Satellite video moving target detection method based on low-rank sparse matrix decomposition

Abstract

The invention discloses a satellite video moving target detection method based on low-rank sparse matrix decomposition, and relates to the technical field of data processing. According to the scheme, the influence of the dynamic change background in the video on moving target detection is considered, and the singular value part and the minimum norm are used for representing the low-rank property of the dynamic change background, so that the background image under dynamic change is more accurately recovered; on the basis, a non-convex # imgabs0 # norm is utilized to constrain a sparse foreground target so as to enhance the robustness to noise, and meanwhile, a total variation regular term is utilized to suppress dynamic background pixels so as to construct a moving target detection model, so that the background and the target can be better distinguished, and the accuracy of target image detection in satellite video image data is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method for detecting moving targets in satellite videos based on low-rank sparse matrix decomposition. Background Art

[0002] Moving target detection for video satellite data aims to separate the target from the complex background. Since the background in satellite video data usually changes slowly, the background area usually has a certain degree of autocorrelation and continuity, and the background is considered to have low rank. Moving targets occupy a very small part of satellite video images, and the targets are considered to be sparse.

[0003] Therefore, the detection problem of moving targets in satellite videos can be transformed into a low-rank sparse reconstruction problem. Although many satellite video moving target detection methods based on the classic principal component analysis method have achieved good detection performance, they cannot guarantee high detection accuracy in dynamically changing backgrounds, because the dynamically changing background is similar to the target change in the time domain.

[0004] In view of the above, how to solve the problem that the current satellite video moving target detection cannot guarantee high detection accuracy under a dynamically changing background is an urgent problem to be solved by technicians in this field. Summary of the invention

[0005] The purpose of this application is to provide a satellite video moving target detection method based on low-rank sparse matrix decomposition to solve the problem that the current satellite video moving target detection cannot guarantee high detection accuracy under a dynamically changing background.

[0006] In order to solve the above technical problems, the present application provides a satellite video moving target detection method based on low-rank sparse matrix decomposition, comprising:

[0007] Acquire satellite video image data, and construct a satellite video matrix data model based on the satellite video image data;

[0008] Constructing a total variation regularization term model according to the satellite video image data;

[0009] According to the satellite video matrix data model, a background model based on the singular value part and the minimized norm and a background model based on the non-convex Norm foreground model;

[0010] According to the background model based on the singular value part and the minimized norm, the non-convex The foreground model of the norm and the total variation regularization term model are used to construct a moving target detection algorithm model;

[0011] Solve the moving target detection algorithm model to obtain the target image of the satellite video image data.

[0012] On the one hand, constructing a satellite video matrix data model according to the satellite video image data includes:

[0013] Vectorizing the satellite video image data;

[0014] The vectorized satellite video image data is decomposed into a background matrix and a foreground matrix to obtain the satellite video matrix data model.

[0015] On the other hand, a total variation regularization term model is constructed according to the satellite video image data, including:

[0016] Obtaining differential values ​​of pixel values ​​in the satellite video image data in the horizontal direction, the vertical direction and the time direction;

[0017] Vectorizing each of the differential values ​​to obtain a spatial image long-dimension differential matrix, a spatial image wide-dimension differential matrix, and a time-dimension differential matrix;

[0018] The total variation regularization term model is constructed according to the spatial image long-dimensional difference matrix, the spatial image wide-dimensional difference matrix and the time-dimensional difference matrix.

[0019] On the other hand, the formula of the moving target detection algorithm model includes:

[0020] ;

[0021] in, is the background model based on the singular value part and the minimization norm, Based on the non-convex norm of the foreground model, is the total variation regularization model, is the satellite video matrix data model, is the background matrix, is the foreground matrix, is the moving background and dynamically changing background matrix caused by local misalignment and illumination changes, is the number of singular values, and are the hyperparameters of the constrained foreground regularization term and the total variation regularization term, respectively.

[0022] On the other hand, solving the moving target detection algorithm model to obtain the target image of the satellite video image data includes:

[0023] Constructing a Lagrangian function according to the moving target detection algorithm model;

[0024] The Lagrangian function is decomposed by an alternating direction multiplier method to obtain the target image.

[0025] On the other hand, the decomposition of the Lagrangian function by the alternating direction multiplier method comprises:

[0026] Update the background matrix based on the Lagrangian function;

[0027] updating the foreground matrix based on the Lagrangian function;

[0028] Update the motion background and dynamically changing background matrices based on the Lagrangian function;

[0029] Update the Lagrangian operator of the Lagrangian function according to the updated background matrix, the updated foreground matrix, and the updated moving background and dynamically changing background matrices;

[0030] Determine whether the preset number of iterations is reached or the preset conditions are met;

[0031] If yes, output the updated background matrix, the updated foreground matrix, and the updated moving background and dynamically changing background matrices;

[0032] If not, the process returns to the step of updating the background matrix based on the Lagrangian function.

[0033] On the other hand, after obtaining the target image, the method further includes:

[0034] When there is a single moving target in the target image, the moving target is continuously tracked and the corresponding movement trajectory and state change are recorded;

[0035] When there are multiple moving targets in the target image, each of the moving targets is tracked simultaneously, and the relative positions and movement relationships between the moving targets are analyzed.

[0036] On the other hand, after obtaining the target image, the method further includes:

[0037] When there is a moving target in the target image, continuously tracking the moving target;

[0038] Determine whether the motion trajectory and state of the moving target have changed according to a preset period;

[0039] If so, output warning information indicating that the moving target is abnormal.

[0040] On the other hand, after obtaining the target image, the method further includes:

[0041] When there is a moving target in the target image, acquiring the appearance features of the moving target;

[0042] The type and attribute of the moving object are determined according to the appearance features.

[0043] On the other hand, after determining the type and attribute of the moving object according to the appearance feature,

[0044] The information of each type of the sports target is stored in a database in the form of reports and / or charts.

[0045] The satellite video moving target detection method based on low-rank sparse matrix decomposition provided in the present application obtains satellite video image data and constructs a satellite video matrix data model according to the satellite video image data; constructs a total variation regularization term model according to the satellite video image data; constructs a background model based on the singular value part and the minimized norm and a background model based on the non-convex norm foreground model; based on the background model based on the singular value part and the minimization norm, based on the non-convex The foreground model of the norm and the total variation regularization term model are used to construct the moving target detection algorithm model; the moving target detection algorithm model is solved to obtain the target image of the satellite video image data. It can be seen that this scheme takes into account the impact of the dynamically changing background in the video on the moving target detection, and uses the singular value part and the minimized norm to characterize the low rank of the dynamically changing background, so as to more accurately restore the background image under dynamic changes; on this basis, the non-convex The norm constrains the sparse foreground targets to enhance the robustness to noise. At the same time, the total variation regularization term is used to suppress dynamic background pixels. Then, a moving target detection model is constructed, which can better distinguish between background and target and improve the accuracy of target image detection in satellite video image data. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 A flowchart of a satellite video moving target detection method based on low-rank sparse matrix decomposition provided in an embodiment of the present application;

[0048] Figure 2 A first schematic diagram of minimizing the nuclear norm provided in an embodiment of the present application;

[0049] Figure 3A second schematic diagram of minimizing the nuclear norm provided in an embodiment of the present application;

[0050] Figure 4 Provided for the embodiments of this application , and One-dimensional illustration of the norm. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0052] The core of this application is to provide a satellite video moving target detection method based on low-rank sparse matrix decomposition to solve the problem that the current satellite video moving target detection cannot guarantee high detection accuracy under a dynamically changing background.

[0053] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0054] At present, many satellite video moving target detection methods based on low-rank sparse matrix decomposition based on the classic principal component analysis method have achieved good detection performance, but they cannot guarantee high detection accuracy under dynamically changing backgrounds, because the dynamically changing background is similar to the target change in the time domain. Therefore, in order to solve the above problems, the present application provides a satellite video moving target detection method based on low-rank sparse matrix decomposition.

[0055] Figure 1 The present invention provides a flowchart of a method for detecting moving targets in satellite video based on low-rank sparse matrix decomposition. Figure 1 As shown, the method includes:

[0056] S10: Acquire satellite video image data, and construct a satellite video matrix data model according to the satellite video image data.

[0057] In order to realize the moving target recognition of satellite images, it is first necessary to obtain satellite video image data. It should be noted that the satellite video image data is a video sequence. At the same time, a satellite video matrix data model is constructed based on the satellite video image data. Among them, the satellite video matrix data model is to convert the satellite video image data into a matrix data. In this embodiment, there is no restriction on the construction process of the satellite video matrix data model.

[0058] S11: Construct a total variation regularization term model based on satellite video image data.

[0059] Furthermore, a total variation regularization term model is constructed based on the satellite video image data. The total variation (TV) regularization term is a regularization method commonly used in image processing and optimization problems. Its main function is to maintain the smoothness of the image, suppress noise and maintain the edge information of the image by minimizing the sum of the differences between adjacent pixels in the image. In this embodiment, there is no restriction on the construction process of the total variation regularization term model.

[0060] S12: Based on the satellite video matrix data model, a background model based on the singular value part and the minimized norm is constructed and a non-convex Norm foreground model.

[0061] When there is a large amount of measurement noise interference in the image, the recovery performance of the Nuclear Norm (NN) convex relaxation algorithm will drop sharply, and the recovered solution will seriously deviate from the original solution of the rank minimum optimization problem, and the estimated rank will have an over-shrink phenomenon. Figure 2 This is a first schematic diagram of minimizing the nuclear norm provided in an embodiment of the present application. Figure 2 As shown in the figure, the horizontal axis X1 represents the ground truth value, and the vertical axis X2 represents the estimated value. The green represents the true value, and the true value subspace (green) is a one-dimensional straight line with sparse outliers and noise. Figure 2 , the estimated subspace is biased towards the estimated axis which has a smaller nuclear norm but the second singular value is larger than the ground truth coordinate. Figure 3 A second schematic diagram of minimizing the nuclear norm provided in an embodiment of the present application. Figure 3 In , the horizontal axis X1 represents the ground truth value, and the vertical axis X2 represents the estimated value. Some internal values ​​located on the true subspace are regarded as outliers to obtain smaller singular values.

[0062] In order to more reasonably utilize the prior knowledge represented by different singular values ​​and treat each rank component differently to improve the accuracy of low-rank recovery, instead of using the same weights as in Nuclear Norm Minimization (NNM), this scheme adopts the Partial Sum Minimization of Singular Values ​​(PSSV), which discards the smallest - singular values, keeping the largest unchanged, among which is the number of singular values, is the rank of the matrix. The PSSV norm is defined as:

[0063] ;

[0064] in, Background matrix The target rank, It is before indivual The singular values ​​of Background matrix of The largest singular value, for and The upper limit of the ratio is also the rank of the matrix, and The background matrix is The number of rows and columns.

[0065] The PSSV norm can minimize the variance of the residual rank while maintaining large singular values, thereby retaining the main components of the data, which is conducive to improving the accuracy of dynamic satellite video background image restoration. Compared with the traditional nuclear norm minimization method, the PSSV norm minimization method takes some singular values, making the soft threshold more reasonable. Therefore, the difference between the target motion and the background motion becomes obvious and measurable, making it easier for the detection model to remove the background.

[0066] On the other hand, although Norm Replacement The norm can better solve sparse terms, however The norm may cause the result to fall into the local minimum instead of the global minimum, which will affect the effect of sparse constraints. In addition, it does not consider the spatial correlation of foreground pixels. In order to improve the foreground smoothness caused by the constraint, we use the performance better than of norm to enforce the sparsity constraint. Figure 4 Provided for the embodiments of this application , and One-dimensional diagram of the norm. Figure 4 As shown in the figure, the horizontal axis X represents the absolute value of the one-dimensional variable, and the vertical axis Y represents the norm value. Norm and A one-dimensional example of the norm. It can be seen from the figure Norm Ratio The curve of the norm is closer to Norm. The norm is proved to not only fully utilize spatial correlation and feature similarity to restore sparse targets and obtain a sparser solution, but also use fewer iterations to reach convergence while reducing the amount of calculation. Therefore, the foreground matrix is ​​defined as:

[0067] ;

[0068] in, is the foreground matrix, For the Line The pixel value of the column.

[0069] In summary, this application specifically constructs a background model based on the singular value part and the minimized norm and a background model based on the non-convex The foreground model based on the norm can more accurately restore the background image under dynamic changes, and can constrain the sparse foreground targets to enhance the robustness to noise.

[0070] S13: Based on the background model based on the singular value part and the minimized norm, based on the non-convex The foreground model and total variation regular term model of the norm are used to build a moving target detection algorithm model.

[0071] S14: Solve the moving target detection algorithm model to obtain the target image of the satellite video image data.

[0072] Furthermore, by integrating the background model, the foreground model and the total variation regularization term model, a moving target detection algorithm model is constructed. In this embodiment, there is no restriction on the expression of the moving target detection algorithm model, which depends on the specific implementation situation.

[0073] Finally, the moving target detection algorithm model is solved to obtain the target image of the satellite video image data, thus realizing the moving target recognition of satellite images.

[0074] In this embodiment, satellite video image data is obtained, and a satellite video matrix data model is constructed according to the satellite video image data; a total variation regularization term model is constructed according to the satellite video image data; a background model based on the singular value part and the minimized norm and a background model based on the non-convex norm foreground model; based on the background model based on the singular value part and the minimization norm, based on the non-convex The foreground model of the norm and the total variation regularization term model are used to construct the moving target detection algorithm model; the moving target detection algorithm model is solved to obtain the target image of the satellite video image data. It can be seen that this scheme takes into account the impact of the dynamically changing background in the video on the moving target detection, and uses the singular value part and the minimized norm to characterize the low rank of the dynamically changing background, so as to more accurately restore the background image under dynamic changes; on this basis, the non-convex The norm constrains the sparse foreground targets to enhance the robustness to noise. At the same time, the total variation regularization term is used to suppress dynamic background pixels. Then, a moving target detection model is constructed, which can better distinguish between background and target and improve the accuracy of target image detection in satellite video image data.

[0075] On the basis of the above embodiments, in some embodiments, a satellite video matrix data model is constructed according to satellite video image data, including:

[0076] S101: vectorizing satellite video image data;

[0077] S102: Decomposing the vectorized satellite video image data into a background matrix and a foreground matrix to obtain a satellite video matrix data model.

[0078] Specifically, assuming that the given satellite image video sequence is , which is vectorized into .in, , and are the height, width and number of frames of the video image respectively. In general, the vectorized video sequence can be decomposed into foreground and background. Therefore, the satellite video matrix data model is as follows:

[0079] ;

[0080] in, is the satellite video matrix data model, is the background matrix; is the foreground matrix, which contains the moving target.

[0081] In this way, the construction of the satellite video matrix data model is realized.

[0082] On the basis of the above embodiments, in some embodiments, a total variation regularization term model is constructed according to satellite video image data, including:

[0083] S111: Obtaining differential values ​​of pixel values ​​in the satellite video image data in the horizontal direction, the vertical direction and the time direction;

[0084] S112: performing vectorization processing on each difference value to obtain a spatial image long dimension difference matrix, a spatial image wide dimension difference matrix and a time dimension difference matrix;

[0085] S113: Construct a total variation regularization term model according to the spatial image long dimension difference matrix, the spatial image wide dimension difference matrix and the time dimension difference matrix.

[0086] It can be understood that the derivative can be used to measure the sensitivity of a function to change. For discrete functions, the differential operator is an approximation of the derivative. Therefore, in order to construct the total variation regularization term model, it is necessary to calculate the differential values ​​of the pixel values ​​in the video sequence image in the horizontal, vertical and time directions. Specifically, the differential values ​​of the pixel values ​​in the satellite video image data in the horizontal, vertical and time directions are obtained:

[0087] ;

[0088] in, They are respectively the length, width and time frame number of the video data image.

[0089] Then these results are matrix-vectorized to obtain the long-dimensional difference matrix of the spatial image , spatial image wide dimensional difference matrix and the time dimension difference matrix .Will , and Merge into one matrix to get the total difference matrix , the total variation regularization model is as follows:

[0090] ;

[0091] in, , and Represents the intensity changes on the horizontal, vertical and time axes respectively.

[0092] In this way, the construction of the total variation regularization term model is realized.

[0093] On the basis of the above embodiment, in this embodiment, the formula of the moving target detection algorithm model includes:

[0094] ;

[0095] in, is a background model based on the singular value part and the minimization norm, Based on non-convex norm of the foreground model, is the total variation regularization model, is the satellite video matrix data model, is the background matrix, is the foreground matrix, is the moving background and dynamically changing background matrix caused by local misalignment and illumination changes, is the number of singular values, and are the hyperparameters of the constrained foreground regularization term and the total variation regularization term, respectively. It should be noted that when The value is fixed When it becomes small enough, the model will degenerate into a special case, namely, moving target detection in a static background.

[0096] Based on the above embodiments, in some embodiments, solving the moving target detection algorithm model to obtain the target image of the satellite video image data includes:

[0097] S141: constructing a Lagrangian function according to a moving target detection algorithm model;

[0098] S142: Decomposing the Lagrangian function by an alternating direction multiplier method to obtain a target image.

[0099] In order to solve the moving target detection algorithm model, we first need to establish the Lagrangian function:

[0100] ;

[0101] in, is the Lagrange multiplier matrix, is a positive penalty scalar, represents the inner product of the matrix, is the Frobenius norm.

[0102] Since it is difficult to optimize all the above variables at the same time, in this embodiment, the Lagrangian function is decomposed by the alternating direction method of multipliers (ADMM), which is specifically decomposed into the following sub-problems:

[0103] (1) Background matrix Updates;

[0104] Specifically, the background matrix is ​​updated based on the Lagrangian function. It should be noted that the update , while keeping other variables unchanged.

[0105] ;

[0106] therefore, The solution formula is:

[0107] ;

[0108] in, is the partial singular value threshold operator (PSVT), k is the number of iterations. According to the relevant public theories, we can get:

[0109] ;

[0110] ;

[0111] in, is a soft threshold operator. It can be solved using Singular Value Decomposition (SVD), and can be considered as the sum of two matrices, namely:

[0112] ;

[0113] in, , , , , , , , and are all singular vector matrices.

[0114] (2) Foreground Matrix Updates;

[0115] Specifically, the foreground matrix is ​​updated based on the Lagrangian function. It should be noted that the update , but also keep other variables unchanged.

[0116] ;

[0117] It can be understood that the above formula is a typical Regular term constraint problem. Therefore, the optimal solution can be obtained through element-level contraction operation:

[0118] ;

[0119] in, is the element-wise contraction operator.

[0120] (3) Moving background and dynamically changing background matrix Updates;

[0121] Specifically, the moving background and dynamically changing background matrices are updated based on the Lagrangian function. During the optimization process, The update can be done through the following steps:

[0122] ;

[0123] Using TV norm regularization, we can get the solution:

[0124] ;

[0125] in, It is the TV operator.

[0126] Furthermore, according to the updated background matrix, the updated foreground matrix, and the updated moving background and dynamically changing background matrices, the Lagrangian operator of the Lagrangian function is updated:

[0127] ;

[0128] During the optimization process, all variables are iteratively updated, and it is determined whether the preset number of iterations is reached or the preset conditions are met; if so, the updated background matrix, updated foreground matrix, and updated motion background and dynamically changing background matrices are output; if not, it returns to the step of updating the background matrix based on the Lagrangian function.

[0129] It should be noted that the pre-conditions are:

[0130] ;

[0131] in, .

[0132] In summary, the moving target detection algorithm model is solved and the target image is obtained.

[0133] Based on the above embodiment, in some embodiments, after obtaining the target image, the method further includes:

[0134] S15: When there is a single moving target in the target image, the moving target is continuously tracked and the corresponding movement trajectory and state change are recorded;

[0135] S16: When there are multiple moving targets in the target image, the moving targets are tracked simultaneously, and the relative positions and movement relationships between the moving targets are analyzed.

[0136] After detecting the target image in the satellite video image data, target tracking can be further performed, which can provide in-depth insights into the target's motion patterns and behavioral characteristics. Target tracking can be divided into two main forms: single target tracking and multi-target tracking.

[0137] Single target tracking focuses on continuous monitoring of a single detected moving target. By recording the target's movement trajectory and state changes, such as position, speed, and acceleration, single target tracking can reveal the individual behavioral characteristics of the target. The advantage of this tracking method is that its precise focus on a specific target enables detailed analysis of the target's short-term and long-term movement trends. For example, in wildlife conservation, single target tracking can help researchers understand the migration routes and living habits of animals.

[0138] Multi-target tracking involves tracking multiple moving targets in a video at the same time and analyzing their relative positions and motion relationships. This tracking method can provide a broader field of view and reveal the behavior patterns and interaction mechanisms of the target group. The advantage of multi-target tracking is its adaptability to complex scenes and its ability to handle occlusion, intersection, and separation between targets. In traffic monitoring, multi-target tracking can be used to count traffic flow, detect traffic incidents, and provide data support for traffic planning and management. In addition, in military and security applications, multi-target tracking can help monitoring personnel detect and deal with potential threats in a timely manner.

[0139] In this embodiment, target tracking, as a subsequent step of moving target detection, provides strong support for research and practice in various fields by recording and analyzing the target's motion trajectory and state changes. Whether it is precise focus on a single target or group analysis of multiple targets, target tracking can reveal the target's behavioral characteristics and interaction mechanism, thereby providing decision support and scientific basis for related applications.

[0140] Based on the above embodiment, in some embodiments, after obtaining the target image, the method further includes:

[0141] S17: When there is a moving target in the target image, the moving target is continuously tracked.

[0142] S18: Determine whether the motion trajectory and state of the moving target have changed according to a preset period; if so, proceed to step S19.

[0143] S19: Outputting warning information indicating that there is an abnormality in the moving target.

[0144] After detecting moving targets in satellite video image data, anomaly detection and early warning are important applications in moving target tracking and analysis. Specifically, when there is a moving target in the target image, the moving target is continuously tracked. By monitoring the target's motion trajectory and behavior pattern, deviations from normal patterns, such as sudden acceleration and deviation from the planned route, are identified. These abnormal behaviors may indicate potential problems or threats that require further attention and processing. Anomaly detection algorithms are usually based on statistical methods or machine learning techniques, which can automatically identify unusual patterns from large amounts of data.

[0145] Furthermore, the motion trajectory and state of the moving target are determined according to the preset period. If so, based on the results of anomaly detection, the security warning system can issue an alarm in time to remind relevant personnel to pay attention or take necessary measures. This warning mechanism is crucial in many fields, including traffic safety, border monitoring, and environmental protection. By combining anomaly detection with real-time warning, the response speed and processing efficiency of potential risks can be greatly improved.

[0146] Based on the above embodiment, in some embodiments, after obtaining the target image, the method further includes:

[0147] S20: When there is a moving target in the target image, the appearance features of the moving target are obtained.

[0148] S21: Determine the type and attribute of the moving target according to the appearance features.

[0149] After detecting moving targets in satellite video image data, target classification and recognition are key follow-up steps. Target type recognition divides targets into different categories, such as airplanes, cars, ships, etc., by analyzing the target's appearance characteristics, movement patterns and other information. This recognition not only relies on the static characteristics of the target, such as shape and color, but also considers the dynamic behavior of the target, such as movement speed and trajectory. Accurate target type recognition is crucial for many applications.

[0150] Furthermore, target attribute analysis digs deeper into more details of the target, including attributes such as size, shape, and color. These attributes provide additional information about the target and can be used for more sophisticated analysis and decision-making. For example, in environmental monitoring, analyzing the size and shape of wild animals can help determine the species, while color information may be used to identify specific individuals. Target attribute analysis enriches the understanding of the target and provides more comprehensive support for subsequent data mining and applications.

[0151] In addition, after determining the type of moving target based on appearance characteristics, the information of each type of moving target can be stored in the database in the form of reports and / or charts, which can not only more comprehensively display the trends and patterns behind the data, but also be stored as historical data to facilitate subsequent possible data analysis.

[0152] The above is a detailed introduction to a satellite video moving target detection method based on low-rank sparse matrix decomposition provided by the present application. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the present application.

[0153] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A satellite video moving target detection method based on low-rank sparse matrix decomposition, characterized in that: include: Acquire satellite video image data, and construct a satellite video matrix data model based on the satellite video image data; Constructing a total variation regularization term model according to the satellite video image data; According to the satellite video matrix data model, a background model based on the singular value part and the minimized norm and a background model based on the non-convex Norm foreground model; According to the background model based on the singular value part and the minimized norm, the non-convex The foreground model of the norm and the total variation regularization term model are used to construct a moving target detection algorithm model; Solve the moving target detection algorithm model to obtain the target image of the satellite video image data.

2. The satellite video moving target detection method based on low-rank sparse matrix decomposition according to claim 1 is characterized in that: Constructing a satellite video matrix data model according to the satellite video image data includes: Vectorizing the satellite video image data; The vectorized satellite video image data is decomposed into a background matrix and a foreground matrix to obtain the satellite video matrix data model.

3. The satellite video moving target detection method based on low-rank sparse matrix decomposition according to claim 1 is characterized in that: A total variation regularization term model is constructed according to the satellite video image data, including: Obtaining differential values ​​of pixel values ​​in the satellite video image data in the horizontal direction, the vertical direction and the time direction; Vectorizing each of the differential values ​​to obtain a spatial image long-dimension differential matrix, a spatial image wide-dimension differential matrix, and a time-dimension differential matrix; The total variation regularization term model is constructed according to the spatial image long-dimensional difference matrix, the spatial image wide-dimensional difference matrix and the time-dimensional difference matrix.

4. The satellite video moving target detection method based on low-rank sparse matrix decomposition according to claim 1, characterized in that: The formula of the moving target detection algorithm model includes: ； in, is the background model based on the singular value part and the minimization norm, Based on the non-convex norm of the foreground model, is the total variation regularization model, is the satellite video matrix data model, is the background matrix, is the foreground matrix, is the moving background and dynamically changing background matrix caused by local misalignment and illumination changes, is the number of singular values, and are the hyperparameters of the constrained foreground regularization term and the total variation regularization term, respectively.

5. The satellite video moving target detection method based on low-rank sparse matrix decomposition according to claim 4 is characterized in that: Solving the moving target detection algorithm model to obtain the target image of the satellite video image data includes: Constructing a Lagrangian function according to the moving target detection algorithm model; The Lagrangian function is decomposed by an alternating direction multiplier method to obtain the target image.

6. The satellite video moving target detection method based on low-rank sparse matrix decomposition according to claim 5 is characterized in that: The decomposition of the Lagrangian function by the alternating direction multiplier method comprises: Update the background matrix based on the Lagrangian function; updating the foreground matrix based on the Lagrangian function; Update the motion background and dynamically changing background matrices based on the Lagrangian function; Update the Lagrangian operator of the Lagrangian function according to the updated background matrix, the updated foreground matrix, and the updated moving background and dynamically changing background matrices; Determine whether the preset number of iterations is reached or the preset conditions are met; If yes, output the updated background matrix, the updated foreground matrix, and the updated moving background and dynamically changing background matrices; If not, the process returns to the step of updating the background matrix based on the Lagrangian function.

7. The method for satellite video moving target detection based on low-rank sparse matrix decomposition according to any one of claims 1 to 6, characterized in that: After obtaining the target image, the method further includes: When there is a single moving target in the target image, the moving target is continuously tracked and the corresponding movement trajectory and state change are recorded; When there are multiple moving targets in the target image, each of the moving targets is tracked simultaneously, and the relative positions and movement relationships between the moving targets are analyzed.

8. The method for satellite video moving target detection based on low-rank sparse matrix decomposition according to any one of claims 1 to 6, characterized in that: After obtaining the target image, the method further includes: When there is a moving target in the target image, continuously tracking the moving target; Determine whether the motion trajectory and state of the moving target have changed according to a preset period; If so, output warning information indicating that the moving target is abnormal.

9. The satellite video moving target detection method based on low-rank sparse matrix decomposition according to any one of claims 1 to 6, characterized in that: After obtaining the target image, the method further includes: When there is a moving target in the target image, acquiring the appearance features of the moving target; The type and attribute of the moving object are determined according to the appearance features.

10. The satellite video moving target detection method based on low-rank sparse matrix decomposition according to claim 9, characterized in that: After determining the type and attribute of the moving object according to the appearance features, The information of each type of the sports target is stored in a database in the form of reports and / or charts.

Images