Aircraft tracking method, storage medium and electronic device based on background perception correlation filtering

Abstract

The aircraft tracking method, storage medium and electronic device based on background-aware correlation filtering relate to the technical field of remote sensing video tracking and solve the problem of cloud and fog occlusion. The method includes: initializing the BACF correlation filter, determining the object function, calculating the feature maps of CN and HOG respectively and fusing them to obtain the final feature map; converting the BACF correlation filter to the frequency domain, using L2 regularization to constrain the update of the time series, and using the ADMM algorithm to solve the problem after the constraint; calculating the average peak correlation energy and temporarily storing the point with the largest response value as the final result; judging that the average peak correlation energy in the previous step meets the high confidence, updating the model and the target position based on the calculation result; otherwise, using the motion estimator to estimate the motion state, and weightedly fusing the result of the motion estimator and the calculation result to obtain the updated result; S5, repeating S2, S3 and S4 until the video playback ends.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing video tracking, and in particular to an aircraft tracking technology based on background perception correlation filtering. Background Art

[0002] The launch of remote sensing video satellites enables us to observe and measure moving objects on the earth's surface, providing rich information for monitoring surface change events, among which target detection and tracking are key steps in detection. Aircraft, as an important means of transportation in the military, has a very high research value. However, the current research on remote sensing video detection is mainly focused on the detection and tracking of vehicles, and there is not enough research on aircraft, mainly because of the cloud occlusion problem in remote sensing research. In the case of noise, cloud and light changes, it is impossible to reliably detect targets in each frame. In addition, when the detection algorithm is applied to large-size images, the detection speed is very slow, so it can only be used as the initial value of the moving target.

[0003] Aircraft tracking technology needs to meet the following requirements: 1. Fast computing speed. 2. Tracking can be performed based on specific inputs. 3. It can resist cloud occlusion and operate normally under normal circumstances. In the current tracking research on remote sensing videos, the main focus is on how to detect and track moving targets. This type of tracking detection can only detect and track moving targets. If there are other moving objects in the video, the algorithm cannot track specific targets and processes unnecessary targets. There are also some methods based on deep learning in existing research, but after testing, the speed of its tracking detection is difficult to meet actual requirements. The existing research methods for tracking remote sensing targets based on correlation filtering have relatively simple application scenarios, often for small videos after cropping. The video does not contain complex scenes, such as: smoke, clouds, light spots caused by illumination changes, etc., and cannot achieve more stable results in practical applications.

[0004] In summary, neither deep learning methods nor foreground modeling methods can adapt to remote sensing videos with complex environments. When occlusion occurs, the traditional correlation filtering method will gradually update from the target to completely cloud layer, resulting in a complete offset. The problem of cloud occlusion can be solved by a defogging method based on dark channel data, but it will bring other problems and lack good universality. The defogging method based on dark channel data can reduce cloud interference, but it will cause greater losses to normal samples in the same data, causing their eigenvalues ​​to change and making it impossible to track normally. Summary of the invention

[0005] In order to solve the problem that the existing aircraft tracking technology cannot solve the cloud and fog occlusion, the present invention proposes an aircraft tracking method, a storage medium and an electronic device based on background perception correlation filtering.

[0006] The technical solution of the present invention is as follows:

[0007] An aircraft tracking method based on background-aware correlation filtering comprises the following steps:

[0008] S1, BACF correlation filter initialization: initialize the BACF correlation filter through the image of the first frame and the target area, determine the object function, calculate the feature map of CN and the feature map of HOG respectively, and fuse them to obtain the final feature map;

[0009] S2, convert the BACF correlation filter to the frequency domain through Fourier transform, use L2 regularization to constrain the update of the time series, and use the ADMM algorithm to solve the constrained problem;

[0010] S3, relying on the characteristic spectrum and the BACF correlation filter, calculate the average peak correlation energy, and temporarily store the point with the largest response value in the response graph as the final result;

[0011] S4, update model and target position: compare the average peak correlation energy in the previous step with the historical average peak correlation energy. If the confidence level is high, update the model and target position based on the calculation result of step S3; if the confidence level is not high, use the motion estimator to estimate the motion state, and fuse the result of the motion estimator and the calculation result of step S3 by weighting to finally obtain an updated result.

[0012] S5. Repeat steps S2, S3 and S4 until the video playback ends.

[0013] Preferably, the object function in step S1 is expressed in the frequency domain as follows:

[0014]

[0015]

[0016] Where P is a D×T binary matrix, T is the number of pixels, x represents a training image sample, y represents the corresponding output centered on the peak of the target object, h represents the correlation filter, and x∈R T , y∈R T and h∈R D , x[Δτ i ] represents the cyclic shift of x, K is the number of feature channels, is an auxiliary variable, I is an identity matrix, represents the Kronecker product, ^ represents the denotes discrete Fourier transform, F represents the orthogonal matrix of the complex basis vector to map to the Fourier domain of any L-dimensional vector signal, the operator T represents the conjugate transpose,

[0017] Preferably, the L2 regularization is expressed in the frequency domain as follows:

[0018]

[0019]

[0020] Among them, λ and η are two regularization parameters, η ≥ 0, which are used to adjust the role of the target object in the previous frame in the current frame model training, and T represents the transpose operator on the complex vector or matrix to calculate the conjugate transpose operator.

[0021] Preferably, the The specific calculation method is:

[0022] (I) Using the augmented Lagrangian method Rewrite it as follows:

[0023]

[0024] Where ζ represents the complex Lagrange multiplier, μ represents the penalty factor, μ>0;

[0025] (ii) Use the alternating direction multiplication method to iteratively solve the equation and divide the rewritten formula into the following three sub-problems:

[0026]

[0027] (III) Solve the three sub-problems separately:

[0028] (1) Solve for h:

[0029]

[0030] Among them, g and ζ are represented by (2)

[0031]

[0032] Decompose into independent sub-problems and solve them:

[0033]

[0034] in,

[0035]

[0036] (3) Yes Solution:

[0037]

[0038] in, μ is the penalty factor, μ=min(μ max ,βμ) is continuously updated through the alternating direction multiplier algorithm,μ max represents the maximum value of the historical value of μ, and β represents the scaling factor.

[0039] Preferably, the motion estimator in step S4 is a Kalman filter.

[0040] Preferably, the state equation and observation equation of the Kalman filter system are as follows:

[0041] X k =φ x,k-1 X k-1 +W k-1 ,

[0042] Y k =H k X k +V k ,

[0043] Among them, X k and X k-1 are the state vectors of the system at time k and k-1, respectively, x,k-1 is the state transition matrix of the system, H k is the observation matrix of the system, W k and V k is the noise matrix following the Gaussian matrix with covariance matrix Q k and R k Distribution of

[0044] Choose the state vector to be x k =[xs k ,ys k ,xv k ,yv k ], where xs k and ys k are the horizontal and vertical positions of the object at time k, respectively, xv k and yv k are the horizontal and vertical velocities of the object at time k, respectively;

[0045] The state transition matrix is ​​expressed as follows:

[0046]

[0047] The observation vector is Y k =[xw k ,yw k], represents the position of the object observed at time k, H k It is expressed as:

[0048] The motion state estimates are as follows:

[0049]

[0050]

[0051]

[0052]

[0053] P k+1 =(IK k+1 H k+1 ) k+1,k ,

[0054] in, is the optimal state estimation, K is the KF gain matrix, Q and R are the covariance matrices of the noise;

[0055] The hypothetical motion state is used to simulate the real motion state to estimate the motion of the object before the Kalman filter converges. The velocity of the object in the current frame is estimated by the average displacement of the previous frame. The position of the object in the current frame can be estimated using the velocity and the position of the object in the previous frame. The motion state estimate is described by the following equation:

[0056]

[0057]

[0058] P t =AS t-1 n,

[0059] Among them, S t-1 =(x t-1 ,y t-1 ,Δx t-1 ,Δy t-1 ) is the state vector of the object at time t-1, P t =(x t ,y t ) is the position vector of the object at time t, A is a transfer matrix, n is the number of frames used for estimation, n ≥ 30, determined by considering the FPS of the satellite video.

[0060] Preferably, the average peak correlation energy is calculated as follows:

[0061]

[0062] Among them, F max 、F min 、F w，h They represent the maximum value, minimum value and response at the (w, h) position respectively.

[0063] Preferably, the judgment method that satisfies high confidence is:

[0064] In the current frame t, only when y max and E APCE The target center position is judged to have high confidence only when the values ​​of both exceed the historical mean by a certain ratio α and β, that is, the following two conditions need to be met at the same time:

[0065]

[0066] After calculating F i，max and APCE j It is saved in the corresponding sets Sy and SE as one of the historical values ​​for the next judgment.

[0067] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, wherein the computer program executes the aircraft tracking method based on background-aware correlation filtering as described above.

[0068] The present invention also provides an electronic device, including a processor and a memory, wherein the processor and the memory communicate with each other via a communication bus; the memory is used to store computer programs; the processor is used to implement the above-mentioned aircraft tracking method based on background perception correlation filtering when executing the computer program stored in the memory.

[0069] Compared with the prior art, the present invention solves the problem that the prior aircraft tracking technology cannot solve the cloud and fog obstruction, and the specific beneficial effects are:

[0070] 1. This invention proposes a background-aware correlation filtering algorithm based on time series regularization constraints to resist the model drift caused by clouds, and uses motion information for correction, and uses the alternating direction multiplier algorithm to quickly solve the problem in the frequency domain. This invention can effectively alleviate the model drift problem caused by cloud occlusion, so that cloud data that was previously difficult to use can be used under certain conditions.

[0071] 2. The present invention adopts a correlation filter based on background perception to solve the problems caused by complex background changes and rotational deformation. It reduces the basic calculation amount by reducing the search range and avoids the excessive size of single-frame video data. It introduces a Kalman filter with high computational efficiency to add the motion state to our method, further improving the tracking accuracy and success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 The present invention provides a flow chart of the method. DETAILED DESCRIPTION

[0073] In order to make the technical solution of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the specification of the present invention. It should be noted that the following embodiments are only used to better understand the technical solution of the present invention and should not be understood as a limitation to the present invention.

[0074] Example 1.

[0075] This embodiment provides an aircraft tracking method based on background perception correlation filtering, such as Figure 1 As shown, the following steps are included:

[0076] S1, BACF correlation filter initialization: initialize the BACF correlation filter through the image of the first frame and the target area, determine the object function, calculate the feature map of CN and the feature map of HOG respectively, and fuse them to obtain the final feature map;

[0077] S2, convert the BACF correlation filter to the frequency domain through Fourier transform, use L2 regularization to constrain the update of the time series, and use the ADMM algorithm to solve the constrained problem;

[0078] S3, relying on the characteristic spectrum and the BACF correlation filter, calculate the average peak correlation energy, and temporarily store the point with the largest response value in the response graph as the final result;

[0079] S4, update model and target position: compare the average peak correlation energy in the previous step with the historical average peak correlation energy. If the confidence level is high, update the model and target position based on the calculation result of step S3; if the confidence level is not high, use the motion estimator to estimate the motion state, and fuse the result of the motion estimator and the calculation result of step S3 by weighting to finally obtain an updated result.

[0080] S5. Repeat steps S2, S3 and S4 until the video playback ends.

[0081] Example 2.

[0082] This embodiment is a further example of the first embodiment. First, the BACF tracking method of perceiving the background is used to train the classifier by densely sampling the image patches. The object function of the BACF tracker can be written as follows:

[0083]

[0084] Where P is a D×T binary matrix and T is the number of pixels. T , y∈R T and h∈R D , and x[Δτ i ] represents the circular shift of x. In the visual object tracking task, x represents a training image sample, y represents the corresponding output centered on the peak of the target object, and h represents the correlation filter. By using the circular shift operator (Px[Δτi]) on the training example, all possible image patches will be returned. K is the number of feature channels. By using the circle shift operator, the number of samples will increase. In order to improve computational efficiency, the above formula can be expressed in the frequency domain as follows:

[0085]

[0086]

[0087] in, is an auxiliary variable, I is an identity matrix, represents the Kronecker product, ^ represents the denotes discrete Fourier transform, F represents the orthogonal matrix of the complex basis vector to map to the Fourier domain of any L-dimensional vector signal, the operator T represents the conjugate transpose,,

[0088] Example 3.

[0089] This embodiment is a further example of Embodiment 1, and the L2 regularization is expressed in the frequency domain as follows:

[0090]

[0091]

[0092] Among them, λ and η are two regularization parameters, η ≥ 0, which are used to adjust the role of the target object in the previous frame in the current frame model training, and T represents the transpose operator on the complex vector or matrix to calculate the conjugate transpose operator.

[0093] The deformation, occlusion or background clutter of the target object will have a significant impact on the tracking performance. For example, if there is occlusion, the BACF tracker will lose the target object, and even if the occlusion disappears in the subsequent video frame, the tracker still cannot locate the target object. In the motion of the target object, there is a potential relationship between the target object in consecutive frames. In general, there is inevitably similarity information of the moving target object between the current frame and the previous frame. Considering the relationship between the moving target objects in the time series, the improved background-aware correlation filtering algorithm is learned by utilizing the L2 regularization term constraint. This embodiment minimizes the following object function:

[0094]

[0095] st h=P T h,

[0096] Where λ and η are two regularization parameters (λ, η ≥ 0), η is mainly used to adjust the role of the target object of the previous frame in the current frame model training, and the last term in the formula refers to the global time consistency constraint. In order to improve the computational efficiency, the correlation filter is usually converted to the frequency domain by Fourier transform to obtain the representation described in this embodiment.

[0097] Example 4.

[0098] This embodiment is a further illustration of embodiment 3. The specific calculation method is:

[0099] (I) Using the augmented Lagrangian method Rewrite it as follows:

[0100]

[0101] Where ζ represents the complex Lagrange multiplier, μ represents the penalty factor, μ>0;

[0102] (ii) Use the alternating direction multiplication method to iteratively solve the equation and divide the rewritten formula into the following three sub-problems:

[0103]

[0104] (III) Solve the three sub-problems separately:

[0105] (1) Solve for h:

[0106]

[0107] Among them, g and ζ are represented by (2)

[0108]

[0109] Decompose into independent sub-problems and solve them:

[0110]

[0111] in,

[0112]

[0113] (3) Yes Solution:

[0114]

[0115] in, μ is the penalty factor, μ=min(μ max ,βμ) is continuously updated through the alternating direction multiplier algorithm,μ max represents the maximum value of the historical value of μ, and β represents the scaling factor.

[0116] Example 5.

[0117] This embodiment is a further example of Embodiment 1, and the motion estimator in step 4 is a Kalman filter.

[0118] Example 6.

[0119] This embodiment is a further illustration of Embodiment 5. The state equation and observation equation of the Kalman filter system are as follows:

[0120] X k =φ x,k-1 X k-1 +W k-1 ,

[0121] Y k =H k X k +V k ,

[0122] Among them, X k and X k-1 are the state vectors of the system at time k and k-1, respectively, x,k-1 is the state transition matrix of the system, H k is the observation matrix of the system, W k and V k is the noise matrix following the Gaussian matrix with covariance matrix Q k and R k Distribution of

[0123] Choose the state vector to be x k =[xs k ,ys k ,xv k ,yv k ], where xs k and ys k are the horizontal and vertical positions of the object at time k, respectively, xv k and yv k are the horizontal and vertical velocities of the object at time k, respectively;

[0124] The state transition matrix is ​​expressed as follows:

[0125]

[0126] The observation vector is Y k =[xw k ,yw k ], represents the position of the object observed at time k, H k It is expressed as:

[0127] The motion state estimates are as follows:

[0128]

[0129]

[0130]

[0131]

[0132] P k+1 =(IK k+1 H k+1 ) k+1,k ,

[0133] in, is the optimal state estimate, K is the KF gain matrix, Q and R are the noise covariance matrices, which can be adjusted according to the actual situation. P0 is usually initialized with non-zero random data. I is the identity matrix. The calculation of KF only includes 10 matrix multiplications, 5 matrix additions, and the inverse of a 2×2 matrix. Compared with the computational complexity of BACF, the increase in computational complexity is very small because the size of the largest matrix is ​​4×4.

[0134] Kalman filters require a certain amount of data to converge. Experiments show that when moving objects in satellite videos, Kalman filters can converge after 20-40 frames. Kalman filters have high accuracy in estimating the motion state of objects, but Kalman filters are complex and the filter cannot converge until some frames are used to update the filter. In order to estimate the motion of objects before the Kalman filter converges, we propose a method to simulate the real motion state using an assumed motion state. Typical moving objects in satellite videos are motor vehicles, aircraft, and ships. We can assume that over a short period of time, objects move in a uniform straight line, even if the object is in a state of turning, emergency stop, or acceleration. Based on this assumption, the speed of the object in the current frame can be estimated by the average displacement of the previous frame. The position of the object in the current frame can be estimated using the speed and the position of the object in the previous frame. Therefore, the motion state estimate is described by the following equation:

[0135]

[0136]

[0137] P t =AS t-1 n,

[0138] Among them, S t-1 =(x t-1 ,y t-1 ,Δx t-1 ,Δy t-1 ) is the state vector of the object at time t-1, P t =(x t ,y t ) is the position vector of the object at time t, A is a transfer matrix, n is the number of frames used for estimation, n≥30, determined by considering the FPS of the satellite video. If n is too small, the estimation method will be too sensitive to changes in the object's motion state. If n is too large, the above assumption does not hold. Therefore, this value needs to be carefully selected. If n<30, it is considered that the motion Kalman filter estimation result cannot be used as a reference value.

[0139] Example 7.

[0140] This embodiment is a further example of embodiment 1. The calculation method of the average peak correlation energy is:

[0141]

[0142] Among them, F max 、F min 、F w，h They represent the maximum value, minimum value and response at the (w, h) position respectively.

[0143] Example 8.

[0144] This embodiment is a further example of Embodiment 1, and the judgment method that satisfies a high confidence level is:

[0145] In the current frame t, only when y max and E APCE The target center position is judged to have high confidence only when the values ​​of both exceed the historical mean by a certain ratio α and β, that is, the following two conditions need to be met at the same time:

[0146]

[0147] After calculating F i , max and APCE jIt is saved in the corresponding sets Sy and SE as one of the historical values ​​for the next judgment.

[0148] When the tracked target is blocked, has a large deformation, is blurred or lost, the current value of APCE will be significantly lower than the historical mean, indicating that the current response graph oscillates and has multiple peaks. At this time, the confidence of the target center position is considered to be low. Usually, when multi-peak oscillations occur, the response value of the target center position will also be significantly reduced, that is, the peak value Fmax is generally lower than the peak value in normal interference-free conditions. It can be seen that Fymax reflects the confidence of the target center position from the local response graph, and APCE reflects its confidence from the overall response graph. The combination of the two can obtain a higher confidence.

[0149] In order to reduce the amount of calculation of the algorithm, we do not perform position correction on every frame. Instead, when multi-peak oscillation occurs in the current t-th frame and the target center position may be misjudged, that is, when Fmax and APCE do not meet any of the conditions in high-confidence detection, we will introduce motion information to correct the position. At this time, we will fuse the motion information with the relevant filter information through weighted fusion, replace the current prediction result with the fused result, and update the filter model with the current result.

[0150] Example 9.

[0151] This embodiment provides a computer-readable storage medium, which is used to store a computer program, and the computer program executes the aircraft tracking method based on background-aware correlation filtering as described in any one of Embodiments 1-8.

[0152] Example 10.

[0153] This embodiment provides an electronic device, including a processor and a memory, wherein the processor and the memory communicate with each other via a communication bus; the memory is used to store computer programs; the processor is used to implement the aircraft tracking method based on background-aware correlation filtering as described in any one of Embodiments 1-8 when executing the computer program stored in the memory.

[0154] Example 11.

[0155] In order to verify whether the method of the present application is effective in combating cloud and fog occlusion, this embodiment has conducted a comparative test. Although the traditional cloud removal method improves the visibility of the target, it often also suffers from loss in texture features. The visualization effect and accuracy loss of normal data processed by cloud removal are very serious. In order to prove that our method has a significant improvement in resisting cloud and fog occlusion, we tested it in data with thick cloud occlusion, medium cloud occlusion and thin cloud occlusion respectively. In addition, we compared the data processed by the existing efficient cloud removal method and conducted a test. It was found that the traditional cloud removal method has a large impact on the data image. Once applied to other data, it will cause serious data loss. Our method can not only track well in the case of cloud and fog occlusion, but also has no effect on the tracking of general targets, and it has improved the tracking of rotating targets and slowly moving targets.

[0156] In our experiments, in order to make a fair comparison, the state-of-the-art trackers of the same type are also in the list of compared trackers. The state-of-the-art trackers of the same type refer to the top trackers with similar functions to AADT. Therefore, ECO-HC is selected from the trackers with hand-crafted features and is also one of the tracking models with good performance. CSK also performs well for introducing kernel tricks and ridge regression methods on MOSSE. CN can obtain color features well and has good performance in images with obvious color contrast. CSR-DCF proposes spatial reliability and channel reliability methods, among which the image segmentation method more accurately selects the effective tracking target area. MKCFup significantly reduces the negative mutual interference of different particles, and STRCF constrains the effective scope of the filter template to solve the boundary effect. BACF is our improved basic method. We compare the improved method with the above advanced methods.

[0157] Our method not only ensures high accuracy and good robustness, but also maintains high operating efficiency. Our experimental environment is shown in the following table: It is executed on a host based on Windows 10 system. The working environment of this host is as follows: Intel(R) Core(TM) i7-9700 CPU processor, RAM 16G, and the specific performance and programming language are as follows:

[0158] Table 1 - Models and their AUC (area below the success rate curve)

[0159] Model Development Language Auc_all Auc_covered FPS AADT (this application) Python 0.779 0.694 76.84 CSRDCF Python 0.772 0.665 31.65 BACF Python 0.769 0.421 90.00 CN Matlab 0.765 0.267 110.68 CSK Python 0.758 0.264 136.07 ECO c++ 0.751 0.561 18.71 MKCFup Python 0.724 0.365 13.66 KCF Matlab 0.648 0.601 187.86 STRCF Python 0.610 0.429 95.32

[0160] Compared with the improved BACF method, the tracking success rate of the algorithm of this application can be more accurate under normal circumstances. And there is a significant improvement in occlusion conditions, especially in the balance between resisting occlusion and general conditions, which greatly exceeds the existing methods. This kind of occlusion is widely present in remote sensing images. In previous practices, because the existing methods do not have a good ability to deal with cloud occlusion problems, although the targets in these data can be distinguished by the eyes, in practice, this kind of data is often directly discarded. Under the existing methods, these cloud-occluded data have not been well used, and the amount of this data is very considerable. Once used, it has very important significance for research and practical applications.

[0161]

[0162]

[0163] The table above shows the CLE results (in pixels, center location error) obtained by the tracker proposed in this application and other methods on 20 sequences. By comparison, it can be seen that the method proposed in this application has achieved satisfactory performance, with a small average center location error. The results show that the algorithm is highly robust to video sequences with fast motion, object occlusion, and large deformation.

[0164] This application uses correlation filters for position estimation, and when occlusion occurs, its impact will be corrected according to the motion state. Its average center position error is only 4.363 pixels, which is much better than the results of other trackers based on correlation filters.

[0165] In summary, the features of remote sensing images themselves change little, and the main model drift is caused by cloud occlusion. We have successfully solved this type of drift problem by learning a time series and using the running information to correct the target model. In addition, during the general flight of an aircraft, we can also effectively obtain the characteristics of the aircraft, thereby achieving aircraft tracking problems in complex scenarios.

Claims

1. An aircraft tracking method based on background-aware correlation filtering, characterized in that: The following steps are involved: S1, BACF correlation filter initialization: initialize the BACF correlation filter through the image of the first frame and the target area, determine the object function, calculate the feature map of CN and the feature map of HOG respectively, and fuse them to obtain the final feature map; S2, convert the BACF correlation filter to the frequency domain through Fourier transform, use L2 regularization to constrain the update of the time series, and use the ADMM algorithm to solve the constrained problem; S3, relying on the characteristic spectrum and the BACF correlation filter, calculate the average peak correlation energy, and temporarily store the point with the largest response value in the response graph as the final result; S4, update the model and target position: compare the average peak correlation energy in the previous step with the historical average peak correlation energy, and if a high confidence level is met, update the model and target position based on the calculation result of step S3; If the high confidence level is not met, the Kalman filter is used to estimate the motion state, and the result of the Kalman filter and the calculation result of step S3 are fused by weighting to finally obtain an updated result; S5, repeat steps S2, S3 and S4 until the video playback ends; The object function in step S1 is expressed in the frequency domain as follows: Where P is a D×Z binary matrix, Z is the number of pixels, x represents a training image sample, y represents the corresponding output centered on the peak of the target object, h represents the correlation filter, and x∈R T , y∈R T and h∈R D , K is the number of feature channels, is an auxiliary variable, I is an identity matrix, represents the Kronecker product, ^ represents the denotes discrete Fourier transform, F represents the orthogonal matrix of the complex basis vector to map to the Fourier domain of any L-dimensional vector signal, the operator T represents the conjugate transpose, The average peak correlation energy is calculated as follows: Among them, F max 、F min 、F w，h Respectively represent the maximum value, minimum value and response at the (w, h) position of the response; The judgment method that satisfies high confidence is: In the current frame t, only when y max and E APCE The target center position is judged to have high confidence only when the values ​​of both exceed the historical mean in proportion α and β, that is, the following two conditions need to be met at the same time: After calculating F i，max and APCE j It is saved in the corresponding sets Sy and SE as one of the historical values ​​for the next judgment.

2. The aircraft tracking method based on background-aware correlation filtering according to claim 1, characterized in that: The L2 regularization is expressed in the frequency domain as follows: Among them, λ and η are two regularization parameters, η ≥ 0, which are used to adjust the role of the target object in the previous frame in the current frame model training, and T represents the transpose operator on the complex vector or matrix to calculate the conjugate transpose operator.

3. The aircraft tracking method based on background-aware correlation filtering according to claim 2, characterized in that: Said The specific calculation method is: (I) Using the augmented Lagrangian method Rewrite it as follows: Where ζ represents the complex Lagrange multiplier, μ represents the penalty factor, μ>0; (ii) Use the alternating direction multiplication method to iteratively solve the equation and divide the rewritten formula into the following three sub-problems: (III) Solve the three sub-problems separately: (1) Solve for h: Among them, g and ζ are represented by (2) Decompose into independent sub-problems and solve them: in, (3) Yes Solution: in, μ is the penalty factor, μ=min(μ max ,βμ) is continuously updated through the alternating direction multiplier algorithm,μ max represents the maximum value of the historical value of μ, and β represents the scaling factor.

4. The aircraft tracking method based on background-aware correlation filtering according to claim 1, characterized in that: The state equation and observation equation of the Kalman filter are as follows: X k =φ x,k-1 X k-1 +W k-1 , Y k =H k X k +V k , Among them, X k and X k-1 are the state vectors of the system at time k and k-1, respectively, x,k-1 is the state transition matrix of the system, H k is the observation matrix of the system, W k and V k is the noise matrix following the Gaussian matrix with covariance matrix Q k and R k Distribution of Choose the state vector to be x k =[xs k ,ys k ,xv k ,yv k ], where xs k and ys k are the horizontal and vertical positions of the object at time k, respectively, xv k and yv k are the horizontal and vertical velocities of the object at time k, respectively; The state transition matrix is ​​expressed as follows: The observation vector is Y k =[xw k ,yw k ], represents the position of the object observed at time k, H k It is expressed as: The motion state estimates are as follows: P k+1 =(I-K k+1 H k+1 )P k+1,k , in, is the optimal state estimation, K is the KF gain matrix, Q and R are the covariance matrices of the noise; The hypothetical motion state is used to simulate the real motion state to estimate the motion of the object before the Kalman filter converges. The velocity of the object in the current frame is estimated by the average displacement of the previous frame. The position of the object in the current frame can be estimated using the velocity and the position of the object in the previous frame. The motion state estimate is described by the following equation: P t =AS t-1 n, Among them, S t-1 =(x t-1 ,y t-1 ,Δx t-1 ,Δy t-1 ) is the state vector of the object at time t-1, P t =(x t ,y t ) is the position vector of the object at time t, A is a transfer matrix, n is the number of frames used for estimation, n ≥ 30, determined by considering the FPS of the satellite video.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and the computer program executes the aircraft tracking method based on background-aware correlation filtering as described in any one of claims 1 to 4.

6. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the processor and the memory communicate with each other via a communication bus; the memory is used to store a computer program; the processor is used to implement the aircraft tracking method based on background perception correlation filtering as described in any one of claims 1 to 4 when executing the computer program stored in the memory.

Images