Moving target detection method under dynamic background based on non-convex space-time low-rank robust principal component analysis

Abstract

The invention discloses a method for detecting a moving target under a dynamic background based on non-convex space-time low-rank robust principal component analysis. The method comprises the following steps: acquiring video information of the moving target; converting the video information of the moving target into a three-dimensional observation matrix; decomposing the observation video matrix O into a low-rank static background matrix B, a sparse moving target matrix S and a dynamic background matrix E according to the background of the moving target and the characteristics of the moving target; constraining the decomposed matrix, and constructing a moving target detection model under the dynamic background based on non-convex space-time low-rank robust principal component analysis; converting a moving target detection model of the non-convex space-time low-rank robust principal component analysis under the dynamic background into an augmented Lagrange function, and solving a moving target detection problem of the model under the dynamic background by using an alternating multiplier method; obtaining an extracted moving target according to a sub-problem solving result, and performing binarization masking on the extracted moving target to realize detection of the moving target; according to the invention, the moving target can be completely detected in a dynamic background.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis. Background Art

[0002] The basic task of moving target detection under dynamic background is to extract the foreground moving target from the complex dynamic background. Moving target detection under dynamic background is a complex and challenging task, because the dynamic background (such as fluctuating water surface, shaking leaves, rain, snow, etc.) will confuse the moving target and affect the detection results. Moving target detection methods can be divided into background subtraction method, frame difference method, optical flow method, deep learning method and RPCA method. The frame difference method identifies moving objects by calculating the absolute difference between consecutive frames and applying a threshold, but improper threshold selection will cause "holes" and "ghosting" phenomena in the detection results. The background subtraction method compares the current frame with the background image, but it will face interference when dealing with dynamic backgrounds, resulting in inaccurate background modeling and misidentifying the dynamic background as a moving target. The optical flow method detects moving objects and background areas separately according to different optical flow fields, but the calculation is complex. Deep learning methods require a large number of data sets.

[0003] In recent years, moving target detection based on RPCA has been widely used. Traditional RPCA believes that the observed image data can be divided into background data and foreground data. Static background has low-rank characteristics, while moving targets usually only occupy a small part of the image pixels and can be regarded as sparse features. Traditional RPCA uses rank function constraints for low-rank backgrounds and L1 norm constraints for foreground targets. Although the nuclear norm and L1 norm have significant advantages in background modeling and moving target detection, this method has some limitations in the actual video processing. Real scenes have a certain degree of complexity. There are dynamic water ripples, leaf jitters, and shadow and illumination changes in the background, which cause dynamic background pixels to be mistakenly identified as moving targets, resulting in a decrease in the integrity of the moving target and the accuracy of detection. Therefore, it is of great research significance to further study the inherent characteristics of the foreground and background parts and introduce more prior knowledge to enhance the modeling of background and moving targets. Summary of the invention

[0004] In order to solve the existing problems of motion target detection under dynamic background interference, the technical solution adopted by the present invention is: a motion target detection method under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis, comprising the following steps:

[0005] Obtain video information of moving targets;

[0006] Converting the video information of the moving target into a three-dimensional observation matrix;

[0007] According to the background and characteristics of the moving target, the observation video matrix O is decomposed into a low-rank static background matrix B, a sparse moving target matrix S and a dynamic background matrix E;

[0008] Constrain the decomposed matrix and construct a moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis;

[0009] The moving target detection model under dynamic background of non-convex spatiotemporal low-rank robust principal component analysis is transformed into augmented Lagrangian function, and the moving target detection problem under dynamic background of the model is solved by using alternating multiplier method.

[0010] The extracted moving target is obtained according to the result of solving the sub-problem, and the extracted moving target is subjected to a binary mask to realize the detection of the moving target.

[0011] Further: the process of constraining the decomposed matrix is ​​as follows:

[0012] The γ norm is used to approximate the low-rank static background matrix to effectively constrain the static background;

[0013] The regularization term L1 is used to constrain the dynamic background matrix E;

[0014] Structured sparsity and 3D total variation priors are used to constrain the sparse motion target matrix S.

[0015] Further: the γ norm |B|| γ The expression is as follows:

[0016]

[0017] Among them, σ i (B) represents the i-th singular value of B, and γ is a parameter between 0 and 1.

[0018] Further: the structured sparsity induced norm Ω(S) is expressed as:

[0019]

[0020] Where G is a set of predefined group distributions, where the distribution of each group is set to g; Represents a sparse vector s formed by the elements in g as indexes j A subset of j is the j-th column of the matrix S; ||*|| ∞ represents the infinite norm; b is the maximum number of columns.

[0021] Furthermore: the 3D total variation norm is expressed as:

[0022]

[0023] Where f represents the vectorization operator of F, V h ,V v ,V t Represent the operators of the horizontal, vertical and time directions of F respectively.

[0024] Further: the moving target detection model is:

[0025] min||B|| γ +λ1Ω(S)+λ2||S|| 3D-TV +λ3||E||1

[0026] stO=B+S+E

[0027] Among them, O is the observation video matrix, B is the low-rank static background matrix, S is the sparse moving target matrix, and E is the dynamic background matrix. In order to simplify the calculation, the variables of the 3D total variation term are replaced by F, as follows:

[0028]

[0029] Further: the process of converting the moving target detection model under dynamic background of non-convex spatiotemporal low-rank robust principal component analysis into an augmented Lagrangian function and solving the moving target detection sub-problem under dynamic background of the model by using the alternating multiplier method is as follows:

[0030] 1) According to the moving target detection model, the corresponding augmented Lagrangian function is determined, which is expressed as:

[0031]

[0032] Among them, X, Z are Lagrange multipliers, μ is the penalty factor, <,> is the inner product of the matrix, is the F-norm;

[0033] 2) Set initial parameters:

[0034] B 0 =O,S 0 =0,E 0 =0,F 0 = 0, draw a Lagrange multiplier Z 0 =0, ADMM parameter ρ = 5.5, maximum penalty factor μ max =μ 0 ×10 7 , relative error tol = 10 -5 , iterative parameter k = 0, initial value of Lagrange multiplier Initial value of penalty factor Regularization parameters λ1>0,λ2>0,λ3>0;

[0035] 3) Update variable B;

[0036] B k+1 =Udiag{σ *}V T

[0037] Among them, U, V are orthogonal matrices, and the diagonal elements σ * is the singular value of B, non-negative and arranged in descending order. B is the static background matrix, and the singular value threshold method is used to solve the variable B; B k+1 is the static background matrix at K+1 step length;

[0038] 4) Update variable S;

[0039]

[0040] Where prox represents the nearest neighbor operator, and the quadratic minimum optical flow method is used to solve the S subproblem; S is the moving target matrix, S k+1 The sparse moving target matrix when the step length is K+1

[0041] 5) Update variable E;

[0042]

[0043] in, represents the soft threshold shrinkage operator; E k+1 is the dynamic background matrix at K+1 step length;

[0044] 7) Update variable F;

[0045]

[0046] In order to simplify the operation, a new variable is introduced K = Vf;

[0047]

[0048] Among them: F k+1 represents the auxiliary variable at k+1 step length, μ k is the penalty factor for the k-step, Z k is the k-step Lagrange multiplier;

[0049] The transformed augmented Lagrangian function is:

[0050]

[0051] Among them, ζ is the penalty parameter, T is the Lagrange multiplier, and the variables t and k are the number of inner and outer loops respectively;

[0052]

[0053] Let q = {1} and update the auxiliary variable K as follows:

[0054]

[0055] Update the parameter T as follows:

[0056] T t+1 =T t +ζ t (K t+1 -Vf t+1 )

[0057] By (μ k I+ζ t V T V) to solve for variable f; reshape variable F into a three-dimensional matrix, and the exact analytical solution of F is as follows:

[0058]

[0059] in, denote inverse Fourier transform and Fourier transform respectively;

[0060] 7) Update the multiplier X as follows;

[0061] X k+1 =X k +μ k (OB k+1 -E k+1 -S k+1 )

[0062] Where: X k+1 represents the update multiplier for k+1 steps, X k represents the update multiplier for k steps,

[0063] 8) Update the multiplier Z as follows;

[0064] Z k+1 =Z k +μ k (S k+1 -F k+1 )

[0065] Where: Z k+1 represents the update multiplier for k+1 steps;

[0066] 9) Update the penalty factor μ as follows:

[0067] μ k+1 =min(ρμ k ,μmax )

[0068] Among them, ρ is the step size; μ max is the maximum value of μ;

[0069] Let k=k+1 and repeat steps 2)-9) until the termination condition is met, the iteration is terminated, and the low-rank matrix B corresponding to the static background in the video information, the sparse matrix S corresponding to the moving target, and the matrix E corresponding to the dynamic background are output.

[0070] Further: The termination condition is:

[0071] A moving target detection device under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis, comprising:

[0072] Acquisition module: used to acquire video information of moving targets;

[0073] Conversion module: used for converting the video information of the moving target into a three-dimensional observation matrix;

[0074] Decomposition module: used to decompose the observation video matrix O into a low-rank static background matrix B, a sparse moving target matrix S and a dynamic background matrix E according to the background and characteristics of the moving target;

[0075] Construction module: used to constrain the decomposed matrix and construct a moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis;

[0076] Solution module: used to transform the dynamic background moving target detection model of non-convex spatiotemporal low-rank robust principal component analysis into an augmented Lagrangian function, and solve the dynamic background moving target detection problem of the model using the alternating multiplier method;

[0077] One-area module: It is used to obtain the extracted moving target according to the result of solving the sub-problem, perform binary masking on the extracted moving target, and realize the detection of the moving target.

[0078] A computer device comprises: a processor and a memory, wherein the memory stores a program module, and the program module runs on the processor to implement any one of the methods described above.

[0079] The present invention provides a method for detecting moving targets under dynamic backgrounds based on non-convex spatiotemporal low-rank robust principal component analysis. The method can remove the influence of point dynamic background on moving target detection under the interference of different complex background environments, such as dynamic backgrounds such as tree shaking, wave fluctuations, and shadow changes, and finally identify the moving target completely and accurately.

[0080] Through comprehensive analysis of evaluation indicators such as F-measure, Recall and Precision, the prior combination designed by the present invention can detect moving targets more completely in a dynamic background. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0082] Figure 1 is a flow chart of the present invention;

[0083] Figure 2 The following is a comparison chart of the simulation test results of the present invention and other models. Among them, (a) is the real color original image; (b) is the true value image; (c) is the result obtained by the model proposed by the present invention; (d) is the detection result of the SGSM algorithm; (e) is the result obtained by the TVRPCA algorithm; (f) is the moving target detection result of the LSD model. DETAILED DESCRIPTION

[0084] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0085] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] A method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis comprises the following steps:

[0087] Obtain video information of moving targets;

[0088] Converting the video information of the moving target into a three-dimensional observation matrix;

[0089] According to the background and characteristics of the moving target, the observation video matrix O is decomposed into a low-rank static background matrix B, a sparse moving target matrix S and a dynamic background matrix E;

[0090] Constrain the decomposed matrix and construct a moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis;

[0091] The moving target detection model under dynamic background of non-convex spatiotemporal low-rank robust principal component analysis is transformed into augmented Lagrangian function, and the moving target detection problem under dynamic background of the model is solved by using alternating multiplier method.

[0092] The extracted moving target is obtained according to the result of solving the sub-problem, and the extracted moving target is subjected to binary masking to achieve the detection of the moving target. By binarizing the moving target detection result, the moving target area is white, that is, the detection of the moving target is achieved;

[0093] Further: the process of constraining the decomposed matrix is ​​as follows:

[0094] The γ norm is used to approximate the low-rank static background matrix to effectively constrain the static background;

[0095] The regularization term L1 is used to constrain the dynamic background matrix E;

[0096] Structured sparsity and 3D total variation priors are used to constrain the sparse motion target matrix S.

[0097] Further: the γ norm ||B|| γ The expression is as follows:

[0098]

[0099] Among them, σ i (B) represents the i-th singular value of B, γ is a parameter between 0 and 1, and the recovery effect of low-rank background is best when it is 0.002;

[0100] The moving target matrix S includes overlapping blocks. For a moving target, the matrix moves piece by piece to detect the target, and there should be overlapping parts during the movement.

[0101] Further: the structured sparsity induced norm Ω(S) is expressed as:

[0102]

[0103] Where G is a set of predefined group distributions, where the distribution of each group is set to g; Represents a sparse vector s formed by the elements in g as indexes jA subset of j is the j-th column of the matrix S; ||*|| ∞ represents the infinite norm; b is the maximum number of columns.

[0104] Furthermore: the 3D total variation norm is expressed as:

[0105]

[0106] Where f represents the vectorization operator of F, V h ,V v ,V t Represent the operators of the horizontal, vertical and time directions of F respectively.

[0107] Further: the moving target detection model is:

[0108] min||B|| γ +λ1Ω(S)+λ2||S|| 3D-TV +λ3||E||1

[0109] stO=B+S+E

[0110] Among them, O is the observation video matrix, B is the low-rank static background matrix, S is the sparse moving target matrix, and E is the dynamic background matrix. In order to simplify the calculation, the variables of the 3D total variation term are replaced by F, as follows:

[0111]

[0112] Further: the process of converting the moving target detection model under dynamic background of non-convex spatiotemporal low-rank robust principal component analysis into an augmented Lagrangian function and solving the moving target detection sub-problem under dynamic background of the model by using the alternating multiplier method is as follows:

[0113] 1) According to the moving target detection model, the corresponding augmented Lagrangian function is determined, which is expressed as:

[0114]

[0115] Among them, X, Z are Lagrange multipliers, μ is the penalty factor, <,> is the inner product of the matrix, is the F-norm;

[0116] 2) Set initial parameters:

[0117] B 0 =O,S 0 =0,E 0 =0,F 0 = 0, draw a Lagrange multiplier Z 0=0, ADMM parameter ρ = 5.5, maximum penalty factor μ max =μ 0 ×10 7 , relative error tol = 10 -5 , iterative parameter k = 0, initial value of Lagrange multiplier Initial value of penalty factor Regularization parameters λ1>0,λ2>0,λ3>0;

[0118] 3) Update variable B;

[0119] B k+1 =Udiag{σ *}V T

[0120] Among them, U, V are orthogonal matrices, and the diagonal elements σ * is the singular value of B, non-negative and arranged in descending order. B is the static background matrix, and the singular value threshold method is used to solve the variable B; B k+1 is the static background matrix at K+1 step length;

[0121] 4) Update variable S;

[0122]

[0123] Where prox represents the nearest neighbor operator, and the quadratic minimum optical flow method is used to solve the S subproblem; S is the moving target matrix, S k+1 is the sparse moving target matrix with K+1 step length;

[0124] 5) Update variable E;

[0125]

[0126] in, represents the soft threshold shrinkage operator; E k+1 is the dynamic background matrix at K+1 step length;

[0127] 8) Update variable F;

[0128]

[0129] In order to simplify the operation, a new variable is introduced K = Vf;

[0130]

[0131] Among them: F k+1 represents the auxiliary variable at k+1 step length, μ k is the penalty factor for the k-step, Z k is the k-step Lagrange multiplier;

[0132] The transformed augmented Lagrangian function is:

[0133]

[0134] Among them, ζ is the penalty parameter, T is the Lagrange multiplier, and the variables t and k are the number of inner and outer loops respectively;

[0135]

[0136] Let q = {1} and update the variable K as follows:

[0137]

[0138] Update the parameter T as follows:

[0139] T t+1 =T t +ζ t (K t+1 -Vf t+1 )

[0140] By (μ k I+ζ t V T V) to solve for variable f; reshape variable F into a three-dimensional matrix, and the exact analytical solution of F is as follows:

[0141]

[0142] in, denote inverse Fourier transform and Fourier transform respectively;

[0143] 7) Update the multiplier X as follows;

[0144] X k+1 =X k +μ k (OB k+1 -E k+1 -S k+1 )

[0145] Where: X k+1 represents the update multiplier for k+1 steps, X k represents the update multiplier for k steps,

[0146] 8) Update the multiplier Z, as follows;

[0147] Z k+1 =Z k +μ k (S k+1 -F k+1 )

[0148] Where: Z k+1 represents the update multiplier for k+1 steps;

[0149] 9) Update the penalty factor μ as follows:

[0150] μ k+1 =min(ρμ k ,μ max )

[0151] Among them, ρ is the step size; μ max is the maximum value of μ;

[0152] Let k=k+1 and repeat steps 2)-9) until the termination condition is met, the iteration is terminated, and the low-rank matrix B corresponding to the static background in the video information, the sparse matrix S corresponding to the moving target, and the matrix E corresponding to the dynamic background are output.

[0153] Further: the termination condition is:

[0154] A moving target detection device under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis, comprising:

[0155] Acquisition module: used to acquire video information of moving targets;

[0156] Conversion module: used for converting the video information of the moving target into a three-dimensional observation matrix;

[0157] Decomposition module: used to decompose the observation video matrix O into a low-rank static background matrix B, a sparse moving target matrix S and a dynamic background matrix E according to the background and characteristics of the moving target;

[0158] Construction module: used to constrain the decomposed matrix and construct a moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis;

[0159] Solution module: used to transform the dynamic background moving target detection model of non-convex spatiotemporal low-rank robust principal component analysis into an augmented Lagrangian function, and solve the dynamic background moving target detection problem of the model using the alternating multiplier method;

[0160] One-area module: It is used to obtain the extracted moving target according to the result of solving the sub-problem, perform binary masking on the extracted moving target, and realize the detection of the moving target.

[0161] A computer device comprises: a processor and a memory, wherein the memory stores a program module, and the program module runs on the processor to implement any one of the methods described above.

[0162] In order to verify the effectiveness of the proposed model, the present invention conducts simulation experiments on several video data segments, and compares the model of the present invention with the SGSM, TVRPCA, and LSD models.

[0163] The present invention adopts "Highway", "Fall", "Canoe" and

[0164] The comparative experiment is conducted on the "Pets2006" data set, and the experimental objects are as follows;

[0165] "Highway" contains 180 frames (480×720);

[0166] "Fall" contains 180 frames (480×720);

[0167] "Canoe" contains 180 frames (240×320);

[0168] "Pets2006" contains 300 frames (576×720);

[0169] The parameters in the simulation test are set as follows (Adjust the value appropriately according to the specific video), using Precision, Recall and F-measure as evaluation indicators;

[0170] Figure 2 The following is a comparison chart of the simulation test results of the present invention and other models. (a) is the real color original image; (b) is the true value image; (c) is the result obtained by the model proposed by the present invention; Figure 2 (d) is the detection result of the SGSM algorithm; (e) is the result obtained by the TVRPCA algorithm; (f) is the moving target detection result of the LSD model;

[0171] In order to verify the performance of the proposed model, a comparative test was conducted on the experimental objects. The experimental results are shown as follows: Figure 2 (c) to (f).

[0172] The “Highway” dataset contains video sequences recorded on a highway with multiple moving objects of different shapes and sizes. The methods from (d) to (f) detect moving vehicles incompletely, and the dynamic foliage background is mistaken for moving targets. In “Highway” (c), the present invention effectively ensures the integrity of the foreground car extraction while greatly reducing the interference of the dynamic background.

[0173] There is a moving object on the road in the "Fall" dataset. The TVRPCA and LSD methods can remove most of the dynamic background, but the detection results are still incomplete. The SGSM method can fully detect the moving vehicle in the dataset, but cannot reduce the interference of swaying leaves. The results of the present invention are shown in "Fall" (c), where there is almost no dynamic background interference in the detection and the outline of the moving vehicle can be fully recognized.

[0174] There are people rowing in the "Canoe" dataset. The SGSM method is easily disturbed by waves when detecting rowers and boats. The sunlight reflected from the water surface may be mistaken for a false target, and the real moving target is surrounded by the dynamic sea surface. Therefore, the detection results of other methods are also incomplete, such as "Canoe" (e) and "Canoe" (f). The present invention distinguishes between dynamic water surfaces and moving target ships by analyzing the orderly changes of moving targets in consecutive frames. Against the dynamic sea surface background of "Canoe" (c), the present invention can correctly and effectively detect moving objects. However, a small part of the hull is still not recognized.

[0175] The "Pets2006" dataset represents an indoor environment with shadow changes. There are light and shadow changes when pedestrians move, so the background can be considered dynamic. In the detection results "Pets2006" (d), (e), and (f), the comparison method can only detect the outline of the moving pedestrians and cannot suppress the interference of dynamic shadows. In the detection result "Pets2006" (c), the present invention can fully detect two targets without the interference of moving shadows, but cannot fully detect the third pedestrian carrying a backpack. The specific reasons are as follows: 3D-TV constraints retain edge features but reduce image gradients, making it difficult to detect relatively small moving people, so the new method still needs further improvement.

[0176] In order to further verify the effectiveness of the present invention, the above methods are compared from a quantitative perspective.

[0177] Table 1 Comparison of evaluation indicators of various methods

[0178]

[0179] When validating the “Fall” dataset, Figure 2 It can be seen from “Fall” (b) and (d) that the present invention is able to remove more dynamic background in motion target detection, so the Precision index values ​​of the present invention are higher than those of other methods, indicating that the present invention can effectively suppress the interference of dynamic background on motion target detection in various situations.

[0180] In the "Highway", "Canoe" and "Pets2006" datasets, the three indicators of the improved method are all high. In particular, in "Highway" and "Canoe", the evaluation indicators Precision of the two scenes are 0.959 and 0.908 respectively. The two indicators Recision and F-measure also have high values, indicating that the present invention is not only efficient and comprehensive in extracting moving targets under dynamic backgrounds, but also can effectively suppress the interference of point-shaped dynamic backgrounds such as trees and waves on target detection.

[0181] In summary, it can be determined that the processing results obtained by the model proposed in the present invention are better than those of the comparison model in terms of both visual effects and evaluation indicators. Therefore, the present invention has a better processing effect in the detection of moving objects in a dynamic background.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A moving target detection method under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis, characterized by: The following steps are involved: Obtain video information of moving targets; Converting the video information of the moving target into a three-dimensional observation matrix; According to the background and characteristics of the moving target, the observation video matrix O is decomposed into a low-rank static background matrix B, a sparse moving target matrix S and a dynamic background matrix E; Constrain the decomposed matrix and construct a moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis; The moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis is transformed into an augmented Lagrangian function, and the moving target detection problem under dynamic background of the model is solved by using the alternating multiplier method. The extracted moving target is obtained according to the result of solving the sub-problem, and the extracted moving target is subjected to a binary mask to realize the detection of the moving target.

2. According to claim 1, a method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis is characterized by: The process of constraining the decomposed matrix is ​​as follows: The γ norm is used to approximate the low-rank static background matrix to effectively constrain the static background; The regularization term L1 is used to constrain the dynamic background matrix E; Structured sparsity and 3D total variation priors are used to constrain the sparse motion target matrix S.

3. According to claim 1, a method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis is characterized by: The gamma norm ||B|| γ The expression is as follows: Among them, σ i (B) represents the i-th singular value of B, and γ is a parameter between 0 and 1.

4. According to claim 1, a method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis is characterized by: The structured sparsity induced norm Ω(S) is expressed as: Where G is a set of predefined group distributions, where the distribution of each group is set to g; Represents a sparse vector s formed by the elements in g as indexes j A subset of j is the j-th column of the matrix S; ||*|| ∞ represents the infinite norm; b is the maximum number of columns.

5. According to claim 1, a method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis is characterized by: The 3D total variation norm, its expression is: Where f represents the vectorization operator of F, V h ,V v ,V t Represent the operators of the horizontal, vertical and time directions of F respectively.

6. The method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis according to claim 1, characterized in that: The moving target detection model is: min||B|| γ +λ1Ω(S)+λ2||S|| 3D-TV +λ3||E||1 stO=B+S+E Among them, O is the observation video matrix, B is the low-rank static background matrix, S is the sparse moving target matrix, and E is the dynamic background matrix. In order to simplify the calculation, the variables of the 3D total variation term are replaced by F, as follows: stO=B+S+E,S=F.

7. The method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis according to claim 1, characterized in that: The process of converting the moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis into an augmented Lagrangian function and solving the moving target detection problem under dynamic background of the model by using the alternating multiplier method is as follows: 1) According to the moving target detection model, the corresponding augmented Lagrangian function is determined, which is expressed as: Among them, X, Z are Lagrange multipliers, μ is the penalty factor, <,> is the inner product of the matrix, is the F-norm; 2) Set initial parameters: B 0 =O,S 0 =0,E 0 =0,F 0 =0, initial value of Lagrange multiplier Z 0 =0, ADMM parameter ρ = 5.5, maximum penalty factor μ max =μ 0 ×10 7 , relative error tol = 10 -5 , iterative number k = 0, initial value of Lagrange multiplier Initial value of penalty factor Regularization parameters λ1>0,λ2>0,λ3>0; 3) Update variable B; B k+1 =Udiag{σ * }V T Among them, U, V are orthogonal matrices, and the diagonal elements σ * is the singular value of B, non-negative and arranged in descending order, B is the static background matrix, and the singular value threshold method is used to solve the variable B; B k+1 is the static background matrix at K+1 step length; 4) Update variable S; Where prox represents the nearest neighbor operator, and the quadratic minimum optical flow method is used to solve the S subproblem; S k+1 is the sparse moving target matrix with K+1 step length; 5) Update variable E; in, represents the soft threshold shrinkage operator; E k+1 is the dynamic background matrix at K+1 step length; 6) Update variable F; In order to simplify the operation, a new variable is introduced K=Vf; f is a vector; Among them: F k+1 represents auxiliary variables, μ k is the penalty factor for the k-step, Z k is the k-step Lagrange multiplier; The transformed augmented Lagrangian function is: Among them, ζ is the penalty parameter, T is the Lagrange multiplier, and the variables t and k are the number of inner and outer loops respectively; Let q = {1} and update the auxiliary variable K as follows: Where: K t+1 Represents the updated auxiliary variables at t+1; Update the parameter T as follows: T t+1 =T t +ζ t (K t+1 -Vf t+1 ) Through μ k I+ζ t V T The inverse matrix of V is used to solve the variable f; reshaping the variable F into a three-dimensional matrix, the exact analytical solution of F is as follows: in, denote inverse Fourier transform and Fourier transform respectively; 7) Update the multiplier X as follows; X k+1 =X k +m k (OB k+1 -E k+1 -S k+1 ) X k+1 represents the update multiplier for k+1 steps, X k represents the update multiplier for k steps, 8) Update the Lagrange multiplier Z as follows; WITH k+1 =Z k +μ k (S k+1 -F k+1 ) Z k+1 represents the update multiplier for k+1 steps; 9) Update the penalty factor μ as follows: m k+1 =min(ρμ k ,m max ) Among them, ρ is the step size; μ max is the maximum value of μ; Let k=k+1 and repeat steps 2)-9) until the termination condition is met, the iteration is terminated, and the low-rank matrix B corresponding to the static background in the video information, the sparse matrix S corresponding to the moving target, and the matrix E corresponding to the dynamic background are output.

8. The method for detecting moving targets in a dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis according to claim 7, characterized in that: The termination conditions are:

9. A moving target detection device under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis, characterized in that: include: Acquisition module: used to acquire video information of moving targets; Conversion module: used for converting the video information of the moving target into a three-dimensional observation matrix; Decomposition module: used to decompose the observation video matrix O into a low-rank static background matrix B, a sparse moving target matrix S and a dynamic background matrix E according to the background and characteristics of the moving target; Construction module: used to constrain the decomposed matrix and construct a moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis; Solution module: used to transform the moving target detection model under dynamic background based on non-convex spatiotemporal low-rank robust principal component analysis into an augmented Lagrangian function, and solve the moving target detection sub-problem under dynamic background of the model using the alternating multiplier method; One-area module: It is used to obtain the extracted moving target according to the result of solving the sub-problem, perform binary masking on the extracted moving target, and realize the detection of the moving target.

10. A computer device comprising: A processor and a memory, wherein the memory stores a program module, wherein the program module runs on the processor to implement the method according to any one of claims 1 to 8.

Images