Nonlinear transformation invariant cluster target matching method based on local topological characteristics

Abstract

A non-linear transformation invariant cluster target matching method based on local topological features comprises the steps that multiple groups of source tetrads and target tetrads are screened from a seeker view point set according to the nearest neighbor principle, after effective description of point targets is constructed through a topological structure, the point targets are matched according to non-linear transformation among views, and the target targets are matched with the source tetrads and the target tetrads. A target point set is mapped to a linear space by using a nonlinear compensation factor, and accurate matching of point targets among different views is realized through matching point pair search and cost matrix optimization. According to the method, the non-linearly distorted point set is compensated, and the base vector linear representation is combined to ensure that the point target description has invariance for the non-linear transformation, so that the high-robustness matching effect is realized, and the method can be widely applied to unmanned aerial vehicle cluster matching, air defense missile multi-target identification and multi-view point set matching tasks in complex scenes.

Description

Technical Field

[0001] The present invention relates to a technology in the field of image processing, specifically a non-linear transformation invariant cluster target matching method based on local topological features. Background Art

[0002] The existing method of using combined multi-missile seeker detection views to complete the matching task of UAV cluster targets has the following three characteristics: First, the field of view of air defense missiles has the sky as the background, and there are no landmark objects as references in most cases; Second, when the cluster targets are at a relatively long distance, it is difficult for air defense missiles to obtain their individual visual features. Third, due to different coordinate systems, there are non-linear distortions in the cluster topology of UAV clusters in different seeker detection views. Summary of the Invention

[0003] In view of the above deficiencies of the prior art, the present invention proposes a non-linear transformation invariant cluster target matching method based on local topological features, which constructs a descriptor for each target by using the local topology around each point target. While ensuring the specificity of the point target representation within the same missile view by the disorder of the cluster target itself and the parallax effect caused by two-dimensional projection, linear compensation is performed on two-dimensional non-linear distortion, so that the topological coding is invariant to two-dimensional non-linear transformation, ensuring the repeatability of the same point target representation in different missile views and realizing the matching of two-dimensional cluster targets between different perspectives.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to a non-linear transformation invariant cluster target matching method based on local topological features, including:

[0006] Step 1: Processing of point sets in two seeker detection views: Based on the nearest neighbor principle, source quadruples and target quadruples are respectively constructed for the point sets in the detection views of seeker A and seeker B, specifically including:

[0007] 1.1 Select several combinations of four adjacent points from the point set in the detection view of seeker A according to the nearest neighbor principle to form source quadruples, which are used to construct the triple for the basis vector and the unit group for solving the non-linear compensation factor. Specifically: For all points in the point set in the detection view of seeker A, find four points with the closest Euclidean distance and at least three non-collinear points to form a source quadruple.

[0008] 1.2 Expand the source quadruples obtained in step 1.1: For all points in the point set in the detection view of seeker A, find N groups of four points with the closest Euclidean distance and at least three non-collinear points to form several source quadruples, which are used to extract the interference of invalid points introduced by the error of the cluster point set from the seeker detection view.

[0009] 1.3 Construct a quadruple for each point in the detection view point set of seeker B according to the nearest neighbor principle, namely the target quadruple, which is used for constructing the triple for the basis vector and the unit group for solving the nonlinear compensation factor. Specifically: for each point in the detection view point set of seeker B, sort the other points according to their Euclidean distances from this point, and select the three points closest to this point to form a quadruple with this point.

[0010] 1.4 Expand the target quadruple obtained in step 1.3: find the N points closest to each point in the detection view point set of seeker B, and take three of them in a traversal manner to form several target quadruples with the current point. At this time, one point corresponding to one quadruple is expanded to one point corresponding to multiple quadruples.

[0011] Step 2: Search for matching source quadruples and target quadruples in a double-loop manner. Specifically: take the traversal of the source quadruple as the outer loop, and perform point target description for each point in the detection view point set of seeker A according to this source quadruple; for each source quadruple, take the traversal of the target quadruple as the inner loop, that is, after performing point target description for each point in the detection view point set of seeker B according to the target quadruple, generate a cost matrix according to the difference in point target descriptions between the two seeker detection view point sets, and use the Hungarian algorithm to obtain the corresponding optimal matching point pairs under the current source quadruple and target quadruple. If the number M of the optimal matching point pairs is greater than the threshold ξ, it is considered that the current source quadruple and target quadruple are matched, and the corresponding optimal matching point pairs are the mutually matched cluster targets; at this time, the double-loop will terminate.

[0012] The above-mentioned point target description is based on the cluster topology and is obtained through the following method: after decomposing the source quadruple and the target quadruple into a triple and a unit group respectively, use the basis vector composed of the triple to linearly represent all point targets in the seeker detection view point set; use this linear representation coefficient as the corresponding point target description; it has been proved that the linear representation coefficient is invariant to linear transformation. For non-linear distortion, the present invention uses the non-linear compensation factor composed of the unit group to perform non-linear compensation, so that the linear representation coefficient is invariant to two-dimensional non-linear transformation.

[0013] The specific content of step 2 includes:

[0014] 2.1 For each source quadruple constructed in step 1.2, take its triple to describe all other points except this source quadruple. Specifically, take one point in the triple as the origin to construct a two-dimensional space basis vector, use this origin to vectorize all other points except this source quadruple to generate vectors for all other points, and then use this basis vector to linearly represent the vectors for all other points except this source quadruple; take the linear representation coefficient corresponding to each vector as the description of the target of the point corresponding to this vector.

[0015] 2.2 For each source quadruple, traverse all target quadruples and determine the non-linear compensation factor λ x and λ y , specifically including: for the target quadruple whose corresponding point is U' in the current traversal, assume it matches the current source quadruple whose corresponding point is point U. After vectorization respectively, form and For the source quadruple, use its triple to linearly represent the unit group, and find the linear representation coefficients α U and β U , that is After that, solve the following two equations simultaneously to find the non-linear compensation factor λ x and λ y : -p U'x +α U p A'x +β U p B'x -(α U +β U -1)p O'x =λ x [p Ux p U'x -α U p Ax p A'x -β U p Bx p B'x +(α U +β U -1)p Ox p O'x +λ y [p Uy p U'x -α U p Ay p A'x -β U p By p B'x +(α U + β U -1)p Oy p O'x ,-pU'y +α U p A'y +β U p B'y -(α U +β U -1)p O'y =λ x [p Ux p U'y -α U p Ax p A'y -β U p Bx p B'y +(α U +β U -1)p Ox p O'y +λ y [p Uy p U'y -α U p Ay p A'y -β U p By p B'y +(α U +β U -1)p Oy p O'y ,where: p ·x represents the x - coordinate value of point · in the detection view of seeker A, and p ·y represents the y - coordinate value of point · in the detection view of seeker A; p ·'x represents the y - coordinate value of point · in the detection view of seeker B, and p ·'y represents the y - coordinate value of point · in the detection view of seeker B.

[0016] 2.3 For the currently traversed source quadruple and target quadruple, use the nonlinear compensation factors λ x and λ y to perform nonlinear compensation on all points in the point set of the detection view of seeker B, so as to map them onto a linear space that satisfies the invariance of the linear representation coefficients. Specifically: for all points C’ in the detection view of seeker B, multiply both its horizontal and vertical coordinates by the compensation coefficient λ x ·p C'x +λ y ·p C'y +1 to form the linearized point set of the detection view of seeker B, that is, the positions where all points in the detection view of seeker B should appear after linearization; the corresponding linearized result of the target quadruple is the linearized target quadruple.

[0017] 2.4 For the currently traversed source quadruple and target quadruple, the linear representation coefficient αC and β C are applied to the basis vectors formed by the linearized target quadruples to generate the actual positions where all points in the detection view of seeker B appear after linearization when the current source quadruple and the target quadruple match.

[0018] 2.5 Use the distance between the actual positions where all points in the detection view of seeker B appear after linearization and the positions where all points in the detection view of seeker B should appear as the cost matrix and use the Hungarian algorithm to obtain the optimal matching point pairs.

[0019] The Hungarian algorithm mentioned above adopts but is not limited to the technical implementation described in Kuhn H W. The Hungarian method for the assignment problem([J]. Naval research logistics quarterly, 1955, 2(1 - 2): 83 - 97).

[0020] 2.6 If the number M of the optimal matching point pairs corresponding to the current source quadruple and the target quadruple is ≥ ξ, it is determined that the current source quadruple and the target quadruple can be matched, and the corresponding optimal matching point pairs are the mutually matching cluster targets, where: ξ is the experimental threshold, which is mainly related to the errors in the upstream weak target recognition task, and it can also be extended to an adaptive threshold, that is, this value is gradually decreased during the search process. In subsequent experiments, ξ = 0.8min{N A , N B}, where: N A is the total number of points in the detection view of seeker A, and N B is the total number of points in the detection view of seeker B. Technical effects

[0021] The present invention compensates the detection view point sets of the seeker with a non - linear compensation factor to map them onto a linear space that satisfies the invariance of the linear representation coefficients; thereby realizing the non - linear transformation invariance of the point sets. Compared with the prior art, the present invention solves the non - linear distortion of the detection view point sets of the seeker caused by the different seeker coordinate systems. Brief description of the drawings

[0022] Figure 1 is a schematic diagram of the point sets extracted after weak target recognition for the detection views of seeker A and seeker B;

[0023] Figure 2 is a flow chart of topological coding based on vector linear representation;

[0024] In the figure: O, A, B are triples, and C is any other point in the detection view point set of the seeker except the quadruples;

[0025] Figure 3 This is the flowchart of the present invention;

[0026] Figure 4 This is a schematic diagram of the spatial distribution of the two-dimensional point set A;

[0027] Figure 5 This is a schematic diagram of the spatial distribution of the two-dimensional point set B;

[0028] Figure 6 This is a schematic diagram of the matching result of the affine transformation; Detailed implementation manners

[0029] As Figure 3 shown, the present invention relates to a non-linear transformation invariant cluster target matching method based on local topological features, including:

[0030] Step 1: Processing of the detection view point sets of two seekers: Based on the nearest neighbor principle, a source quadruple and a target quadruple are respectively constructed for the detection view point sets of seeker A and seeker B;

[0031] Step 2: Searching for matching source quadruples and target quadruples in a double-loop manner. Specifically: Taking the traversal of the source quadruple as the outer loop, point target descriptions are made for each point in the detection view point set of seeker A according to this source quadruple; for each source quadruple, taking the traversal of the target quadruple as the inner loop, that is, after making point target descriptions for each point in the detection view point set of seeker B according to the target quadruple, a cost matrix is generated according to the differences in point target descriptions between the detection view point sets of the two seekers, and the Hungarian algorithm is used to obtain the corresponding optimal matching point pairs under the current source quadruple and target quadruple. If the number M of the optimal matching point pairs is greater than the threshold ξ, it is considered that the current source quadruple and target quadruple are matched, and the corresponding optimal matching point pairs are the mutually matching cluster targets; at this time, the double-loop will terminate.

[0032] As Figure 2 shown, take points O, A, and B to form a triple. Taking point O as the origin, points A and B form two-dimensional basis vectors and At this time, the linear representation of any point C other than the triple is the two-dimensional basis vector where: the linear representation coefficients α C and β C are the descriptions of point C.

[0033] 1000 sets of samples were collected for specific experiments. The specific settings for sample collection are as follows: Any set of point coordinates randomly generated in a uniform distribution is used to simulate the point set A corresponding to the current view; the number of points is also generated in a uniform distribution, and its range is within [8, 50]. Then, a 3×3 transformation matrix is randomly generated in a uniform distribution. As long as the last row of this transformation matrix is not [0 0 1], it is a non-linear transformation for two-dimensional coordinates. This transformation matrix is used to transform the randomly generated set of point coordinates to simulate the point set B corresponding to another view.

[0034] For ease of understanding, in this embodiment, one sample with the greatest matching difficulty is selected from the randomly generated 1000 samples for visualization and description: The number of point targets in the point set A is 10, and its coordinate distribution in the two-dimensional space is as Figure 4 shown and all points are numbered.

[0035] The point set A is transformed using this matrix to obtain a two-dimensional point set B as shown in Figure 5 shown. As shown in the figure, after the non-linear transformation, the topology of the point set is distorted, and its distribution is linear; numbering processing corresponding to Figure 4 is performed on the two-dimensional point set B, and the final matching result is as shown in Figure 6 shown. It can be seen that the present invention can effectively match data with such topological distortion.

[0036] The statistical results of the matching success rates of the above 1000 sets of samples are shown in Table 1. To verify the robustness of the algorithm, a small number of outliers are added to the point set A or the point set B to simulate errors in the upstream recognition task. At the same time, in all scenarios, the number of effectively matched points is guaranteed to be greater than or equal to 6.

[0037] Table 1

[0038] Compared with the prior art, for a linear distribution structure as shown in Figure 5 which is regarded as having lost its original effective topological structure and thus unable to achieve effective matching for it, the present method can effectively solve such problems.

[0039] The above specific implementation can be locally adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation. All implementation solutions within its scope are subject to the constraints of the present invention.

Claims

1. A nonlinear transformation invariant cluster target matching method based on local topological features, characterized in that: include: Step 1: Processing of detection view point sets of two seekers: constructing source quadruple and target quadruple for the detection view point sets of seeker A and seeker B respectively based on the nearest neighbor principle; Step 2, searching for matching source quadruple and target quadruple in a double loop, specifically: using the traversal of the source quadruple as the outer loop, describing each point in the detection view point set of the seeker A according to the source quadruple; for each source quadruple, using the traversal of the target quadruple as the inner loop, that is, describing each point in the detection view point set of the seeker B according to the target quadruple, generating a cost matrix according to the difference in the point target descriptions between the two seeker detection view point sets, and using the Hungarian algorithm to obtain the corresponding optimal matching point pairs under the current source quadruple and the target quadruple, if the number M of the optimal matching point pairs is greater than the threshold ξ, it is considered that the current source quadruple and the target quadruple are matched, and the corresponding optimal matching point pairs are the mutually matching cluster targets; at this time, the double loop will terminate.

2. The nonlinear transformation invariant cluster target matching method based on local topological features according to claim 1 is characterized in that: The step 1 specifically includes: 1.1 According to the nearest neighbor principle, a combination of several adjacent four points is selected from the detection view point set of the seeker A to form a source four-tuple, which is used to construct a triple of basis vectors and a unit group for solving the nonlinear compensation factor. Specifically, for all points in the detection view point set of the seeker A, four points with the closest distance and at least three points not collinear are found according to the Euclidean distance to form a source four-tuple; 1.2 Expand the source quadruple obtained in step 1.1: for all points in the detection view point set of the seeker A, find N groups of four points with the shortest distance and at least three points not collinear according to the Euclidean distance to form several source quadruple groups, which are used to extract the invalid point interference introduced by the cluster point set error from the detection view of the seeker; 1.3 According to the nearest neighbor principle, a quaternion is constructed for each point in the detection view point set of the seeker B, that is, a target quaternion, which is used to construct a triplet of basis vectors and a unit group for solving the nonlinear compensation factor. Specifically, for each point in the detection view point set of the seeker B, other points are sorted according to the Euclidean distance between the point and other points, and the point and the three points closest to it are selected to form a quaternion; 1.4 Expand the target quadruple obtained in step 1.3: find the N points closest to each point in the detection view point set of the seeker B, and take three of them and the current point in a traversal manner to form several target quadruple groups about the point. At this time, one point corresponding to one quadruple is expanded to one point corresponding to multiple quadruple groups.

3. The nonlinear transformation invariant cluster target matching method based on local topological features according to claim 1 is characterized in that: The point target description is based on cluster topology and is obtained in the following way: after decomposing the source quadruple and the target quadruple into triplets and unigrams respectively, all point targets in the detection view point set of the seeker are linearly represented using basis vectors composed of triplets; and the linear representation coefficients are used as the corresponding point target description.

4. The nonlinear transformation invariant cluster target matching method based on local topological features according to claim 3 is characterized in that: For nonlinear distortion, a nonlinear compensation factor consisting of a tuple is used to perform nonlinear compensation, so that the linear representation coefficient is invariant to the two-dimensional nonlinear transformation.

5. The nonlinear transformation invariant cluster target matching method based on local topological features according to claim 1 or 3, characterized in that: The step 2 specifically includes: 2.1 For each source quadruple, take its triplet to describe all the remaining points except the source quadruple, specifically including: constructing a two-dimensional space basis vector with one point in the triplet as the origin, using the origin to vectorize all the other points except the source quadruple to generate vectors about all the other points, and then using the basis vector to linearly represent the vectors about all the other points except the source quadruple; taking the linear representation coefficient corresponding to each vector as the description of the point target corresponding to the vector; 2.2 For each source quadruple, traverse all target quadruple and determine the nonlinear compensation factor λ x and λ y Specifically, for the target quadruple whose corresponding point is U', it is assumed that it matches the current source quadruple whose corresponding point is point U, and vectorization is performed to form and For the source quaternary group, use its triplet to linearly represent the unit group and find the linear representation coefficient α corresponding to the unit group U and β U ,Right now Then, the nonlinear compensation factor λ is obtained by combining the following two equations: x and λ y :-p U'x +α U p A'x +β U p B'x -(α U +β U -1)p O'x =λ x [p Ux p U'x -α U p Ax p A'x -β U p Bx p B'x +(α U +β U -1)p Ox p O'x ]+λ y [p Uy p U'x -α U p Ay p A'x -β U p By p B'x +(α U +β U -1)p Oy p O'x ], -p U'y +α U p A'y +β U p B'y -(α U +β U -1)p O'y =λ x [p Ux p U'y -α U p Ax p A'y -β U p Bx p B'y +(α U +β U -1)p Ox p O'y ]+λ y [p Uy p U'y -α U p Ay p A'y -β U p By p B'y +(α U +β U -1)p Oy p O'y ], where: p ·x represents the x-coordinate value of point · in the detection view of seeker A, p ·y represents the y coordinate value of point · in the detection view of seeker A; p ·'x represents the y coordinate value of point · in the detection view of seeker B, p ·'y represents the y coordinate value of point · in the detection view of seeker B; 2.3 For the currently traversed source quadruple and target quadruple, use the nonlinear compensation factor λ x and λ y Nonlinear compensation is performed on all points in the detection view point set of the seeker B to map them to a linear space that satisfies the invariance of the linear representation coefficient. Specifically, for all points C' in the detection view of the seeker B, their horizontal and vertical coordinates are multiplied by the compensation coefficient λ x ·p C'x +λ y ·p C'y +1, to form a linearized point set of the detection view of the seeker B, that is, the positions where all points of the detection view of the seeker B should appear after linearization; the corresponding linearization result of the target four-tuple is the linearized target four-tuple; 2.4 For the currently traversed source quadruple and target quadruple, the linear representation coefficient α C and β C Apply to the basis vectors formed by the linearized target quadruple to generate the actual positions of all points in the detection view of the linearized seeker B when the current source quadruple matches the target quadruple; 2.5 The distance between the actual position of all points in the linearized detection view of seeker B and the position where all points in the linearized detection view of seeker B should appear is used as the cost matrix and the Hungarian algorithm is used to obtain the optimal matching point pair; 2.6 If the number of optimal matching point pairs corresponding to the current source quadruple and the target quadruple is M≥ξ, the current source quadruple and the target quadruple are determined to be matched, and the corresponding optimal matching point pairs are the mutually matching cluster targets, where: ξ is the threshold.

6. The nonlinear transformation invariant cluster target matching method based on local topological features according to claim 5 is characterized in that: The threshold ξ=0.8min{N A ,N B }, where: N A is the total number of detection view points of seeker A, N B is the total number of detection view points of seeker B.

Images