CN104408401B - A kind of In-flight measurement method of time critical target - Google Patents

Abstract

The present invention is an on-orbit detection method of a time-sensitive target. The method includes: step S1: selecting various time-sensitive target training areas on historical images, and extracting high-dimensional multi-scale concentricity from each pixel of each training image. Circular cluster direction gradient features, offline learning of the structure dictionary of various time-sensitive targets; Step S2: Extract high-dimensional multi-scale concentric circular cluster directional gradient features at each pixel on the image of each time phase currently in orbit, Use the structure dictionary to solve the sensitive target type indicator vector, identify the position and type of the suspicious target according to the structural sparseness of the time-sensitive target type indicator vector, and extract the suspicious target area; Step S3: Detect on the on-orbit images in different time phases The trajectory of the suspicious target area is analyzed, and the time-sensitive target is identified on-orbit according to the singularity of the motion trajectory; Step S4: The image of the on-orbit time-sensitive target is used as the training image for the on-orbit incremental update of the structure dictionary, and returns to step S1.

Description

A method for on-orbit detection of time-sensitive targets

technical field

The invention relates to the technical fields of on-orbit image processing, target detection, target recognition, target monitoring, etc., in particular to an on-orbit detection method of a time-sensitive target.

Background technique

Compared with ordinary targets, time-sensitive targets have strong timeliness, and time-sensitive targets must be identified within a limited time window, which is fleeting. At the same time, time-sensitive targets are often very important targets, once the opportunity of identification is lost, it will cause heavy losses. Therefore, the detection and recognition of time-sensitive targets has important research significance, but at the same time it is more challenging.

With the development of high spatial resolution and high temporal resolution remote sensing satellites, it becomes possible to detect and identify time-sensitive targets using satellite images in orbit. Compared with other means of data acquisition, satellite images have a large range, which is conducive to accurate and long-term tracking of time-sensitive targets.

The difficulty of time-sensitive target detection mainly lies in the complexity of the time-sensitive target. The time-sensitive target only presents the characteristics of the time-sensitive target when the state changes or the trajectory changes at a certain point in time, and this key time point is difficult to be captured. . For the on-orbit detection of time-sensitive targets, there are few prior knowledge and data that can be used. How to use the latest data to automatically adjust the target model trained in the offline state is the key to time-sensitive target detection. However, the key technologies mentioned above are very immature at present, which limits the practical application of online detection of time-sensitive targets.

Contents of the invention

The purpose of the present invention is to provide an effective on-orbit detection method for time-sensitive targets in view of the characteristics of on-orbit processing and the requirements of practical applications.

In order to achieve the above object, the on-orbit detection method of the time-sensitive target of the present invention, the method comprises steps as follows:

Step S1: Select various time-sensitive target training areas on historical images, extract high-dimensional multi-scale multi-scale concentric circular cluster direction gradient features from each pixel of each training image for each type of time-sensitive target, and learn offline A structure dictionary for various time-sensitive objects;

Step S2: Extract the high-dimensional multi-scale directional gradient features of concentric circular clusters from each pixel of the current on-orbit image at each time phase, and use the structure dictionary to solve the projection coefficients of the multi-scale concentric circular cluster directional gradient features. Target type indicator vector, identifying the position and type of the suspicious target according to the structural sparseness of the time-sensitive target type indicator vector, and extracting the suspicious target area according to the similarity of the time-sensitive target type indicator vector;

Step S3: Analyze the trajectory of the suspicious target area detected on the on-orbit images of different time phases, and identify the time-sensitive target on-orbit according to the singularity of the motion trajectory;

Step S4: Use the image of the on-orbit time-sensitive target as the training image for the on-orbit incremental update of the structure dictionary, and return to step S1.

The method of the present invention has important significance for improving the universality and automation degree of time-sensitive target on-orbit detection, and its main advantages are as follows:

The present invention combines the historical training data with the latest current data, embodies the prior constraints of the time-sensitive target through the historical training data, and ensures that the requirements and data characteristics can be well combined in the on-orbit processing environment without human intervention Combine the features of the time-sensitive target contained in the historical data and the new features of the current image, and improve the representation ability of the dictionary and greatly save the amount of calculation through the incremental update of the dictionary on orbit.

In the target detection stage, the present invention uses the time-sensitive target type indicator vector to indicate the target type to which the pixel belongs, which overcomes the uncertainty of the scalar representation method; extracts the target area according to the similarity of the time-sensitive target type indicator vectors between pixels, and improves the noise immunity. and occlusion robustness.

The present invention utilizes the distance metric based on the generalized eigenvalue of the covariance matrix of the target area in the abnormal detection stage of the motion state, which has good robustness to the change of the viewing angle, and reduces the false alarm rate of time-sensitive target recognition; In the first stage, the space-time trajectory change curve is converted to polar coordinate space, which effectively depicts the motion singularity of time-sensitive targets and improves the accuracy of time-sensitive target recognition.

Benefiting from the above advantages, the present invention makes the on-orbit detection of time-sensitive targets possible, greatly improves the timeliness, robustness and automation of time-sensitive target detection and identification, and can be widely used in time-sensitive target discovery and monitoring , Target monitoring and other systems.

Description of drawings

Fig. 1 is a flowchart of an on-orbit detection method for a time-sensitive target according to the present invention.

Figure 2 is an anomaly detection diagram of spatio-temporal trajectories.

detailed description

The technical problems involved in the technical solution of the present invention will be described below in conjunction with the accompanying drawings. It should be pointed out that the described embodiments are only intended to facilitate the understanding of the present invention, rather than limiting it in any way.

As shown in Figure 1, the present invention proposes an on-orbit detection method of a time-sensitive target. The implementation steps are as follows:

Step S1: Select various time-sensitive target training areas on historical images, extract high-dimensional multi-scale multi-scale concentric circular cluster direction gradient features from each pixel of each training image for each type of time-sensitive target, and learn offline A dictionary of structures for various types of time-sensitive objects.

The multi-scale concentric ring cluster direction gradient feature takes the sampling point as the center and samples the image block with the sampling scale as the radius, and constructs three concentric ring structures with different radii, and the corresponding sampling points are located at the concentric On the ring, 8 sampling points are extracted at equal angular intervals of 45° on each concentric ring. The sampling points on the same radius have the same Gauss scale value, and the Gauss scale values of the sampling points on different radii are different. The specific process of the multi-scale concentric circular cluster direction gradient feature extraction is as follows:

Step S01: Calculate the 8 direction gradients of each pixel (u, v) of the image block with the sampling point as the center and the sampling scale Σ as the radius, and then use Gaussian kernel convolution to obtain the direction at (u, v) The gradient feature vector h _∑ (u, v) is expressed as follows:

u and v are the row number and column number of the pixel respectively, and T represents the transposition of the vector, Indicates the gradient vector obtained by convolution of the m-th directional gradient with a Gaussian kernel, m is the direction number, m=1, 2,...,8.

Step S02: The multi-scale concentric circular cluster direction gradient feature D(u, v) is the union of a series of related vectors describing each position in the local support area of the sampling point (u, v), and D(u, v) The representation is as follows:

Wherein, l _m2 (u, v, R _n2 ) represents the coordinates of the m2 sampling point on the n2 concentric ring of the pixel point (u, v), Represents the local orientation gradient histogram of the m2th sampling point on the n2th concentric ring of the pixel point (u, v), n is the sampling scale number, n2 is the concentric ring number, and m2 is the sampling point number.

The structure dictionary learning is to learn a low-dimensional and divisible dictionary from the high-dimensional multi-scale concentric circular cluster direction gradient feature vector set and the corresponding target type number. set image is the i-th training image of target type j, The number of pixels is N _i , then from the image N _i multi-scale concentric ring cluster direction gradient feature vectors can be extracted, and the union of these N _i feature vectors is used as the feature of the target type j. For the convenience of description, the training feature set of the jth type of target is recorded as X _j ＝{x _k ＝{j, f _k }|1≤k≤A _j }, and X _k ＝{j, f _k } represents the k training samples, j is the number of the target type, f _k is the multi-scale concentric circular cluster direction gradient feature vector corresponding to the kth training sample, and A _j is the number of elements in the jth class target training feature set. The structure dictionary learning model of the present invention is as follows:

Among them, matrix X is the set of multi-scale concentric circular cluster direction gradient feature vectors obtained from all training images, the dimension of matrix X is M rows and N columns, and M is the dimension of multi-scale concentric circular cluster direction gradient feature vectors, N is the number of all training samples. The matrix D is a structure dictionary, and the dimension of the structure dictionary D is M rows and K columns. Each column of the structure dictionary D is called a dictionary atom, and K is the number of structure dictionary atoms. Matrix Z=[z ₁ ^T ; z ₂ ^T ; ...; z _a ^T ; ...; z _N ^T ] ^T is the projection coefficient matrix obtained by using the structure dictionary of matrix X, and T represents the transposition of vector or matrix. The K-dimensional vector z _a represents the ath column of the matrix Z, 1≤a≤N. ||·|| _F , ||·|| ₁ and ||·|| ₂ represent the Frobenius norm, 1 norm and 2 norm of the matrix, and λ ₁ and λ ₂ are regularization coefficients, respectively controlling the projection coefficient Sparsity and separability. W _{i1, i2} represent similar weights of training samples z _i1 and z _i2 , if z _i1 and z _i2 are the projection coefficients of training samples of the same type of target, then W _{i1, i2} = 1, if z _i1 and z _i2 are different The projection coefficient of the training sample of the type target, then w _{i1, i2} =0.

The specific process of solving the structure dictionary learning model is as follows:

Step S11 initial value setting. Set the regularization coefficients λ ₁ and λ ₂ , where λ ₁ =λ ₂ =0.01 in the present invention. Perform principal component analysis on the multi-scale concentric circular cluster direction gradient feature set of each type of target to obtain the eigenvector corresponding to the significant eigenvalue, and the union of the eigenvectors of different types of targets is used as the initial value D of the structure dictionary D ⁽⁰⁾ , the initial value Z ⁽⁰⁾ of the projection coefficient Z =([D ⁽⁰⁾ ] ^T D ⁽⁰⁾ ) ^-1 [D ⁽⁰⁾ ] ^T X. Significant eigenvalues refer to the first L eigenvalues that exceed 90% of the energy of all eigenvalues after the eigenvalues are arranged in descending order. The number of all significant eigenvalues is the number K of dictionary atoms.

Step S12 Alternate iterative updating of the structure dictionary and the projection coefficient matrix. Let D ^(t) and Z ^(t) be the solutions of the structure dictionary D and the projection coefficient matrix Z at the tth generation, and alternately iteratively update the structure dictionary and the projection coefficient matrix according to the following formula:

Among them, S is a diagonal matrix, and the elements on the diagonal The value of row b and column l of matrix W is w _b,1 , 1≤b≤N, 1≤1≤N. D _{r, c} ^(t+1) and D _{r, c} ^(t) respectively represent the value of the rth row and the cth column of the solution at the (t+1)th and tth iteration of the structure dictionary; Z _{r, c} ^(t) represents the value of row r and column c of Z ^(t) , r and c are the row number and column number of the matrix, 1≤r≤M, 1≤c≤K. [Z ^(t) ] ^T represents the transpose of the matrix Z ^(t) . The criterion for the update stop of alternating iterations is t<100 or mse(D ^(t+1) Z ^(t+1) , D ^(t) Z ^(t) )<ε, mse(D ^(t+1 )Z ^{(t+ 1} ), D ^(t) Z ^(t) ) represents the mean square error of two adjacent iterations, ε is a threshold, and ε=0.1 in the present invention.

Step S2: Extract the high-dimensional multi-scale directional gradient features of concentric circular clusters from each pixel of the current on-orbit image at each time phase, and use the structure dictionary to solve the projection coefficients of the multi-scale concentric circular cluster directional gradient features. The target type indicator vector identifies the position and type of the suspicious target according to the structural sparseness of the time-sensitive target type indicator vector, and extracts the suspicious target area according to the similarity of the time-sensitive target type indicator vector. The specific process is as follows:

Step S21 constructs a new projection matrix according to the structure dictionary, and projects the multi-scale concentric circular cluster direction gradient feature vector of each pixel of the image currently on orbit at each time phase to obtain the time-sensitive target type indicator vector of each pixel, The specific process is as follows:

Let F _{r2, c2} be the multi-scale concentric circular cluster direction gradient feature vector of row r2 and column c2 on the current image, then the time-sensitive target type target type indicator vector corresponding to F _{r2, c2} is CE _{r2, c2} = PF _{r2 , c2} , CE _{r2, c2} is a K-dimensional vector. Where P=(D ^T D+λI) ^-1 D ^T , λ is a correction coefficient, and λ=0.1 in the present invention. I is the identity matrix.

Step S22 determines the suspicious target according to the structural sparsity of the time-sensitive target type indicator vector reflecting the type information of the target to which the pixel belongs. The specific process is as follows:

Assuming that the number of salient features of the j-th type of target is K _j , then the corresponding K _j dictionary atoms in the obtained dictionary represent the features of the j-th type of target. Correspondingly, the time-sensitive target type indication vector CE _{r2, c2} contains the type information of the target. Specifically, the energy and sparseness of the corresponding K _j -th segment components in the vector CE _r2,c2 reflect the probability that the target belongs to the j-th type of target. If CE _{r2, c2} has the largest energy of the K _jth segment component and the other segment components are very sparse, it means that the r2th row and c2th column pixel on the current image is a part of the jth type of target. If the energy and sparseness of each segment component are not much different, it means that the pixel is a background area. Compared with the traditional scalar object type representation method, this vector object type representation method is more robust.

Step S23 is to segment the image of each time phase on the track according to the similar target type indicator vectors of the pixels located in the same area, and extract the suspicious target area. The image segmentation of the present invention is a segmentation algorithm based on graph theory. To this end, first construct an undirected graph G=(V; E), each pixel point pi in the image corresponds to a vertex vi in V, V is the vertex set of the undirected graph G, and E is the vertex of the undirected graph G A collection of edges; the weight d((vi, vj)) of edge (vi, vj) is the difference between the time-sensitive target type indicator vectors of pixels pi and pj. Target area extraction is the specific process of merging and splitting the undirected graph G. The steps are as follows:

S231: Initialize. Arrange E in ascending order according to the weight, and get an ordered set of edges π=(o ₁ ,..., o _m3 ), the weight of edge o ₁ is the smallest, and the weight of edge o _m3 is the largest; calculate the initial segmentation S ⁰ , each region Contains only one vertex vi, m3 is the number of edges in the edge set E of the undirected graph G.

S232: Let S ^q-1 and S ^q represent the segmentation results including q-1 edges and q edges respectively, q represents the sequence number of the edges in the ordered edge set π, 1≤q≤m3. For q (1≤q≤m3), perform the following operations: Given S ^q-1 and the vertices vi and vj connected to its qth edge, let with are respectively the set of regions containing vi and vj in S ^q-1 , construct S ^q as follows: if the region and the weight merge regions and area if area or weight Then S ^q =S ^q-1 ; where: express with The minimum value of the function collection of control areas The overall similarity of wood inventions in Indicates the area The number of pixels in , tt=vi or tt=vj. The value of α reflects the observation scale, and α=200 in the present invention. Indicates the area The internal dissimilarity based on the maximum weight representation of the minimum spanning tree, that is Indicates the area Based on the minimum spanning tree, d(e) represents the weight of the edge on the minimum spanning tree.

S233 The final segmentation result is S=S ^m3 .

Step S3: Analyze the trajectory of the suspicious target area detected on the on-orbit images of different time phases, and identify the time-sensitive target on-orbit according to the singularity of the motion trajectory. The singularity of the motion trajectory is mainly manifested in the following two situations: abnormal motion state (sudden appearance or disappearance of suspicious objects, drastic changes in appearance), and abnormal space-time trajectory. The present invention recognizes time-sensitive targets according to these two singularities. The specific process of on-orbit identification of the time-sensitive target is as follows:

Step S31: Use the distance difference of the covariance matrix of the time-sensitive target type indicator vector set of the suspicious target area to find the nearest neighbor at the approaching moment for each suspicious target, if the direction gradient of the multi-scale concentric circle cluster between the nearest neighbor areas The covariance matrix between the eigenvector sets is still the nearest neighbor, which means that the target’s motion state is not abnormal, that is, it is a time-insensitive target; If the covariance matrix between is not the nearest neighbor, then the abnormal motion state of the target is a time-sensitive target. The specific process is as follows:

Assume that the number of j-th type targets detected on the images of two adjacent moments t ₁ and t ₂ are respectively with with Respectively represent the covariance matrix of the multi-scale concentric ring cluster direction gradient feature vector set and the time-sensitive target type indicator vector set of the j-th type of β-th target at time _tc (c=1, 2), let the time-sensitive Target type indicator vector covariance matrix In the set of covariance matrices The nearest and next neighbors in are respectively with like then means The status or appearance of the target indicated has changed, and the suspicious target is judged as a time-sensitive target; if then means Indicates a target that is not time-sensitive. In the present invention, τ ₁ =0.9 and τ ₂ =0.6. For two covariance matrices A and B, Indicates the nth generalized eigenvalue of matrices A and B, and μ represents the number of generalized eigenvalues of covariance matrices A and B.

Step S32: Project the time-space trajectory change curves of non-time-sensitive targets in different phases into polar coordinates, and identify time-sensitive targets according to the direction changes of adjacent time-space trajectories in polar coordinates. The specific process is as follows:

Assume that the number of non-time-sensitive targets obtained on the current on-orbit multi-temporal image according to step S31 is N1, and the coordinates of the center of mass of the gamma-th target at time g are ζ is the number of multi-temporal images currently in orbit, and its spatio-temporal trajectory change curve is expressed as:

The abnormality of the space-time trajectory usually manifests as a sudden change in the direction of motion. For this reason, Mu invented to convert its space-time trajectory change curve into the form of polar coordinates {(ρ ² , θ ² ), (ρ ³ , θ ³ ),…, (ρ ^ζ , θ ^ζ )}, where Indicates the speed of the space-time trajectory change at time 2≤k1≤ζ, Indicates the direction of the change of the space-time trajectory at time kl.

If θ ^k1 >π/2, it indicates that the moving direction of the suspicious target at time k1 is opposite to the original moving direction, and the space-time trajectory is abnormal, and the suspicious target is a time-sensitive target. Fig. 2 is an anomaly detection diagram of spatio-temporal trajectories, in which the target corresponding to the change track a is a normal target, and the target corresponding to the change track b is a time-sensitive target.

Step S4: Use the image of the on-orbit time-sensitive target as the training image for the on-orbit incremental update of the structure dictionary, and return to step S1. The specific process is as follows:

Step S41: When there are many training samples, in order to meet the timeliness requirements of on-orbit processing, the multi-scale concentric circular cluster direction gradient features of detected on-orbit time-sensitive targets and suspicious targets are used as new training samples.

Step S42: According to the current image and the training image of the on-orbit incremental update of the structure dictionary obtained by the current time-sensitive target, use the new training sample to update the on-orbit incremental update of the structure dictionary at the previous moment. The on-orbit incremental update method of the structural dictionary is similar to the dictionary learning method in step S1, the difference lies in the setting of the initial dictionary. The initial dictionary in step S11 comes from the basis vector of the principal component analysis method, and the initial dictionary in this step comes from an existing dictionary. The updated dictionary combines state-of-the-art image features and time-sensitive object features with stronger representation capabilities.

The above is only a specific implementation mode in the present invention, but the scope of protection of the present invention is not limited thereto. Anyone familiar with the technology can understand the conceivable transformation or replacement within the technical scope disclosed in the present invention. All should be covered within the protection scope of the present invention, therefore, the protection scope of wood invention should be as the criterion with the protection scope of claims.

Claims (10)

1. An on-orbit detection method of a time-sensitive target, the method comprising steps as follows: Step S1: Select various time-sensitive target training areas on historical images, extract high-dimensional multi-scale multi-scale concentric circular cluster direction gradient features from each pixel of each training image for each type of time-sensitive target, and learn offline A structure dictionary for various time-sensitive objects; Step S2: Extract the high-dimensional multi-scale directional gradient features of concentric circular clusters from each pixel of the current on-orbit image at each time phase, and use the structure dictionary to solve the projection coefficients of the multi-scale concentric circular cluster directional gradient features. Target type indicator vector, identifying the position and type of the suspicious target according to the structural sparseness of the time-sensitive target type indicator vector, and extracting the suspicious target area according to the similarity of the time-sensitive target type indicator vector; Step S3: Analyze the trajectory of the suspicious target area detected on the on-orbit images of different time phases, and identify the time-sensitive target on-orbit according to the singularity of the motion trajectory; Step S4: Use the image of the on-orbit time-sensitive target as the training image for the on-orbit incremental update of the structure dictionary, and return to step S1. 2. The method according to claim 1, wherein the structure dictionary is a low-dimensional one learned from a set of high-dimensional multi-scale concentric ring cluster direction gradient feature vectors and corresponding time-sensitive target type numbers Orientation gradient features of multiscale concentric ring clusters. 3. The method according to claim 2, wherein, according to the sparsity of the time-sensitive target type indicator vector and the similarity between the time-sensitive target type indicator vectors at different pixels, the learning model of the structure dictionary is constructed. 4. method according to claim 2, it is characterized in that, carry out principal component analysis to the multi-scale concentric ring cluster direction gradient feature set of each type of time-sensitive target, obtain and will be combined with the characteristic vector corresponding to significant eigenvalue Set as the initial dictionary, and then iteratively update the initial dictionary and projection coefficients alternately to obtain the structure dictionary. 5. The method according to claim 1, wherein the step of extracting the suspicious target area comprises the following steps: Step S21: Construct a new projection matrix according to the structure dictionary, and project the multi-scale concentric circular cluster direction gradient feature vector of each pixel of the image currently on orbit at each time phase to obtain the time-sensitive target type indicator vector of each pixel ; Step S22: Determine the suspicious target according to the structural sparsity of the time-sensitive target type indicator vector reflecting the type information of the target to which the pixel belongs; Step S23: According to the similar target type indicator vectors of the pixels located in the same area, the image of each time phase on the track is segmented, and the suspicious target area is extracted. 6. The method according to claim 1, wherein the step of identifying a time-sensitive target on the track comprises the following steps: Step S31: Use the distance difference of the covariance matrix of the time-sensitive target type indicator vector set of the suspicious target area to find the nearest neighbor at the approaching moment for each suspicious target, if the direction gradient of the multi-scale concentric circle cluster between the nearest neighbor areas The covariance matrix between the eigenvector sets is still the nearest neighbor, which means that the target’s motion state is not abnormal, that is, it is a time-insensitive target; If the covariance matrix between is not the nearest neighbor, then the abnormality of the motion state of the target is a time-sensitive target; Step S32: Project the time-space trajectory change curves of non-time-sensitive targets in different phases into polar coordinates, and identify time-sensitive targets according to the direction changes of adjacent time-space trajectories in polar coordinates. 7. The method according to claim 6, wherein the distance difference between the suspicious target areas is measured by the difference between the covariance matrices of the time-sensitive target type indicator vector sets of the suspicious target areas Difference; uses the sum of squares of the generalized eigenvalues of the covariance matrices to measure the distance difference between the covariance matrices. 8. The method according to claim 6, characterized in that, the distance difference between the suspicious target areas is the difference between the covariance matrices of the multi-scale concentric ring cluster direction gradient eigenvector sets of the suspicious target areas. The difference is used to measure the distance difference; the sum of squares of the generalized eigenvalues of the covariance matrix is used to measure the distance difference between the covariance matrices. 9. The method according to claim 6, wherein the on-orbit identification of the time-sensitive target is to project the space-time trajectory change curve under polar coordinates, and compare the speed changes and direction changes of non-time-sensitive targets in different phases. Separated to obtain the curve of the direction change of the space-time trajectory, which is used to better describe the anomaly of the space-time trajectory. 10. The method according to claim 1, wherein the step of incrementally updating the training image on the track of the structure dictionary comprises the following steps: Step S41: When there are many training samples, in order to meet the timeliness requirements of on-orbit processing, the detected on-orbit time-sensitive targets and multi-scale concentric circular cluster direction gradient features of suspicious targets are used as new training samples; Step S42: According to the current image and the training image of the on-orbit incremental update of the structure dictionary obtained by the current time-sensitive target, use the new training sample to update the on-orbit incremental update of the structure dictionary at the previous moment.