Moving Target Detection Method Based on Weighted Low-Rank and Neighborhood Group Sparse Decomposition

Abstract

The present invention discloses a moving target detection method based on weighted low-rank and neighborhood group sparse decomposition, including: first, taking the grayscale image of each frame and normalizing the pixels to the interval [0, 1]; second, constructing an overlapping local image group for each video frame according to the circular neighborhood of the pixels and giving an expression of the neighborhood group sparse induced norm; third, decomposing D into a low-rank background matrix B, a structured sparse foreground S, and a Gaussian noise matrix E; fourth, constructing a weight W based on the structural information of S and the modeling result of the residual noise after background subtraction; fifth, establishing a final weighted low-rank and neighborhood group sparse decomposition model according to the above process; sixth, solving the model in the fifth step by the alternating direction multiplier method, updating the weight during the iteration process, and finally obtaining the foreground target. The present invention can detect foreground objects that are stationary for a short or long time, can handle complex dynamic backgrounds, and enhance the integrity of foreground detection, and is superior to the detection effects of other models in terms of the F-measure index and visual effects.

Description

Technical Field

[0001] The technical solution of the present invention relates to a moving target detection method, and in particular to a moving target detection method based on weighted low-rank decomposition and neighborhood group sparse decomposition. Background Art

[0002] Generally speaking, the research targets in videos are in continuous or intermittent motion, such as pedestrians and vehicles. These moving targets (foregrounds) need to be extracted and interpreted. The goal of moving target detection is to separate the foreground from the background and ultimately detect the foreground. Moving target detection is fundamental to video analysis and a preprocessing step for a range of high-level visual tasks, such as target tracking and behavior recognition. It is not only an important part of computer vision research, but also widely used in real-life applications, such as in intelligent transportation, intelligent surveillance, and healthcare.

[0003] Low-Rank and Sparse Decomposition (LRSD) is a commonly used method in video modeling that can simultaneously separate the foreground and background of a video frame. Its basic assumption is that, for a fixed camera, the background of the video frame has a strong correlation between frames and can be approximated as low-rank. Foreground pixels are considered outliers and occupy a small portion of the video background, so the foreground has a sparse outlier characteristic.

[0004] Low-rank sparse decomposition is applied to the task of moving target detection. The basic steps are: given a video sequence of n frames, each frame contains m pixels, and then these frames are vectorized and stacked into a matrix D∈R m×n Each frame contains a static background component B∈R m×n and foreground component S∈R m×n The observation matrix D can be decomposed into D = B + S. LRSD can be achieved by minimizing the following problem:

[0005]

[0006] where ||·|| *The nuclear norm and the L1 norm are defined by ||·||* and ||·||1 respectively. α regulates the sparsity of the sparse matrix F. This problem is called RPCA-PCP. Through experiments, it is found that the model can basically detect moving targets in simple video scenarios. In recent years, most scholars have improved the L1 sparse constraint of the foreground, taking into account the structural distribution characteristics and spatio-temporal continuity characteristics of the foreground. The overlapping group sparse induced norm is one of the commonly used methods for structuring the foreground. It promotes the estimated foreground to maintain spatial continuity and has high performance in adapting to dynamic backgrounds and improving the integrity of foreground detection. Currently, most use a fixed-size rectangular sliding window to construct local image groups, which cannot accurately capture the pixel neighborhood relationship.

[0007] The advantage of the LRSD-based method lies in the joint solution of the background and the foreground, with the two interacting with each other. Specifically, in terms of background modeling, the foreground can be subtracted from the original video frame to reduce its interference with background modeling, and the remaining part is then used to solve the background through low-rank constraints; the same applies during foreground detection. Its disadvantage is that it ignores the influence of the structural information of the foreground on background modeling, resulting in an inability to model a "clean" background when the foreground briefly stalls, affecting foreground detection.

[0008] Weighted Low-Rank Matrix Factorization (WLRMF) is used to simply model the background and detect the foreground through post-processing methods such as background subtraction and threshold segmentation. The basic mathematical expression is:

[0009]

[0010] where U ∈ R m×r , V ∈ R r×n , r << min{m, n}. W and p need to be defined and solved according to the specific model. In the WLRMF method, other auxiliary information can be added to the weight W to make W play a role in correcting the background. For example, the low-rank decomposition based on mixed noise modeling proposed by Meng Deyu et al., where the weight comes from the result of mixed noise modeling. However, this method only solves or improves background modeling and does not explicitly model the foreground, and cannot use the structural characteristics of the foreground to correct the background model. Summary of the Invention

[0011] Objective of the Invention: The present invention provides a moving target detection method based on Weighted Low-Rank and Neighborhood Group Sparsity Decomposition (WLNGS). This method uses weighted low-rank decomposition to model the background, where the weights are from the structured information of the foreground and the residual modeling results after background subtraction; it uses the neighborhood group sparsity induced norm to model the foreground, where the local image groups are constructed through the circular neighborhoods of pixels to enhance the expression of the neighborhood relationship within the group. This method can detect moving targets with short-term stagnation, alleviate or eliminate the interference of dynamic backgrounds, and improve the integrity of foreground detection.

[0012] The technical solution adopted by the present invention to solve this technical problem is: to provide a moving target detection method based on weighted low-rank and neighborhood group sparsity decomposition, and the steps of this method are as follows:

[0013] First step: Take the grayscale image of each frame and normalize the pixels to the interval [0, 1]. Given n grayscale images, each frame contains m pixels, and then vectorize and stack these frames into a matrix D ∈ R m×n ;

[0014] Second step: According to the circular neighborhoods of pixels, construct overlapping local image groups for each video frame and give the expression of the neighborhood group sparsity induced norm;

[0015] Third step: Decompose the observed video D into a low-rank background matrix B, a structured sparse foreground S, and a Gaussian noise matrix E, where the E component absorbs the dynamic background and other noises;

[0016] Fourth step: Based on the structural information of the foreground S and the residual noise modeling results after background subtraction, construct the expression of the weight W;

[0017] Fifth step: According to the above expression of the neighborhood group sparsity induced norm, the decomposition method of the observed video, and the expression of the weight, establish the final weighted low-rank and neighborhood group sparsity decomposition model;

[0018] Sixth step: Solve the model in the fifth step by the Alternating Direction Method of Multipliers (ADMM), update the weight during the iteration process, and finally obtain the foreground target.

[0019] The above neighborhood group sparsity induced norm, its expression is:

[0020]

[0021] where g ij ={(p, q)|(p - i)2 +(q - j) 2 ≤r N 2} defines the index set of pixels in the circular neighborhood. In this paper, the "neighborhood" is defined on the two-dimensional space unfolded from each frame. ||·|| ∞ represents the infinity norm, that is, the maximum absolute value of all elements in the vector. s k is the k-th column of the matrix S. is a sparse vector formed by using the elements in the set g ij as indices, which is a subset of s k . represents the weights of each group. In the present invention,

[0022] The decomposition objective of the above WLNGS model is as follows:

[0023] D = B + S + E where E represents Gaussian noise. In this paper, it is considered that other interference noises such as dynamic background are generated by the component E. Here, without explicit modeling of the noise term E, B and S are reconstructed by minimizing the decomposition error.

[0024] The above weight matrix W ∈ R m×n has the following expression:

[0025]

[0026] where d i2j2 , b i2j2 , s i2j2 and w i2j2 represent the elements of D, B, S, and W at the position (i2, j2) respectively, and b i2j2 is calculated during each iterative solution process. δ 2 is the scale parameter, which controls the numerical size of the non-zero weights. The weights represent the contribution degree of each pixel position to the background recovery. When s i2j2 = 0, the position (i2, j2) is a background pixel and is affected by Gaussian noise, and the weight w i2j2 is constructed with a Gaussian kernel; when s i2j2 ≠0, the position (i, j) is a foreground pixel and has no contribution to the background recovery, so the weight w i2j2 is 0. The neighborhood group sparse induced norm constrains the foreground matrix S, which promotes the spatial distribution of zero elements in W to have a structural feature. Then the weights integrate the foreground structure information and the modeling result of the residual noise after background subtraction.

[0027] The above WLNGS model uses weighted low-rank matrix factorization for the background B. At the same time, the foreground matrix S is constrained by the neighborhood group sparse norm. Therefore, the mathematical model of WLNGS is as follows:

[0028]

[0029] where U ∈ R m×r and V ∈ R r×n are low-rank matrices, and r << min{m, n}. In this paper, the low-rank property of the background B is directly described in the form of B = UV. The parameter λ > 0 controls the group sparse penalty degree for S.

[0030] The present invention uses ADMM to solve the above-mentioned WLNGS model, which specifically includes the following steps:

[0031] (1) Initialization: Given the observed video matrix D ∈ R m×n , set the initial parameters λ > 0, μ > 0, and ρ > 1.0, and randomly initialize the matrices B (0) , S (0) , U (0) , V (0) and W (0) , and the multiplier Z (0) = 0. The iteration number t = 0;

[0032] (2) Update the variables U and V:

[0033]

[0034] (3) Update the variable B:

[0035]

[0036] (4) Update the variable S:

[0037]

[0038] (5) Update the variable W:

[0039]

[0040] (6) Update the multiplier Z: Z (t+1) = Z (t) + μ (t) (B (t+1) - U (t+1) V (t+1) );

[0041] (7) Update the variable μ: μ (t+1) = ρυ (t) ;

[0042] (8) The algorithm termination condition is Or reach the maximum number of iterations 100. If the termination condition is satisfied, the iteration ends; otherwise, let t = t + 1 and return to step (2).

[0043] Beneficial effects: The present invention first combines weighted low-rank decomposition and neighborhood group sparse decomposition to establish a moving object detection model named WLNGS. Regarding the improvement of low-rank constraints, most of the work focuses on improving the nuclear norm, such as weighted nuclear norm and truncated nuclear norm. In the weighted low-rank decomposition framework of the present invention, only matrix factorization is used for low-rank constraints, and the weight term that can correct the background in the weighted low-rank decomposition model is mainly improved. That is, the weight of WLNGS combines the structured distribution information of the foreground and the residual noise modeling result after background subtraction, and can more accurately restore the static low-rank background, and can detect foreground objects that have stagnated for a short or long time. WLNGS uses neighborhood group sparse constraints for the foreground and forms local groups on the image in a circular neighborhood manner, enhancing the pixel neighborhood relationship within the image group. These two aspects of improvement enable the model to handle complex dynamic backgrounds and improve the integrity of foreground detection. The foreground detected by the present invention is superior to the detection effects of other models in terms of F-measure index and visual effect. Description of the Drawings

[0044] Figure 1 is the flowchart of the present invention;

[0045] Figure 2 is the system construction diagram of the present invention;

[0046] Figure 3 is the comparison diagram (i.e., simulation diagram) of the foreground detection effects between the present invention and other models. Among them, Figure 3 (a) is the original video frame; Figure 3 (b) is the actual foreground; Figure 3 (c) is the result after PCP processing; Figure 3 (d) is the result after DECOLOR processing; Figure 3 (e) is the result after ROUTE processing; Figure 3 (f) is the result after E-LSD processing; Figure 3 (g) is the result after WLNGS processing. Detailed Embodiments

[0047] The motion target detection method of the present invention focuses on the motion target detection of weighted low-rank and neighborhood group sparse decomposition. Because the structural information of the foreground and the residual modeling result after background subtraction play a joint correction role on the background model, the present invention integrates these two aspects into the weights of weighted low-rank decomposition. Taking into account that the current overlapping group sparse induced norm uses a fixed-size sliding window to form a local image group, and the neighborhood relationship of pixels in the group is not clearly expressed, the present invention improves the grouping method and uses the circular neighborhood of pixels to construct the local image group, thereby enhancing the expression of the neighborhood relationship of pixels in the group and enriching the shape and size of the local image group. Then, the proposed model is solved using the alternating direction multiplier method to finally obtain the foreground detection result.

[0048] like Figure 1-2 As shown, first obtain the video data, take the grayscale image of each frame, and normalize the pixels to the [0,1] interval. Given n frames of grayscale images, each frame contains m pixels, and then vectorize these frames and stack them into the observation video matrix D∈R m×n .

[0049] Then, based on the characteristics of foreground structured distribution, the present invention uses the overlapping group sparse induced norm to constrain the foreground matrix and improves the grouping method with pixel circular neighborhood, which is specifically expressed as:

[0050]

[0051] where g ij ={(p,q)|(pi) 2 +(qj) 2 ≤r N 2} defines the index set of pixels in the circular neighborhood. The "neighborhood" in this paper is defined by re-expanding it into two-dimensional space in each frame. ∞ Represents the infinity norm, i.e., the maximum absolute value of all elements in the vector, which forces pixels in the same group to have similar values, thereby improving the completeness of foreground detection. k is the k-th column of the matrix S. The set g ij The elements in the matrix are indices, and the sparse vector formed is s k A subset of . Represents the weight of each group. In the present invention,

[0052] The decomposition objectives of the WLNGS model of the present invention are as follows:

[0053] D=B+S+ED,B,S,E∈R m×n

[0054] where R m×nDenote the real space of m×n, D∈R m×n Denote the observed video matrix, and E denote the Gaussian noise matrix. In this paper, it is considered that other interference noises such as dynamic background are generated by the component E. Here, without considering the explicit modeling of the noise term E, the low-rank background matrix B and the structured sparse foreground matrix S are reconstructed by minimizing the decomposition error. The model adopts weighted low-rank matrix decomposition for the background B. At the same time, the structured sparse foreground matrix S is constrained by the neighborhood group sparse norm. Therefore, the WLNGS model is as follows:

[0055]

[0056] where U∈R m×r and V∈R r×n are low-rank matrices, r << min{m,n}. In this paper, the low-rank property of the background B is directly described in the form of B = UV. The parameter λ>0 controls the group sparse penalty degree for S.

[0057] The present invention uses the alternating direction multiplier method to solve the above model. First, the augmented Lagrangian optimization problem of the model is obtained:

[0058]

[0059] where Z∈R m×n is the Lagrange multiplier, μ>0 is the penalty factor, <·> is the inner product of matrices, and the specific solution process is as follows:

[0060] Update the variables U and V:

[0061]

[0062] Let By taking the partial derivatives of U and V respectively, their closed-form solutions are calculated

[0063] U (t+1) =(MV (t)T )(V (t) V (t)T ) -1 ;

[0064] V (t+1) =(U (t+1)T U (t+1) ) -1 (U (t+1)T M);

[0065] Update the variable B:

[0066]

[0067] To intuitively reflect the correction effect of the weight on the background modeling, the present invention first optimizes this sub-problem in scalar form:

[0068]

[0069] where \(s\) ij and \(z\) ij are the elements of matrices \(S\) and \(Z\) at the \((i, j)\) position respectively, \(u\) i: is the \(i\)-th row of matrix \(U\), and \(v\) :j is the \(j\)-th column of matrix \(V\). Secondly, let \(p\) ij = \(d\) ij - \(b\) ij . The above equation is equivalent to

[0070]

[0071] where \(q\) ij = \((\mu d\) ij - \(\mu(u\) i: \(v\) :j ) + \(z\) ij + \(s\) ij ) / (1 + \(\mu\)). Then, take the first-order partial derivative of \(p\) ij and set it to 0 to calculate the solution of \(p\) ij . Finally, the closed-form solution of \(B\) is derived:

[0072]

[0073] Update variable \(S\):

[0074]

[0075] The form in the above equation is the proximity operator associated with the group sparse-inducing norm, and the solution of this operator can be obtained by solving the quadratic minimum-cost flow problem.

[0076] Update multiplier \(Z\): \(Z\) (t+1) = \(Z\) (t) + \(\mu\) (t) \((B\) (t+1) - \(U\) (t+1) \(V\) (t+1) );

[0077] Update variable \(\mu\): \(\mu\) (t+1) = \(\rho\mu\) (t) .

[0078] In summary, the algorithm steps for solving the WLNGS model using the alternating direction multiplier method are as follows:

[0079] (1) Initialization: Given the observed video matrix \(D\in\mathbb{R}\) m×n , set the initial parameters \(\lambda\gt0\), \(\mu\gt0\), and \(\rho\gt1.0\), and randomly initialize matrices \(B\) (0) , \(S\) (0) , \(U\) (0) , \(V\)(0) and W (0) , multiplier Z (0) = 0. The number of iterations t = 0;

[0080] (2) Update variables U and V:

[0081]

[0082] (3) Update variable B:

[0083]

[0084] (4) Update variable S:

[0085]

[0086] (5) Update variable W:

[0087]

[0088] where d i2j2 , b i2j2 , s i2j2 and w i2j2 represent the elements of D, B, S, and W at the position (i2, 2j) respectively, and b i2j2 is calculated in each iteration solution process; δ 2 is a scale parameter that controls the numerical size of non-zero weights; the weights represent the contribution degree of each pixel position to the background restoration; when s i2j2 = 0, the position (i2, j2) is a background pixel and is affected by Gaussian noise, and the weight w i2j2 is constructed with a Gaussian kernel; when s i2j2 ≠ 0, the position (i2, j2) is a foreground pixel and has no contribution to the background restoration, and the weight w i2j2 is 0.

[0089] (6) Update multiplier Z: Z (t+1) = Z (t) + μ (t) (B (t+1) - U (t+1) V (t+1) );

[0090] (7) Update variable μ: μ (t+1) = ρυ (t) ;

[0091] (8) The algorithm termination condition is or reaching the maximum number of iterations 100. If the termination condition is satisfied, the iteration ends; otherwise, let t = t + 1 and return to step (2).

[0092] To verify the effectiveness of the proposed model, the present invention conducts simulation experiments on several video data segments respectively, and compares the model of the present invention with the PCP, DECOLOR, ROUTE-LRMF (ROUTE) and E-LSD models.

[0093] The present invention uses six video data segments, namely "WaterSurface", "Fountain", "Curtain", "Lobby" provided by the I2R dataset and "canoe", "highway" provided by the CDnet dataset, to conduct comparative experiments on them. It should be emphasized that there are situations where the foreground stagnates briefly in "WaterSurface", "Curtain" and "Lobby".

[0094] In the simulation experiment, it is set to process 220 frames in each batch, and the model parameters are: r N = 1.5, δ 2 = 0.5, r = 1. For scenes with strong interference, λ = 0.12, and for scenes with weak interference and static background, λ = 0.07 (adjust the value of λ appropriately according to the specific video), ρ = 0.31, and the F-measure is used as the performance evaluation index.

[0095] As Figure 3 shown, in order to verify the performance of the WLNGS model proposed by the present invention, a comparative experiment is conducted on the experimental object. The results of the simulation experiment are respectively as Figure 3 (a) to Figure 3 (g) shown. Among them, Figure 3 (a) is the original frame; Figure 3 (b) is the actual foreground; Figure 3 (c) is the result after PCP processing; Figure 3 (d) is the result after DECOLOR processing; Figure 3 (e) is the result after ROUTE processing; Figure 3 (f) is the result after E-LSD processing; Figure 3 (g) is the result after WLNGS processing.

[0096] For videos containing dynamic backgrounds, Figure 3 (c) PCP cannot adapt to this kind of scene, and a large amount of noise appears in the detected foreground; Figure 3 (d) DECOLOR can suppress the interference of the dynamic background, but there is a problem of excessive foreground smoothing, and it is easy to misdetect the area around the moving target as the foreground; neither PCP nor DECOLOR can handle the foreground with brief stagnation; Figure 3(e) ROUTE models the background through weighted low-rank decomposition. Compared with the previous two methods, it can adapt to the situation where the foreground stagnates briefly. However, without imposing structural constraints on the foreground, there are more false detections and missed detections. Figure 3 (f) E-LSD forms local image groups using a fixed-size rectangular sliding window, and applies an overlapping group sparse induction norm to the foreground. It can effectively suppress dynamic backgrounds, and the effect is better in simple scenes. However, there are still deficiencies in improving the integrity of the foreground. It can be seen that the method in this paper Figure 3 (f) WLNGS can significantly suppress the interference of dynamic backgrounds and other noises compared with other methods. Moreover, it can detect foreground objects with short-term stagnation and improve the integrity of detected foregrounds. Therefore, the method proposed in this paper has better performance and universality.

[0097] To further quantitatively compare the foreground detection performance of each method, the F-measure indicators obtained by the proposed model WLNGS and PCP, DECOLOR, ROUTE-LRMF (ROUTE), and E-LSD in this paper are compared, as shown in Table 1. Based on the comprehensive data in the table, the method in this paper has the highest F-measure value in all seven videos, with an average of 0.91, which is 0.06 higher than that of E-LSD, which has the highest average performance among the comparison methods. In dynamic backgrounds such as "WaterSurface", "Fountain", "Curtain", "canoe", and "blizzard", the F-measure of WLNGS is 0.05, 0.06, 0.06, 0.10, and 0.03 higher than that of the model with the highest performance among other methods, respectively. It can be seen that WLNGS shows better detection performance compared with other methods. For simple scenes, WLNGS can also maintain a relatively high detection level. The F-measure value in "highway" has reached 0.99, further indicating the universality of WLNGS in this paper.

[0098] Table 1 Comparison of F-measure indicators of each method

[0099] experimental subject PCP DECOLOR ROUTE E-LSD WLNGS WaterSurface 0.78 0.84 0.85 0.89 0.94 Fountain 0.68 0.83 0.73 0.83 0.89 Curtain 0.69 0.78 0.79 0.83 0.89 canoe 0.12 0.26 0.43 0.82 0.92 blizzard 0.88 0.85 0.90 0.87 0.93 Lobby 0.65 0.61 0.71 0.74 0.84 highway 0.79 0.88 0.75 0.96 0.99 average 0.66 0.72 0.74 0.85 0.91

[0100] As shown in Table 1 above, through experimental verification, the foreground detected by the WLNGS model proposed in this invention is superior to that of other models in terms of F-measure indicators and visual effects, that is, it can suppress the interference of dynamic backgrounds and other noises, detect foregrounds with short-term stagnation, and improve the integrity of detected foregrounds.

Claims

1. A moving target detection method based on weighted low-rank and neighborhood group sparse decomposition, characterized in that: It includes the following steps: S1. Obtain the grayscale images of n frames, where each frame contains m pixels. Normalize the pixels to the interval [0, 1], and then vectorize these frames and stack them into an observation video matrix D ∈ R m×n ; S2. Construct an overlapping local image group for each video frame according to the circular neighborhood of pixels, and give an expression for the neighborhood group sparse induced norm; S3. Decompose the observed video matrix D into a low-rank background matrix B, a structured sparse foreground matrix S, and a Gaussian noise matrix E, where the E component absorbs dynamic background and other noises; S4. Construct an expression for the weight matrix W based on the structure information of the structured sparse foreground matrix S and the modeling result of the residual noise after background subtraction; S5. Establish a final weighted low-rank and neighborhood group sparse decomposition WLNGS model according to the above neighborhood group sparse induced norm expression, the decomposition method of the observed video matrix D, and the expression of the weight matrix W; S6. Solve the WLNGS model in step S5 by the alternating direction method of multipliers ADMM, update the weights during the iteration process, and finally obtain the foreground target; The expression for the neighborhood group sparse induced norm in step S2 is: where g ij = {(p, q)|(p - i) 2 + (q - j) 2 ≤ r N 2}, representing the index set of pixels in the circular neighborhood; ||·|| ∞ represents the infinity norm; s k is the k-th column of matrix S, is the sparse vector s ij formed with elements in set g[[ID=I8]] k as indices; represents the weights of each group; The weight matrix W in the step S4 belongs to R m×n and its expression is: where d i2j2 , b i2j2 , s i2j2 and w i2j2 represent the elements of D, B, S, and W at the position (i2, j2) respectively, and b i2j2 is calculated in each iteration solving process; δ 2 is a scale parameter that controls the numerical value of non-zero weights; when s i2j2 = 0, the position (i2, j2) is a background pixel and is affected by Gaussian noise, and the weight w i2j2 is constructed with a Gaussian kernel; when s i2j2 ≠0, the position (i2, j2) is a foreground pixel and makes no contribution to the background restoration, and the weight w i2j2 is 0; The WLNGS model in step S5 is specifically: where \(U\in\mathbb{R}\) m×r and \(V\in\mathbb{R}\) r×n are low-rank matrices, \(r\ll\min\{m,n\}\), with the constraint \(B = UV\), and the parameter \(\lambda>0\) that controls the group sparsity penalty on \(S\); Step S6 specifically includes the following steps: S61. Initialization: Given the observation video matrix D ∈ R m×n , set the initial parameters λ > 0, μ > 0, and ρ > 1.0, and randomly initialize the matrices B (0) , S (0) , U (0) , V (0) , and W (0) , the multiplier Z (0) = 0, and the iteration number t = 0; S62. Update variables U and V; S63. Update variable B; S64. Update variable S; S65. Update variable W; S66. Update multiplier Z: Z (t+1) = Z (t) + μ (t) (B (t+1) - U (t+1) V (t+1) ); S67. Update variable μ: μ (t+1) = ρυ (t) ; S68. The algorithm termination condition is or the maximum number of iterations 100 is reached; if the termination condition is met, the iteration ends; otherwise, set t = t + 1 and return to step S62.

2. The moving target detection method based on weighted low-rank and neighborhood group sparse decomposition according to claim 1, characterized in that: The said 3. The moving target detection method based on weighted low-rank and neighborhood group sparse decomposition according to claim 1, characterized in that: The decomposition of the observed video matrix D in step S3 is specifically: D = B + S + E; D, B, S, E ∈ R m×n 。

Images