the intersection area of the surface elements a and b, min is the minimum value function; the measurement value domain is [0, 1], and the greater the value, the more similar.

[0112] Step 203, training the identification model: after creating samples and extracting similarity features, a Gaussian kernel function is selected to construct an OCSVM classification model as the matching relationship identification model.

[0113] In a specific implementation, the process of constructing the OCSVM classification model includes: using the maximum test set F1 score to obtain the Gaussian kernel bandwidth g∈(0, 1) and the relaxation factor v∈(0, 1], then constructing and solving the optimization problem to obtain the solution α i and a positive component α j , and calculating ρ; calculating the decision function f(x).

[0114] The optimization problem calculation formula is as follows:

[0115]

[0116] Where, x i For training samples, ρ is the offset, and φ(·) is the value of x. i The kernel function maps to the high-dimensional feature space, where ω is the weight vector, i.e., the normal vector of the separating hyperplane. This case uses the Gaussian kernel function K(x). i x j )=exp(-||x i -x j || 2 / σ 2 ), where σ is the Gaussian kernel bandwidth. v∈(0,1] is the regularization coefficient, which controls the trade-off between the hyperplane and outliers. ξ i It is the slack variable corresponding to the i-th training sample.

[0117] From the expression ω of the hyperplane T φ(x i From the formula for calculating the decision function f(x) and the formula for calculating the optimization problem, we can obtain the following formula:

[0118] f(x)=sgn[ω T φ(x i )-ρ]

[0119] In one possible embodiment, step 2, which involves predicting the matching relationship based on the trained matching relationship recognition model, includes: strict matching relationship prediction and general matching relationship prediction.

[0120] Strict matching prediction includes: using the predicted matching results from the strict matching model as a high-precision 1:1 matching relationship, also known as the initial matching result C1. Samples to be matched are selected from reference data A and target data B to obtain candidate matching pairs; the position, orientation, area, and shape similarity of the candidate matching pairs are predicted using a trained strict classification model, with outputs representing the prediction results 1 and -1 for matching and non-matching, respectively; the output may contain multiple matching relationships, indicating that multiple candidate matching pairs of the same entity are all identified as matches.

[0121] General matching relationship prediction includes: predicting general matching relationships based on the initial matching result C1, optimizing the initial matching result, and achieving a 1:1 match with low overlap. It iterates through the unmatched feature sets A1 and B1 in the initial matching results, assuming the set of buildings to be matched is a = {a1, a2, ..., a...}. n b = {b1, b2, ..., b} n}, exhaustively b and a certain to be matched elements a i , calculate the similarity of the elements to be matched characteristic value, get the final prediction results using the general matching relationship identification model, determine the matching relationship, and update the matched surface elements to the original data, provide context information for the remaining matching; iteration in turn, until all elements are traversed, get the matching result of the surface element pair with a matching relationship of 1:1.

[0122] Step 3, using the road network, class Voronoi diagram and 1:1 matching result as the segmentation boundary, the experimental area is segmented into several closed domains, and then the control polygon in the closed domain is obtained according to the importance of the polygon, and the data with overall deviation in the closed domain is aligned according to the control polygon; according to the offset direction and distance of the control polygon, the data with overall deviation in the closed domain is aligned.

[0123] In a possible embodiment, step 3 includes:

[0124] Step 301, obtaining the surface element matching result of each road network closed domain, using hierarchical clustering algorithm (method='single', metric='euclidean') to cluster the matched data sets A m and B m , obtaining the clustering results A mc ={a c1 , a c2 , …, a cj} and B mc ={b c1 , b c2 , …b ci}.

[0125] Wherein, the matched data sets A m ={a1, a2, …, a j} and B m ={b1, b2, …b i}, the unmatched data sets are A n ={a1, a2, …, a j} and B n ={b1, b2, …b i}.

[0126] Step 302, using the respective cluster centers and directions of A mc and B mc , combined with the class Voronoi diagram to generate the boundary.

[0127] In step 303, the inner boundary combined with the road network divides the unmatched surface elements into multiple groups, and the important polygons of the groups, i.e., control polygons, are obtained by using shape similarity, shape complexity and matching degree indexes. According to the offset direction and distance of the control polygons, the data with overall deviation in the closed domain is aligned.

[0128] In a possible implementation manner, the calculation formula of the shape similarity is as follows:

[0129]

[0130] wherein e (·) is the cumulative value of the corner of the surface element, max is the maximum value function, and the measurement value range is [0, 1]. The greater the value is, the more similar it is.

[0131] The calculation formula of the shape complexity is as follows:

[0132]

[0133] wherein the shape complexity is calculated by the ratio of the perimeter (P EAR ) of the equal-area rectangle (EAR) of the original polygon to the perimeter (P R ) of the original polygon. P MABR is the perimeter of the minimum area circumscribed rectangle, A R is the area of the original polygon, and A MABR is the area of the minimum area circumscribed rectangle.

[0134] The calculation formula of the matching degree index is as follows:

[0135]

[0136] wherein f(x h，k ) is the normalized value d f of the decision function value of the surface element h and the surface element k, d f ∈[-1, 1], and d f is the matching degree of the two polygons.

[0137] According to the above, the calculation formula of the important polygon is as follows:

[0138]

[0139] wherein S represents the shape similarity, C represents the average shape complexity, M is the decision function matrix, and l mn represents the polygon importance.

[0140] In a possible implementation manner, the process of the inner boundary generation in step 3 includes:

[0141] Step 301, skeletonize the blank area of the surface element to obtain a Voronoi-like diagram.

[0142] Step 302, perform hierarchical clustering (method='single', metric='euclidean') on the matched surface elements in each region of the road network to obtain the cluster center and direction.

[0143] Step 303, take the cluster center as the starting point, and according to the cluster direction, merge the Voronoi-like diagram where the matched surface elements are located.

[0144] Step 304, when the Voronoi-like diagram where the matched surface elements are located cannot be merged, determine whether it is adjacent to other classes, if so, continue to merge, if not, merge along the direction until it is closed with the outer boundary (road network) or generates an inner ring.

[0145] The distribution of surface elements in the digital terrain map shows different degrees of linear distribution from the perspective of overall structure and local structure, which can well segment the unmatched surface elements. Figure 5 is an example of generating a segmentation boundary based on matched surface elements. First, cluster the matched surface elements in the region, that is, the two dark surface element groups in the figure, and obtain the closed boundary for segmenting the unmatched surface elements according to the inner boundary generation logic, such as the road network and the dark Voronoi diagram in the figure, which can divide the surface element data A and data B in the region into two parts; then align the to-be-matched surface elements according to the translation variable of the landmark surface elements, such as Figure 6 In the figure, according to the two landmarks of data A represented by dark color, the two groups of unmatched surface elements of data A can be aligned respectively, and the two landmarks of data B represented by dark color can also be aligned to align the two groups of unmatched surface elements of data B. The alignment result is shown in Figure 6 (right). It can be seen that the accuracy of the aligned data is obviously improved, which helps to improve the subsequent matching effect.

[0146] Step 4, obtain the candidate matching pairs of the to-be-matched surface elements of various matching relationships, and identify the candidate matching pairs of each to-be-matched surface element using the matching relationship identification model.

[0147] In one possible implementation manner, step 4 includes:

[0148] Step 401, obtain all candidate matching sets of the to-be-matched surface elements after alignment, and perform exhaustive combination on the to-be-matched surface elements under the constraint of the proximity relationship, to ensure that all matching cases are considered. Compared with simple exhaustive combination, the exhaustive combination under the constraint reduces a lot of unreasonable cases and reduces the complexity of the algorithm.

[0149] Figure 7 Exhaustive combination of all candidate matching sets is shown, for example, matching pairs of (a3, a5, a 7-11 , a 14-21 ; b3, b5, b 7-12 ) may have (a9; b 11 , b 12 ), (a9, a 10 ; b 11 , b 12 ), (a9, a 10 , a 14 ; b 11 , b 12 ), etc., and the possibility of (a9, a 11 ; b 11 , b 12 ) and other violations of proximity relationships is ruled out.

[0150] Step 402, using the Concave Hull merging method, the non-empty matching and non-1:1 matching relationship building group is aggregated into a single element, converted into 1:1 matching relationship, and the calculation of feature indicators is facilitated.

[0151] Figure 7 (d) The example of two groups of M:N matching pairs after merging into 1:1 matching is shown. Compared with simple convex hull merging, the boundary contour features are more similar to the original graph, the error in similarity feature index calculation is smaller, and it is beneficial to identify complex M:N matching relationships.

[0152] Step 403, repeat the prediction matching relationship processing method of step 2 for the merged data, and use the strict identification model and the general identification model to identify the 1:N, M:1 and M:N matching relationship of the face elements to be matched.

[0153] However, there may be some multiple matching cases based on the OCSVM classifier output, which are manifested as multiple candidate matching pairs of the same face element being identified as matching. For this case, the final matching relationship is obtained according to the decision distance value of OCSVM. Specifically, if there are matching pairs containing the same elements in the results predicted by the classifier, the matching pair with the highest matching degree is selected as the final matching result.

[0154] Step 404, count the face elements whose matching relationship is 1:1 obtained in step 2 and the matching relationship is 1:N, M:1 and M:N obtained in step 403, and identify the face elements that have not been matched after step 2 and step 403 as 1:0 or 0:1 matching relationship.

[0155] The remaining multiple matching relationships of step 5 are achieved by strict identification model and general identification model for primary identification and secondary identification respectively, and in each subsequent iteration, the surface elements of the matching relationship that have been determined are updated to the original data to provide context information for the remaining matching. The iteration is sequentially performed until all elements are traversed, and finally the matching result containing all matching relationships is obtained. Through the above steps, 1:1, 1:N, M:1 and M:N and empty matching association relationships of the surface elements in the heterogeneous topographic map data can be identified.

[0156] The vector surface element matching method based on entity alignment provided by the embodiment of the application introduces a single-class classifier and designs a hierarchical matching strategy of simple and complex surface elements and strict and general classification models. First, a small amount of positive samples is used to train the classification model, and then hierarchical matching is performed from the matching type and building data, which takes into account the bias of the positive and negative sample feature expression based on learning and the generalization problem of feature learning; (2) In view of the problem that the heterogeneous surface elements with position deviation are prone to errors and omissions in the general method matching, the surface entity alignment strategy is introduced, not only the simple feature comparison is considered, but also the comparison of the structure in the closed domain is proposed, the polygon alignment surface element group is controlled, the position deviation is reduced, and the problem that the traditional geometric feature description and local adjacent information support cannot comprehensively describe the surface elements with large feature deviation is alleviated.

[0157] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0158] Those skilled in the art should understand that the embodiments of the application can be provided as a method, a system, or a computer program product. Therefore, the application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0159] The application is described with reference to flowcharts and / or block diagrams according to the method, device (system), and computer program product of the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a machine that implements the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 an apparatus for performing the functions specified in the flowchart

[0160] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flowchart Figure 1 one or more processes and / or blocks Figure 1 an apparatus for performing the functions specified in the flowchart

[0161] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart Figure 1 one or more processes and / or blocks Figure 1 an apparatus for performing the functions specified in the flowchart

[0162] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such additional variations and modifications as fall within the scope of the application.

[0163] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. An entity alignment based vector map feature matching method, characterized in that, The matching method comprises: Step 1, constructing a training sample set containing multiple pairs of candidate matches of surface elements and calculating the similarity indexes of each pair of candidate matches ; i and j represent serial numbers; Step 2, a matching relationship identification model is constructed by using an OCSVM classifier, the training sample set is input into the matching relationship identification model for learning, and a to-be-matched surface element is input into the trained matching relationship identification model, and a surface element pair with a matching relationship of 1:1 is screened from an output result of the matching relationship identification model. Step 3, the experimental area is segmented into a plurality of closed domains by using a road network, a Vornoi diagram and the 1:1 matching result as a segmentation boundary, a control polygon in the closed domain is obtained according to the importance of the polygon, and data with overall deviation in the closed domain is aligned according to the control polygon; and data with overall deviation in the closed domain is aligned according to the offset direction and distance of the control polygon. Step 4, candidate matching pairs of to-be-matched surface elements of various matching relationships are obtained, and the matching relationship identification model is used to identify the candidate matching pairs of the to-be-matched surface elements. The step 3 comprises: Step 301, obtaining the face element matching result of each road network closed domain, using hierarchical clustering algorithm on the matched data set and carrying out clustering analysis to obtain the clustering result and ; wherein the matched data set and the unmatched data set is and ; Step 302, utilizing and the respective cluster centers and directions, the boundaries are generated in conjunction with a Voronoi diagram. Step 303, the unmatched surface elements are segmented into a plurality of groups by using an inner boundary and a road network, a control polygon of the group is obtained by using shape similarity, shape complexity and a matching degree index, and data with overall deviation in the closed domain is aligned according to the offset direction and distance of the control polygon. In the step 3, The calculation formula of the shape similarity is: wherein is the corner cumulative value of the planar element, max is the maximum function, the measurement value domain is [0, 1], and the greater the value, the more similar. The calculation formula of the shape complexity is: wherein the shape complexity is calculated as a ratio of the perimeter of the equal-area rectangle (EAR) of the original polygon to the perimeter of the original polygon itself wherein is the perimeter of the minimum-area circumscribed rectangle, is the area of the original polygon, is the area of the minimum-area circumscribed rectangle; The calculation formula of the matching degree index is: wherein is a normalized value of the decision function value of the face element h and the face element k , ,1], here as a matching degree of two polygons; The calculation formula of the important polygon is as follows: = where S represents shape similarity, C represents average shape complexity, and M is a decision function matrix, represents the polygon importance.

2. The matching method of claim 1, wherein, The step 1 comprises: heterogeneous surface element set and heterogeneous surface element set heterogeneous surface element set any surface element in the heterogeneous surface element set and heterogeneous surface element set any surface element in the heterogeneous surface element set topological intersection and topological adjacency analysis are performed on the heterogeneous surface element set ; selecting a plurality of the face element candidate matching pairs constructing a training sample set The step 2 learns each of the face element candidate matching pairs in the training sample set by inputting the training sample set into the matching relation identification model of similarity features.

3. The matching method of claim 2, wherein, The process of performing topological intersection and topological proximity analysis in the step 1 comprises: The target surface elements and the reference surface elements are superimposed to obtain a topological intersection relationship of the target surface elements and the reference surface elements; a topological proximity relationship of the surface elements is obtained by using a Delaunay triangulation of the surface element centroids, and the candidate matching pairs of the surface elements with a matching relationship of 1:1 are obtained.

4. The matching method of claim 2, wherein, The similarity index comprises at least one of position similarity, direction similarity, area similarity, shape similarity, context similarity and topological similarity. The calculation formula of the topological similarity index is: ； wherein, nI is the number of adjacent face elements, is a label value indicating whether or not the adjacent match, is a label value indicating whether or not the adjacent mismatch, ; nS denotes the number of neighboring polygons for the topological depth, denotes the closed domain edge length of the neighboring element in the Voronoi-like diagram, denotes the total closed domain edge length of the element to be matched in the Voronoi-like diagram, denotes the decision function value of the th neighbor matching pair. represents the degree of matching support of the adjacent surface element to the matching surface element; topological similarity of the face elements a and b.

5. The matching method of claim 1, wherein, The process of training the matching relationship identification model in the step 2 comprises: Step 201, candidate matching relationships of each building are obtained, buildings with matching relationships are identified, and similarity feature vectors of the buildings are marked as positive samples, and candidate matching pairs without matching relationships are marked as negative samples; Step 202, the training sample set is divided into a strict training sample set and a general training sample set according to area overlap, a strict identification model is trained by using position similarity, direction similarity, area similarity and shape similarity of the strict sample set, and a general identification model is obtained by using position similarity, direction similarity, area similarity, shape similarity, topological similarity and context similarity of the general training sample set; Step 203, a Gaussian kernel function is selected to construct an OCSVM classification model as the matching relationship identification model.

6. The matching method of claim 5, wherein, The process of predicting a matching relationship based on the trained matching relationship identification model in the step 2 comprises strict matching relationship prediction and general matching relationship prediction: The strict matching relationship prediction comprises: selecting samples to be matched from the reference data A and the target data B to obtain a candidate matching pair; and predicting the position, direction, area and shape similarity of the candidate matching pair by using a trained strict classification model, and outputting prediction results 1 and -1 respectively representing matching and non-matching; The general matching relationship prediction comprises: traversing a set of unmatching pair elements in the initial matching result and , assuming that a set of buildings to be matched is 、 , exhaustively a candidate matching condition of a certain element to be matched , calculating a similarity feature value of the element to be matched, obtaining a final prediction result by using a general matching relationship identification model, determining a matching relationship, and updating a matched surface element to original data to provide context information for remaining matching; iteratively, until all elements are traversed, to obtain a matching result of the surface element pair with a matching relationship of 1:

1.

7. The matching method of claim 2, wherein, The process of generating the inner boundary in step 3 comprises: Step 301: skeletonizing the blank area of the surface element to obtain a Voronoi-like partition graph; Step 302: performing hierarchical clustering on the matched surface elements in each region of the road network to obtain a clustering center and a direction; Step 303: taking the clustering center as a starting point, fusing the Voronoi-like graph of the matched surface elements according to the clustering direction, and performing boundary merging; Step 304: when the Voronoi-like graph of the class without matched surface elements cannot be fused, determining whether the class is adjacent to other classes, and if so, continuing to fuse, and if not, fusing along the direction until the boundary is closed or an inner ring is generated.

8. The matching method of claim 1, wherein, The step 4 comprises: Step 401: obtaining a complete candidate matching set of the un-matched surface elements after alignment, and performing exhaustive combination on the un-matched surface elements under the constraint of the adjacent relationship; Step 402: using the Concave Hull merging method to aggregate the building groups with non-empty matching and non-1:1 matching relationship into a single element, and converting the matching relationship into a 1:1 matching relationship; Step 403: repeating the prediction matching relationship processing method of step 2 on the merged data to identify the 1:N, M:1 and M:N matching relationships of the un-matched surface elements by using the strict recognition model and the general recognition model; Step 404: counting the surface elements with the matching relationship of 1:1 obtained in step 2 and the matching relationship of 1:N, M:1 and M:N obtained in step 403, and identifying the surface elements that have not been matched after steps 2 and 403 as 1:0 or 0:1 matching relationship.

Images