Satellite video moving target detection method based on intrinsic decomposition

Abstract

The invention discloses a satellite video moving target detection method based on intrinsic decomposition, relates to the satellite video moving target detection method based on intrinsic decomposition, and belongs to the field of remote sensing image processing. The invention aims to solve the problems that an existing satellite video moving target detection method based on deep learning seriously depends on a large-scale accurately-labeled data set and needs a long training process; and a traditional image processing technology may not fully suppress background noise in a complex scene. The method comprises the following steps: obtaining an optimization problem based on a time-space consistency constraint of a reflectivity component, a local space consistency constraint of a light and shade component and a signal decomposition constraint; solving the optimization problem to obtain a reflectivity component and a light and shade component; tensor robust principal component analysis is carried out on the reflectivity component, and a moving target is extracted; and processing the extracted moving target by using a noise filter, and outputting a final detection result.

Description

Technical Field

[0001] The invention relates to a satellite video moving target detection method based on eigendecomposition, and belongs to the field of remote sensing image processing. Background Art

[0002] In recent years, with the successful launch of video satellites, satellite video research has received increasing attention. Video satellites have staring imaging capabilities and can obtain continuous video data for fixed areas. Since the spatial size of each frame in satellite videos is large, they contain rich and valuable information, such as mountains, rivers, cities, ships, and airplanes, providing rich data for various applications. Detection and tracking of small dynamic targets such as ships and airplanes is an important topic in satellite video applications. Motion target detection involves locating and identifying targets in each frame of satellite video and obtaining quantitative descriptive information about the target, including outline size, location, and confidence, which has a wide range of application value. For example, satellite videos can provide long-term and large-scale monitoring information for commercial areas, estimate traffic density, and evaluate commercial activities.

[0003] Existing satellite video moving target detection methods can be roughly divided into several categories: deep learning-based methods, such as ClusterNet and DSFNet, which show strong detection capabilities but rely heavily on large-scale accurately annotated datasets and require a long training process. Traditional image processing techniques, including frame difference methods, such as D&T and MMB; and background modeling methods, such as GoDec, DECOLOR, E-LSD, and B-MCMD, directly process raw video frames and have high computational efficiency, but may not be able to fully suppress background noise in complex scenes. Summary of the invention

[0004] The purpose of the present invention is to solve the problems that the existing satellite video motion target detection methods based on deep learning are heavily dependent on large-scale accurately labeled data sets and require a long training process; and traditional image processing technology may not be able to fully suppress background noise in complex scenes, and propose a satellite video motion target detection method based on intrinsic decomposition.

[0005] A satellite video moving target detection method based on eigendecomposition has the following specific process:

[0006] Step 1: Set the spatiotemporal consistency constraints of the reflectance component, the local spatial consistency constraints of the brightness and darkness components, and the signal decomposition constraints;

[0007] Step 2: Based on the spatiotemporal consistency constraint of the reflectivity component, the local spatial consistency constraint of the brightness and darkness components, and the signal decomposition constraint, an optimization problem is obtained;

[0008] Step 3: Solve the optimization problem in step 2 to obtain the reflectivity component and the brightness component;

[0009] Step 4: Perform tensor robust principal component analysis on the reflectivity components calculated in step 3 to extract moving targets;

[0010] Step 5: Use a noise filter to process the moving target extracted in step 4 and output the final detection result.

[0011] The beneficial effects of the present invention are:

[0012] The present invention proposes an innovative method for detecting moving targets in satellite videos based on eigendecomposition. Unlike previous methods, the method of the present invention does not rely on a large amount of labeled data and does not require a long training process, thereby reducing the cost and difficulty of data preparation. At the same time, the eigendecomposition technology that considers the characteristics of moving targets and background is used to pre-process the background noise in satellite videos, effectively reducing the background noise, and solving the problem that traditional image processing technology may not be able to fully suppress background noise in complex scenes. The present invention significantly reduces the mean square error of the background and has a good detection accuracy compared with other methods.

[0013] This method uses a process combining intrinsic decomposition with motion information to construct temporal and spatial consistency constraints for the reflectance component (the same background pixels in the reflectance component have the same value in its field and adjacent frames) and local spatial consistency constraints for the light and dark components (the same moving target pixels in the light and dark components are similar to the background). This can effectively focus the moving target on the reflectance component and reduce the background noise caused by illumination changes, laying a solid foundation for subsequent processing. At the same time, there is no need to annotate and train massive amounts of data, saving a lot of labor costs. Experiments were conducted on seven Jilin-1 satellite videos, and the experimental results show that our method has achieved considerable performance in satellite video moving target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the implementation process of the present invention;

[0015] Figure 2 is a frame from the original satellite video of the experiment;

[0016] Figure 3 is a frame in the experimental reflectance component video;

[0017] Figure 4 is a frame in the experimentally separated dynamic reflectivity component video;

[0018] Figure 5This is a comparison of the visualization results of the method of the present invention and other satellite video moving target detection methods. From left to right and from top to bottom, the order is: true value, MMB, D&T, Godec, DECOLOR, ClusterNet, DSFNet, and the visualization result of the method of the present invention. DETAILED DESCRIPTION

[0019] Specific implementation method 1: This implementation method is a satellite video moving target detection method based on intrinsic decomposition. The specific process is as follows:

[0020] Step 1: Based on the principle of intrinsic decomposition and combined with motion information, the spatiotemporal consistency constraints of the reflectivity component, the local spatial consistency constraints of the brightness and darkness components, and the signal decomposition constraints are set to reduce background noise and retain the information of the moving target;

[0021] Step 2: Spatiotemporal consistency constraints based on reflectivity components f ρ , local spatial consistency constraints of light and dark components f L and signal decomposition constraint f D , obtain the optimization problem;

[0022] Step 3: Solve the optimization problem in step 2 to obtain the reflectivity component and the brightness component;

[0023] Step 4: Perform a tensor robust principal component analysis on the reflectivity component calculated in step 3 to extract the moving target (a picture containing many pixels);

[0024] Step 5: Use a noise filter to process the moving target extracted in step 4, further reduce the background noise, and output the final detection result.

[0025] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that: in step 1, based on the principle of intrinsic decomposition and combined with motion information, the spatiotemporal consistency constraint of the reflectivity component, the local spatial consistency constraint of the brightness component and the signal decomposition constraint are set to reduce background noise and retain the information of the moving target;

[0026] The specific process is:

[0027] 1) Based on the eigendecomposition theory and combined with motion information, the spatiotemporal consistency constraint f of the reflectivity component is set ρ , the expression is:

[0028]

[0029] Among them, f ρ (ρ t-1 ,ρ t ) represents the spatiotemporal consistency constraint function of the reflectivity component, ρ t,xRepresents the pixel value of the current pixel in the current frame of the reflectivity component video, ρ t-1,x Represents the pixel value of the current pixel in the previous frame of the reflectivity component video, ρ t-1,x+δ(x) Represents the pixel values ​​of the pixels around the current pixel in the previous frame of the reflectivity component video; x represents the pixel; δ(x) is the range of the pixels around the current pixel (with the current pixel as the center point, a 3×3 area);

[0030] 2) Based on the eigendecomposition theory and combined with motion information, set the local spatial consistency constraint f of the light and dark components L , the expression is:

[0031]

[0032] Among them, f L (L t ) represents the local spatial consistency constraint function of the light and dark components, L t,x+δ(x) Indicates the pixel values ​​of the surrounding pixels of the current pixel in the current frame of the light and dark component video, L t,x Indicates the pixel value of the current pixel in the current frame of the light and dark component video. δ(x) is the range of pixels around the current pixel (a 3×3 area with the current pixel as the center point). x represents the pixel.

[0033] 3) Based on the eigendecomposition theory, set the signal decomposition constraint f D , the original formula using intrinsic information is expressed as follows:

[0034]

[0035] Among them, f D (ρ t ,L t ) represents the signal decomposition constraint function, I t,x represents the pixel value of the current pixel point in the current frame of the satellite video, x represents the pixel point; ρ t,x Indicates the pixel value of the current pixel in the current frame of the reflectivity component video, L t,x Indicates the pixel value of the current pixel point in the current frame of the light and dark component video;

[0036] f p represents the background spatiotemporal consistency constraint of the reflectance component, assuming that the background pixels in the reflectance component remain unchanged in their neighborhood and adjacent frames; L It represents the spatial consistency constraint of moving objects in the light and dark components, assuming that the moving object pixels in the light and dark components are the same as their background neighbors.

[0037] The other steps and parameters are the same as those in the first embodiment.

[0038] Specific implementation method three: This implementation method is different from specific implementation method one or two in that: the spatiotemporal consistency constraint f based on the reflectivity component in step 2 ρ , local spatial consistency constraints of light and dark components f L and signal decomposition constraint f D , get the optimization problem; the specific process is:

[0039] The optimization problem is expressed as:

[0040]

[0041] Among them, ρ t Indicates the reflectivity component video current frame, L t Indicates the current frame of the light and dark component video, I t Indicates the current frame of satellite video, U t Represents optical flow (a two-dimensional vector field that describes the trajectory and speed of pixels over time in continuous frame images);

[0042] f D represents the signal decomposition constraint (the product of two components is equal to the original image), f D (ρ t ,L t |I t ) represents the signal decomposition constraint function, ρ t ,L t |I t It means that the satellite video is given, and the brightness component and reflectivity component are to be solved;

[0043] f ρ represents the spatiotemporal consistency constraint (the same background pixels in the reflectance component have the same value in their neighborhood and adjacent frames), f ρ (ρ t-1 ,ρ t |U t ) represents the spatiotemporal consistency constraint function of the reflectivity component; ρ t-1 ,ρ t |U t Indicates U t is given, ρ t-1 and ρ t It is to be solved;

[0044] f L represents the local spatial consistency constraint (the same moving target pixels in the light and dark components are similar to the background); f L (L t ) represents the local spatial consistency constraint function of the light and dark components.

[0045] The other steps and parameters are the same as those in the first or second embodiment.

[0046] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that: in step 3, the optimization problem of step 2 is solved to obtain the reflectivity component and the brightness component; the specific process is:

[0047] Step 31: Initialization: all pixel values ​​of reflectivity component and brightness component are 0;

[0048] Step 32: Solve the optimization problem of step 2 by using the first-order derivative equal to zero to obtain the reflectivity component and the brightness component;

[0049] The specific process is:

[0050] The expression of reflectivity component is:

[0051] ρ t,x =(1-ε)ρ t,x +0.5ε(W t-1 +W t )ρ t,x+δ(x) +I t,x L t,x

[0052] The expression of light and dark components is:

[0053]

[0054] Where I represents the pixel value set of all pixels in the current frame of the satellite video; B represents the B channel in RGB; ε represents a hyperparameter with a value range of 0-1; ρ t,x Represents the pixel value of the current pixel in the current frame of the reflectivity component video, W t-1 Represents the weight of the pixel in the previous frame, W t Represents the weight of the pixel in the current frame, ρ t,x+δ(x) Represents the pixel values ​​around the current pixel point in the current frame of the reflectivity component video, I t,x Indicates the pixel value of the current pixel point in the current frame of the satellite video, L t Indicates the current frame of the light and dark component video; t represents the current frame, and x represents the pixel point;

[0055] Step 33: Repeat step 32 until M times of obtaining the reflectivity component and the brightness component (the obtained ρ t,x and L t Bring it into the next loop as the ρ on the right side of the equal sign t,x and L t ).

[0056] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0057] Specific implementation mode 5: This implementation mode is different from any one of specific implementation modes 1 to 4 in that: the 80≤M≤120.

[0058] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0059] Specific implementation method 6: This implementation method is different from any one of the specific implementation methods 1 to 5 in that: the weight W of the pixel point of the current frame t The calculation method is as follows:

[0060] W t The calculation method is as follows:

[0061]

[0062] Among them, I t,δ(x),Y Indicates the Y dimension value of the pixels around the current pixel, I t,x,Y Indicates the Y dimension value of the current pixel, I t,δ(x) represents the pixel value of the pixels around the current pixel, δ(x) represents the range of the pixels around the current pixel; D represents the variance operation, d1 represents the variance of the Y dimension in the pixel neighborhood after the image is converted to NTSC format; d2 represents the variance of the three YIQ dimensions in the pixel neighborhood after the image is converted to NTSC format; ‖‖2 represents the two-norm.

[0063] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0064] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that the process of obtaining whether the pixel point x belongs to the foreground or the background is as follows:

[0065] Tensor robust principal component analysis is performed on the satellite video to roughly determine whether each pixel point x in the satellite video belongs to the foreground or the background.

[0066] Pixel categories are determined by moving object location masks obtained by directly performing tensor robust principal component analysis on satellite videos.

[0067] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0068] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that: in step 4, a tensor robust principal component analysis is performed on the reflectivity component to extract a moving target (a picture, including many pixels); the specific process is as follows:

[0069] For the reflectivity component video containing motion information obtained in the previous stage, the tensor robust principal component analysis method is further used to separate the static and dynamic components. Tensor robust principal component analysis divides the image into low-rank parts and sparse parts. In satellite videos, low-rank corresponds to static backgrounds, while sparse corresponds to moving targets.

[0070] The tensor robust principal component analysis model is shown as follows:

[0071]

[0072] st t =ρ 静 +ρ 动

[0073] Among them, |||| * represents the nuclear norm, ||||1 represents the L1 norm, ρ 动 represents the dynamic reflectivity component obtained by performing tensor robust principal component analysis on the reflectivity component, ρ 静 represents the static reflectivity component obtained by performing a tensor robust principal component analysis on the reflectivity component, λ represents a hyperparameter; ρ t Represents the current frame of reflectivity component video;

[0074] The alternating direction multiplier method is used to solve the tensor robust principal component analysis model to obtain the sparse part, namely the moving target (a picture containing many pixels).

[0075] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0076] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that: in step 5, a noise filter is used to process the extracted moving target to further reduce the background noise and output the final detection result; the specific process is:

[0077] Step 51: Calculate a threshold T. Those below the threshold are considered as noise, while those above the threshold are temporarily considered as moving targets. The dynamic reflectivity component obtained in the previous step contains a lot of noise but is weak - most of the pixel values ​​are 1, while the moving targets are few but strong - most of the pixel values ​​are above 10. Combined with the mean and variance values ​​of the dynamic reflectivity component; the threshold is calculated as follows:

[0078] T=mean(ρ 动 )+std(ρ 动 )

[0079] Wherein, mean() represents the mean of all pixel values ​​of the moving target obtained in step 4, and std() represents the standard deviation of all pixel values ​​of the moving target obtained in step 4;

[0080] Step 52: compare all pixel values ​​of the moving target (a picture, including many pixels) obtained in step 4 with the threshold T respectively, regard pixel values ​​greater than or equal to the threshold T as preliminary real moving targets, and regard pixel values ​​less than the threshold T as noise;

[0081] Step 53: Perform morphological opening operation on the preliminary real moving target to obtain the processed real moving target to remove the extremely small noise;

[0082] Step 54: retain the target with C pixels in the processed real moving target as the final moving target, and output the detection result.

[0083] The other steps and parameters are the same as those in Specific Implementation 1 to 8-1.

[0084] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that: the range of C is: 20≤C≤100.

[0085] The targets whose pixel numbers among the processed real moving targets are greater than or equal to 20 and less than or equal to 100 (including 20 and 100) are retained as the final moving targets, and the detection results are output.

[0086] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 9.

[0087] The following examples are used to verify the beneficial effects of the present invention:

[0088] Embodiment 1:

[0089] The data used in the experiment are seven videos taken by the Jilin-1 satellite. The videos in this dataset contain different types of backgrounds, such as woodlands, lakes, cities, and various moving objects (including cars, trains, ships, and airplanes), covering a variety of challenges such as lighting changes, dark backgrounds, and dense objects. It is a diverse and comprehensive dataset. Figure 2 A frame from the original satellite video of the experiment is shown. Figure 3 The experimental reflectivity component video results are shown. It can be seen that the information of the moving target is preserved and the background is Figure 2 The results in Table 1 also confirm that the background noise is reduced. Figure 4 This is a frame of the dynamic reflectivity component video separated in the experiment. It can be seen that the moving target is well extracted. Figure 5 Table 2 is a comparison of the visualization results of the method of the present invention and other satellite video motion target detection methods. Table 3 is a comparison of the quantitative detection results of the method of the present invention and other satellite video motion target detection methods. It can be seen that the method of the present invention has a very good detection result.

[0090] Table 1

[0091]

[0092] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A satellite video moving target detection method based on eigendecomposition, characterized by: The specific process of the method is: Step 1: Set the spatiotemporal consistency constraints of the reflectance component, the local spatial consistency constraints of the brightness and darkness components, and the signal decomposition constraints; Step 2: Based on the spatiotemporal consistency constraint of the reflectivity component, the local spatial consistency constraint of the brightness and darkness components, and the signal decomposition constraint, an optimization problem is obtained; Step 3: Solve the optimization problem in step 2 to obtain the reflectivity component and the brightness component; Step 4: Perform tensor robust principal component analysis on the reflectivity components calculated in step 3 to extract moving targets; Step 5: Use a noise filter to process the moving target extracted in step 4 and output the final detection result.

2. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 1, characterized in that: In step 1, the temporal and spatial consistency constraints of the reflectivity component, the local spatial consistency constraints of the brightness and darkness components, and the signal decomposition constraints are set; the specific process is: 1) Set the spatiotemporal consistency constraint f of the reflectivity component ρ , the expression is: Among them, f ρ (ρ t-1 ,ρ t ) represents the spatiotemporal consistency constraint function of the reflectivity component, ρ t,x Represents the pixel value of the current pixel in the current frame of the reflectivity component video, ρ t-1,x Represents the pixel value of the current pixel in the previous frame of the reflectivity component video, ρ t-1,x+δ(x) Represents the pixel values ​​of the pixels around the current pixel in the previous frame of the reflectivity component video; x represents the pixel; δ(x) is the range of pixels around the current pixel; 2) Set the local spatial consistency constraint f of the light and dark components L , the expression is: Among them, f L (L t ) represents the local spatial consistency constraint function of the light and dark components, L t,x+δ(x) Indicates the pixel values ​​of the surrounding pixels of the current pixel in the current frame of the light and dark component video, L t,x Indicates the pixel value of the current pixel in the current frame of the light and dark component video, δ(x) is the range of pixels around the current pixel; x represents the pixel; 3) Set the signal decomposition constraint f D , the formula is as follows: Among them, f D (ρ t ,L t ) represents the signal decomposition constraint function, I t,x represents the pixel value of the current pixel point in the current frame of the satellite video, x represents the pixel point; ρ t,x Indicates the pixel value of the current pixel in the current frame of the reflectivity component video, L t,x Indicates the pixel value of the current pixel in the current frame of the light and dark component video.

3. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 2, characterized in that: In step 2, based on the spatiotemporal consistency constraint of the reflectivity component, the local spatial consistency constraint of the brightness and darkness component, and the signal decomposition constraint, an optimization problem is obtained; the specific process is: The optimization problem is expressed as: Among them, ρ t Indicates the reflectivity component video current frame, L t Indicates the current frame of the light and dark component video, I t Indicates the current frame of satellite video, U t represents optical flow; f D represents the signal decomposition constraint, f D (ρ t ,L t |I t ) represents the signal decomposition constraint function, ρ t ,L t |I t It means that the satellite video is given, and the brightness component and reflectivity component are to be solved; f ρ represents the spatiotemporal consistency constraint, f ρ (ρ t-1 ,ρ t |U t ) represents the spatiotemporal consistency constraint function of the reflectivity component; ρ t-1 ,ρ t |U t Indicates U t is given, ρ t-1 and ρ t It is to be solved; f L represents the local spatial consistency constraint; f L (L t ) represents the local spatial consistency constraint function of the light and dark components.

4. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 3, characterized in that: In step 3, the optimization problem of step 2 is solved to obtain a reflectivity component and a brightness component; The specific process is: Step 31: Initialization: all pixel values ​​of reflectivity component and brightness component are 0; Step 32: Solve the optimization problem of step 2 by using the first-order derivative equal to zero to obtain the reflectivity component and the brightness component; The specific process is: The expression of reflectivity component is: r t,x =(1-e)r t,x +0.5ε(W t-1 +W t )r t,x+δ(x) +I t,x L t,x The expression of light and dark components is: Where I represents the pixel value set of all pixels in the current frame of the satellite video; B represents the B channel in RGB; ε represents the hyperparameter; ρ t,x Represents the pixel value of the current pixel in the current frame of the reflectivity component video, W t-1 Represents the weight of the pixel in the previous frame, W t Represents the weight of the pixel in the current frame, ρ t,x+δ(x) Represents the pixel values ​​around the current pixel point in the current frame of the reflectivity component video, I t,x Indicates the pixel value of the current pixel point in the current frame of the satellite video, L t Indicates the current frame of the light and dark component video; t represents the current frame, and x represents the pixel point; Step 33: Repeat step 32 until the reflectance component and the brightness component are obtained M times.

5. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 4, characterized in that: The 80≤M≤120.

6. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 5, characterized in that: The weight W of the pixel point in the current frame t The calculation method is as follows: W t The calculation method is as follows: Among them, I t,δ ( x ) ,Y Indicates the Y dimension value of the pixels around the current pixel, I t,x,Y Indicates the Y dimension value of the current pixel, I t,δ(x) represents the pixel value of the pixels around the current pixel, δ(x) represents the range of the pixels around the current pixel; D represents the variance operation, d1 represents the variance of the Y dimension in the pixel neighborhood after the image is converted to NTSC format; d2 represents the variance of the three YIQ dimensions in the pixel neighborhood after the image is converted to NTSC format; || ||2 represents the two-norm.

7. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 6, characterized in that: The process of obtaining whether the pixel point x belongs to the foreground or the background is as follows: Tensor robust principal component analysis is performed on the satellite video to determine whether each pixel point x in the satellite video belongs to the foreground or the background.

8. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 7, characterized in that: In step 4, a tensor robust principal component analysis is performed on the reflectivity component to extract the moving target; the specific process is: The tensor robust principal component analysis model is shown as follows: stp t =ρ 静 +r 动 Among them, || || * represents the nuclear norm, || ||1 represents the L1 norm, ρ 动 represents the dynamic reflectivity component obtained by performing tensor robust principal component analysis on the reflectivity component, ρ 静 represents the static reflectivity component obtained by performing a tensor robust principal component analysis on the reflectivity component, λ represents a hyperparameter; ρ t Represents the current frame of reflectivity component video; The alternating direction multiplier method is used to solve the tensor robust principal component analysis model and obtain the sparse part, i.e. the moving target.

9. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 8, characterized in that: In step 5, a noise filter is used to process the extracted moving target and output the final detection result; the specific process is as follows: Step 51: Calculate a threshold T. The threshold is calculated as follows: T=mean(ρ 动 )+std(r 动 ) Wherein, mean() represents the mean value of all pixel values ​​of the moving target obtained in step 4, and std() represents the standard deviation of all pixel values ​​of the moving target obtained in step 4; Step 52: compare all pixel values ​​of the moving target obtained in step 4 with the threshold T respectively, regard pixel values ​​greater than or equal to the threshold T as preliminary true moving targets, and regard pixel values ​​less than the threshold T as noise; Step 53: Perform morphological opening operation on the preliminary real moving target to obtain the processed real moving target; Step 54: retain the target with C pixels in the processed real moving target as the final moving target, and output the detection result.

10. The method for detecting moving targets in satellite video based on eigendecomposition according to claim 9, characterized in that: The range of C is: 20≤C≤100.

Images