Satellite video dark vehicle tracking method based on adaptive label updating
By using technologies such as adaptive tag update and Retinex local constraints in satellite videos, the problem of difficult tracking of dark vehicles in low-light environments is solved, the visibility and tracking accuracy of vehicles in low-light conditions is improved, and background interference is overcome The impact of
Patent Information
- Application Number
- CN202411158487.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-08-22
AI Technical Summary
In low-light environments, the target characteristics of dark and weak vehicles in satellite videos are weak, have low contrast, and have low signal-to-missive ratios, which are easily disturbed by background, resulting in low tracking accuracy.
Using the low-light satellite video dark and weak vehicle tracking method based on adaptive tag update, the target position is smoothly estimated through single-index smoothly, and the Retinex local constraints and target energy constraints are constructed. Combined with adaptive tag updates, a tracking filter is constructed, and the target energy is redistributed. Enhance target visibility and suppress background interference.
It improves the tracking performance of satellite videos in low-light environments, enhances the visibility of the target under low-light conditions, reduces the impact of background interference, and significantly improves the tracking accuracy and success rate.
Smart Images

Figure CN119048558A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of remote sensing image processing and relates to target tracking of satellite video, in particular to a method for tracking a weak vehicle in satellite video under low light conditions. Background Art
[0002] Object tracking from satellite video is a key research area. However, it faces significant challenges in low-light environments, especially when dealing with dim vehicles. Previous research has focused on tracking methods under favorable lighting conditions, ignoring the complexities introduced by insufficient lighting. Extracting features from objects in low-light environments is more difficult, and these challenges are exacerbated by the sensitivity of dim objects to background noise. In low-light environments, dim vehicles exhibit weak features, low contrast, and low signal-to-noise ratio, making them more susceptible to background interference.
[0003] At present, the existing target tracking methods in the field of satellite video mainly focus on problems such as occlusion, target similarity, and background interference, and mainly focus on bright targets under favorable lighting conditions, without considering the problem of tracking dim targets in low-light environments. In low-light conditions, dim vehicle target features are weaker and more susceptible to background interference. Therefore, tracking dim vehicles in low-light environments is still a problem to be solved. Summary of the invention
[0004] The purpose of the present invention is to solve the problem that the target features of dim vehicles in low-light environments are weak, the signal-to-noise ratio is low, and they are more susceptible to background interference, resulting in low vehicle tracking accuracy. A method for tracking dim vehicles in low-light environments using satellite videos based on adaptive label updates is proposed.
[0005] The satellite video dim vehicle tracking method based on adaptive label update includes the following process:
[0006] Step 1: Estimate the predicted position of the moving target in the satellite video based on exponential smoothing SES;
[0007] Step 2: Construct Retinex local constraints and target energy constraints, and construct an optimization function based on the Retinex local constraints and target energy constraints;
[0008] Step 3: Build a basic tracking framework and adaptive labels, and build a regression target based on the basic tracking framework and adaptive labels;
[0009] Step 4: Solve the regression target constructed in step 3 to obtain the tracking filter.
[0010] The beneficial effects of the present invention are:
[0011] In order to solve the problem of tracking weak vehicles in satellite videos in low-light environments, this paper proposes a correlation filter tracking method embedded with low-light target enhancement. The target visibility is enhanced by redistributing energy through target motion characteristics. In addition, an adaptive label update is proposed to suppress background interference to reduce the impact of background interference. The performance of tracking weak vehicle targets in satellite videos in low-light environments is improved.
[0012] The present invention proposes a method for tracking weak vehicles in satellite video under low light based on adaptive label update. The low-light target enhancement is constructed by using the target motion characteristics and embedded in the correlation filter tracking algorithm. In addition, based on the adaptive label update, a weak vehicle tracking framework is constructed to effectively alleviate the background interference problem and improve the performance of satellite video weak vehicle tracking.
[0013] The present invention proposes a method for tracking weak vehicles in satellite videos under low light based on adaptive label updating, and embeds low-light target enhancement into a correlation filter tracking method. The target visibility is enhanced by redistributing the target energy through the target motion characteristics. In addition, a weak vehicle tracking framework is constructed based on adaptive label updating to suppress the influence of background interference and improve the weak vehicle tracking performance in satellite videos. The implementation process is as follows: (1) input satellite video data, and construct a motion estimation module based on single exponential smoothing; (2) construct a low-light target enhancement module based on intrinsic decomposition; (3) construct a weak vehicle target tracking framework based on adaptive label updating; (4) optimize and solve the tracking filter based on the weak vehicle tracking framework. The present invention can improve the performance of weak vehicle target tracking in low-light environments of satellite videos, and the method is simple and easy to implement.
[0014] The present invention aims to solve the problem of difficulty in tracking weak vehicle targets in satellite videos in low-light environments. Weak target features, low contrast, low signal-to-noise ratio, and severe background interference lead to tracking drift. Combining the target motion characteristics in satellite videos, a satellite video weak vehicle tracking framework is constructed, and a correlation filtering tracking method embedded in low-light target enhancement is proposed. The background interference is suppressed by adaptive label updating, which greatly improves the performance of weak target tracking in low-light environments.
[0015] In order to verify the performance of the algorithm proposed in the present invention, experiments were conducted on video data captured by Jilin-1 video. The experimental results verified the effectiveness of the method for tracking dim vehicles in satellite videos under low light based on adaptive label updating proposed in the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a flow chart of the implementation of the present invention;
[0017] Figure 2 This is the original satellite video data of Jilin-1;
[0018] Figure 3 is the tracking accuracy graph;
[0019] Figure 4 It is a graph of tracking success rate. DETAILED DESCRIPTION
[0020] Specific implementation method 1: The specific process of the satellite video dim vehicle tracking method based on adaptive label update in this implementation method is as follows:
[0021] Step 1: Estimate the predicted position of the moving target in the satellite video based on exponential smoothing SES;
[0022] Step 2: Construct Retinex local constraints and target energy constraints, and construct an optimization function based on the Retinex local constraints and target energy constraints;
[0023] Step 3: Build a basic tracking framework and adaptive labels, and build a regression target based on the basic tracking framework and adaptive labels;
[0024] Step 4: Solve the regression target constructed in step 3 to obtain the tracking filter.
[0025] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that the predicted position of the moving target in the satellite video is estimated based on the exponential smoothing SES in step 1; the specific process is:
[0026] Moving objects in satellite videos usually exhibit regular trajectories, such as vehicles and trains traveling along specific routes. In this case, motion information from the previous frame is crucial for accurate object localization, especially for vehicle targets with linear smooth sequences. Therefore, single exponential smoothing (SES) is chosen to predict the target position and provide a priori position for target enhancement.
[0027] The superscript e indicates that the position P and translation vector Δ are estimated using single exponential smoothing SES. The superscript t indicates that they are derived from the true determined position.
[0028] Estimating the predicted position of the kth frame of a moving target in satellite video using single exponential smoothing (SES) The expression is:
[0029]
[0030]
[0031] in, It indicates that the predicted position of the kth frame of the moving target in the satellite video is estimated using single exponential smoothing SES;
[0032] Indicates the true position of the moving target in the (k-1)th frame in the satellite video;
[0033] represents the translation vector of the kth frame of the moving target in the satellite video estimated using single exponential smoothing SES;
[0034] represents the true translation vector of the moving target in the (k 1)th frame in the satellite video;
[0035] represents the translation vector of the moving target in the (k-1)th frame in the satellite video estimated using single exponential smoothing SES;
[0036] α represents the smoothness constraint;
[0037] The predicted position of the kth frame of the moving target in the satellite video estimated using single exponential smoothing SES Centered on the satellite video, w is used as the width of the moving target, h is used as the height of the moving target in the satellite video, and the predicted position pixel set M of the moving target in the kth frame in the satellite video estimated using single exponential smoothing SES is obtained; w, h are known quantities, which come from the initial position information of the target;
[0038] The target position estimated by SES provides the basic position information required for the subsequent target enhancement process. In addition, using the predicted position as the center of the search area can effectively reduce the impact of background noise. Simple target trajectory prediction is effective for vehicle tracking in satellite videos. By directly predicting the target motion, the tracker can better estimate the future target position. This prediction strategy is particularly effective in satellite videos and improves vehicle tracking performance.
[0039] The other steps and parameters are the same as those in the first embodiment.
[0040] Specific implementation method three: This implementation method is different from specific implementation method one or two in that in step 2, Retinex local constraints and target energy constraints are constructed, and an optimization function is constructed based on the Retinex local constraints and target energy constraints; the specific process is:
[0041] An efficient image intrinsic decomposition method is used to obtain enhanced images, which mitigates the impact of low-light conditions on target features. Compared with traditional low-light image enhancement methods, this method uniquely integrates target position information from motion estimation. By imposing target energy constraints and redistributing target energy, the enhancement requirements of dim vehicles in satellite videos can be more effectively met.
[0042] To improve computational efficiency, this module adopts two related constraints: the Retinex local constraint related to the eigendecomposition theory and the moving target energy constraint related to the target position information. This streamlined constraint configuration enables the tracker to effectively handle target enhancement.
[0043] Step 21: Construct Retinex local constraints; the specific process is:
[0044] According to the eigendecomposition theory, the eigendecomposition divides the image into reflection component and shadow component.
[0045] The loss function can be formulated as the sum of the differences between the reflectance and shading of surrounding pixels.
[0046] Retinex Local Constraint E r The definition is as follows:
[0047]
[0048] Among them, E r (ρ,L) represents the Retinex local constraint; L(p) represents the brightness component of pixel p, L(q) represents the brightness component of pixel q, ρ(p) represents the reflectivity component of pixel p, ρ(q) represents the reflectivity component of pixel q, ω (p,q) represents the balance factor, (p, q) represents the adjacent pixel pair, N represents the set of adjacent pixel pairs; ρ and L represent the reflectance component and the brightness component respectively; in the logarithmic domain, the intrinsic decomposition model I = ρ + L is satisfied, where I represents the search area image;
[0049] Balance factor ω (p,q) By chromaticity is defined as:
[0050]
[0051] Where t is the chromaticity threshold, and denote the chromaticity of pixels p and q respectively; ‖‖2 denotes the two-norm;
[0052] Chroma is an objective color quality criterion used to estimate pixel reflectance;
[0053] Significant changes in chromaticity usually indicate changes in reflectance;
[0054] The decomposition of shading and reflectivity is given by the balancing factor ω (p,q) Determined;
[0055] Step 22: Construct the target energy constraint; the specific process is:
[0056] Intrinsic decomposition separates the image into reflectance and shadow components, dispersing the energy of the original signal. To enhance the focus of the target signal energy, a target energy constraint is introduced using the position information of motion estimation. This constraint can help focus the target energy, thereby enhancing the expressiveness of the target and achieving the desired improvement. Specifically, it aims to minimize the difference between the shading components of the target and the surrounding background. Aligning the shadow component of the target more closely with the background helps to effectively focus the energy in the reflectance.
[0057] Target energy constraint E t Defined as:
[0058]
[0059] Among them, E t (L) represents the energy constraint of the moving target, L(m) represents the light and dark components of the target pixel m, and m represents the The estimated target pixel, M represents the set of predicted position pixels of the moving target in the kth frame of the satellite video estimated using single exponential smoothing SES, L(n) represents the light and dark components of background pixel n, and n represents the background pixel closest to the target pixel m;
[0060] Step 23: Construct an optimization function based on the Retinex local constraint and the target energy constraint;
[0061] Step 24: Optimize and solve the optimization function constructed in step 23.
[0062] The other steps and parameters are the same as those in the first or second embodiment.
[0063] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that in step 23, an optimization function is constructed based on the Retinex local constraint and the target energy constraint; the specific process is:
[0064] The optimization function expression is:
[0065]
[0066] Among them, E(L|I) represents the overall optimization function, λ r represents the target energy constraint penalty coefficient, λ t Represents the Retinex local constraint penalty coefficient.
[0067] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.
[0068] Specific implementation method five: This implementation method is different from any one of the specific implementation methods one to four in that in step 24, the optimization function constructed in step 23 is optimized and solved; the specific process is:
[0069] Based on the eigendecomposition model I = ρ + L in the logarithmic domain, substituting ρ = IL into expression (6), we obtain:
[0070]
[0071] Where I(p) represents the image of pixel p in the search area, I(q) represents the image of pixel q in the search area, ΔL represents the difference between L(p) and L(q), ΔL = L(p) - L(q), ΔI represents the difference between I(p) and I(p), ΔI = I(p) - I(q);
[0072] Based on the polynomial combination in formula (7), the quadratic matrix of the light and dark components L is constructed. The quadratic matrix is as follows:
[0073]
[0074] Among them, A and b in the quadratic matrix are respectively composed of the coefficients of the quadratic term and the coefficients of the linear term in the polynomial of formula (7);
[0075] L T represents the transpose of L, b T represents the transpose of b, c is a constant term;
[0076] Since L(p), L(q) and L(m), L(n) in equation (7) are subject to spatial constraints, the matrix A is a positive definite symmetric matrix;
[0077] Since the quadratic matrix Where A is a positive definite symmetric matrix, and the quadratic matrix is solved according to the convex optimization algorithm Quadratic Matrix The analytical solution is:
[0078] AL=b(9)
[0079] Given matrices A and b, find L;
[0080] Through the light and dark component L, the reflectivity component ρ is obtained:
[0081] ρ=IL (10).
[0082] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.
[0083] Specific implementation method 6: This implementation method is different from any one of the specific implementation methods 1 to 5 in that a basic tracking framework and an adaptive label are constructed in step 3, and a regression target is constructed based on the basic tracking framework and the adaptive label; the specific process is:
[0084] Step 31: Based on the reflectivity component ρ, a basic tracking framework is constructed; the specific process is as follows:
[0085] Spatiotemporal Regularized Correlation Filter (STRCF) is a basic tracking framework based on correlation filtering. The tracking performance is improved by introducing a temporal regularization term. The objective function of the basic tracking framework is described as follows:
[0086]
[0087] in, represents the objective function of the basic tracking framework; x k represents the target sample, represents the dth feature in the target sample, D represents the number of features, d represents a feature in D, * represents the convolution operation, and f d ∈R K×1 represents the filter of the dth feature, R represents a real number, K represents the number of samples, y represents the Gaussian label, w represents the spatial regularization weight, · represents the Hadamard product, μ represents the time regularization hyperparameter, f represents the filter currently to be solved, and f k-1 represents the solved k-1th frame filter; || || 2 represents the square of the norm;
[0088] Step 32: Build adaptive labels;
[0089] Step 33: Based on the basic tracking framework and adaptive labels, construct a regression target.
[0090] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.
[0091] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that an adaptive tag is constructed in step 32; the specific process is as follows:
[0092] Most correlation filter-based trackers rely on fixed regression labels during training, usually using Gaussian functions to represent target responses. However, filters trained with fixed labels are difficult to adapt to environmental interference. Since the interference source and the target have similar appearance, the fixed regression target cannot effectively distinguish them, usually resulting in model drift or tracking failure. In addition, a large portion of the labels in the ideal Gaussian labeling have zero values, indicating that the background information is not fully utilized, thus affecting the tracker performance.
[0093] In order to overcome the limitation of fixed regression labels, an adaptive label update mechanism is introduced. This mechanism uses the historical response map to dynamically adjust the background labels, effectively reducing the impact of background interference on the tracker performance.
[0094] The adaptive label update mechanism is as follows:
[0095] Λ=R k [ψΔ ]-R k-1 [ψ Δ2 ] (12)
[0096] u=y-βB T (e Λ -1) (13)
[0097] Among them, R k and R k-1 are the filter responses of the kth frame and the k-1th frame respectively;
[0098] ψ Δ1 is the filter response R k The target tracking position offset, ψ Δ2 is the filter response R k-1 Target tracking position offset;
[0099] [·] indicates a circular shift operation;
[0100] R k [ψ Δ1 ] represents the filter response R k Cyclic displacement ψ Δ1 , R k-1 [ψ Δ2 ] represents the filter response R k-1 Cyclic displacement ψ Δ2 , the shifted filter response R k [ψ Δ1 ] and R k-1 [ψ Δ2 ]’s peaks are aligned;
[0101] Λ represents the difference between the filter responses after alignment;
[0102] B refers to the cropping matrix, which is used to keep the training label of the target center area unchanged;
[0103] β is the penalty factor parameter;
[0104] y is the Gaussian label;
[0105] u is the adjusted adaptive label.
[0106] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.
[0107] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that in step 33, a regression target is constructed based on a basic tracking framework and adaptive labels; the specific process is as follows:
[0108] The regression target is expressed as:
[0109]
[0110] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.
[0111] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that the regression target constructed in step 3 is solved in step 4 to obtain the tracking filter optimization; the specific process is:
[0112] The solution efficiency is accelerated and the operation speed is increased by using the alternating direction multiplier method (ADMM).
[0113] Step 41: Introduce g=f, and write formula (14) as an extended Lagrangian formula:
[0114]
[0115] in, represents the extended Lagrangian objective function;
[0116] g represents an auxiliary variable; g d represents the dth feature in the auxiliary variable;
[0117] s represents the Lagrange multiplier, s=[s 1 ...,s d ...,s D ]∈R K×D ,s 1 represents the first characteristic of the Lagrange multiplier, s d represents the d-th characteristic of the Lagrange multiplier, s D represents the Dth feature of the Lagrange multiplier, D represents the number of features, and K represents the number of samples;
[0118] γ is the step size parameter;
[0119] * represents the convolution operation, · represents the Hadamard product;
[0120] Step 42: By introducing Formula (15) is simplified to:
[0121]
[0122] in, represents the extended Lagrangian objective function; h represents the Lagrangian multiplier; h d represents the d-th characteristic of the Lagrange multiplier;
[0123] Step 43: Use the ADMM method to decompose formula (16) into three sub-problems.
[0124] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.
[0125] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that the ADMM method is used in step 43 to decompose formula (16) into three sub-problems, as shown below:
[0126]
[0127] Among them, f j+1 represents the j+1th iteration solution of filter f; * represents the convolution operation; g j+1 represents the solution result of the auxiliary variable g at the j+1th iteration; h j+1 represents the j+1th iteration solution of the Lagrange multiplier h;
[0128] Subproblem f: Subproblem f has a high computational complexity. In order to effectively reduce the complexity of the algorithm, the calculation of subproblem f is converted to the frequency domain through Fourier transform to improve the calculation speed;
[0129] right Perform Fourier transform to obtain the sub-problem f after Fourier transform; the expression is:
[0130]
[0131] in, represents the filter f after Fourier transformation; represents the sample characteristics after Fourier transformation, represents the Hadamard product, Represents the Fourier transformed The d-th feature of represents the adaptive label after Fourier transformation, represents the auxiliary variable after Fourier transformation, represents the Lagrange multiplier after Fourier transformation, represents the filter of the k-1th frame after Fourier transformation;
[0132] where the superscript ∧ is the discrete Fourier transform identifier.
[0133] It can be seen from formula (18) that the adaptive label after Fourier transform The i-th element of depends only on D features and filters on The corresponding i-th element of . Represents the vector All D features of the i-th element are taken out and used as a separate vector. Formula (18) is independently decomposed into K sub-problems according to the number of elements. Each sub-problem is described as follows:
[0134]
[0135] in, Representation sample The vectorized vector of the i-th element of ; Representation filter The vectorization vector of the i-th element of Represents adaptive label The vectorization vector of the i-th element of Represents auxiliary variables The vectorization vector of the i-th element of represents the Lagrange multiplier The vectorized vector of the i-th element of Represents the k-1th frame filter The vectorized vector of the i-th element of ;
[0136] When the derivative of equation (19) is 0 The solution, The closed-form solution is:
[0137]
[0138] Where I represents the identity matrix; the superscript T represents the vector transpose;
[0139] The Sherman-Morison formula is used to circumvent the computationally intensive matrix inversion operation, thereby transforming formula (20) into:
[0140]
[0141] Subproblem g: Each element of g can be calculated independently and is easy to solve directly in the time domain;
[0142] The solution of g is obtained by Taking the derivative, we get:
[0143]
[0144] Where w represents the spatial weight matrix, It is the diagonal matrix of the spatial weight matrix w, diag() represents the diagonalization of the matrix, and K represents the number of samples;
[0145] Subproblem h: Based on the solved f and g, through h j+1 =h+f (j+1) -g (j+1) To update and obtain h.
[0146] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.
[0147] Experiment designed by the present invention:
[0148] The data used in the experiment is a set of Jilin-1 satellite video image data in low-light environments. Figure 2 It is the original Jilin-1 satellite video data; Figure 3 is the tracking accuracy graph; Figure 4 is the tracking success rate diagram, and the two diagrams are tracking accuracy indicators, reflecting the tracking effect; Table 1 shows the target tracking accuracy of the method of the present invention. It can be seen that the method proposed by the present invention can effectively improve the tracking performance of dim vehicle targets.
[0149] Table 1 Target tracking accuracy of the method of the present invention in original data
[0150]
[0151] 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 dim vehicle tracking method based on adaptive label update, characterized by: The method comprises the following steps: Step 1: Estimate the predicted position of the moving target in the satellite video based on exponential smoothing SES; Step 2: Construct Retinex local constraints and target energy constraints, and construct an optimization function based on the Retinex local constraints and target energy constraints; Step 3: Build a basic tracking framework and adaptive labels, and build a regression target based on the basic tracking framework and adaptive labels; Step 4: Solve the regression target constructed in step 3 to obtain the tracking filter.
2. The satellite video dim vehicle tracking method based on adaptive label update according to claim 1 is characterized in that: In step 1, the predicted position of the moving target in the satellite video is estimated based on the exponential smoothing SES; the specific process is: Estimating the predicted position of the kth frame of a moving target in satellite video using single exponential smoothing (SES) The expression is: in, It indicates that the predicted position of the kth frame of the moving target in the satellite video is estimated using single exponential smoothing SES; represents the real position of the moving target in the (k1)th frame in the satellite video; represents the translation vector of the kth frame of the moving target in the satellite video estimated using single exponential smoothing SES; represents the true translation vector of the moving target in the (k-1)th frame in the satellite video; represents the translation vector of the moving target in the satellite video frame (k1) estimated using single exponential smoothing SES; α represents the smoothness constraint; The predicted position of the kth frame of the moving target in the satellite video estimated using single exponential smoothing SES With w as the width of the moving target in the satellite video and h as the height of the moving target in the satellite video, the predicted position pixel set M of the kth frame of the moving target in the satellite video estimated using single exponential smoothing SES is obtained.
3. The satellite video dim vehicle tracking method based on adaptive label update according to claim 2 is characterized in that: In step 2, a Retinex local constraint and a target energy constraint are constructed, and an optimization function is constructed based on the Retinex local constraint and the target energy constraint; the specific process is: Step 21: Construct Retinex local constraints; the specific process is: The Retinex local constraint Er is defined as follows: Among them, E r (ρ,L) represents the Retinex local constraint; L(p) represents the brightness component of pixel p, L(q) represents the brightness component of pixel q, ρ(p) represents the reflectivity component of pixel p, ρ(q) represents the reflectivity component of pixel q, ω (p,q) represents the balance factor, (p, q) represents the adjacent pixel pair, N represents the set of adjacent pixel pairs; ρ and L represent the reflectance component and the brightness component respectively; in the logarithmic domain, the intrinsic decomposition model I = ρ + L is satisfied, where I represents the search area image; Balance factor ω (p,q) By chromaticity is defined as: Where t is the chromaticity threshold, and denote the chromaticity of pixels p and q respectively; ‖ ‖2 denotes the two-norm; Step 22: Construct the target energy constraint; the specific process is: Target energy constraint E t Defined as: Among them, E t (L) represents the energy constraint of the moving target, L(m) represents the light and dark components of the target pixel m, and m represents the The estimated target pixel, M represents the set of predicted position pixels of the moving target in the kth frame of the satellite video estimated using single exponential smoothing SES, L(n) represents the light and dark components of background pixel n, and n represents the background pixel closest to the target pixel m; Step 23: Construct an optimization function based on the Retinex local constraint and the target energy constraint; Step 24: Optimize and solve the optimization function constructed in step 23.
4. The satellite video dim vehicle tracking method based on adaptive label update according to claim 3 is characterized in that: In step 23, an optimization function is constructed based on the Retinex local constraint and the target energy constraint; the specific process is: The optimization function expression is: Among them, E(L|I) represents the overall optimization function, λ r represents the target energy constraint penalty coefficient, λ t Retinex local constraint penalty coefficient.
5. The satellite video dim vehicle tracking method based on adaptive label update according to claim 4 is characterized in that: In step 24, the optimization function constructed in step 23 is optimized and solved; The specific process is: Based on the eigendecomposition model I = ρ + L in the logarithmic domain, substituting ρ = IL into expression (6), we obtain: Where I(p) represents the image of pixel p in the search area, I(q) represents the image of pixel q in the search area, ΔL represents the difference between L(p) and L(q), ΔL = L(p) - L(q), ΔI represents the difference between I(p) and I(p), I = I(p) - I(q); Based on the polynomial combination in formula (7), the quadratic matrix of the light and dark components L is constructed. The quadratic matrix is as follows: Among them, A and b in the quadratic matrix are respectively composed of the coefficients of the quadratic term and the coefficients of the linear term in the polynomial of formula (7); represents the transpose of L, represents the transpose of b, and c is a constant term; The matrix A is a positive definite symmetric matrix; Since the quadratic matrix Where A is a positive definite symmetric matrix, and the quadratic matrix is solved according to the convex optimization algorithm Quadratic Matrix The analytical solution is: AL=b(9) Through the light and dark component L, the reflectivity component ρ is obtained: ρ=IIL (10).
6. The satellite video dim vehicle tracking method based on adaptive label update according to claim 5 is characterized in that: In step 3, a basic tracking framework and adaptive labels are constructed, and a regression target is constructed based on the basic tracking framework and the adaptive labels. The specific process is as follows: Step 31: Based on the reflectivity component ρ, a basic tracking framework is constructed; the specific process is as follows: The objective function of the basic tracking framework is described as follows: in, represents the objective function of the basic tracking framework; x k represents the target sample, represents the dth feature in the target sample, D represents the number of features, d represents a feature in D, * represents the convolution operation, and f d ∈R K×1 represents the filter of the dth feature, R represents a real number, K represents the number of samples, y represents the Gaussian label, w represents the spatial regularization weight, · represents the Hadamard product, represents the temporal regularization hyperparameter, f represents the filter currently to be solved, and f k1 represents the solved k1th frame filter; || || 2 represents the square of the norm; Step 32: Build adaptive labels; Step 33: Based on the basic tracking framework and adaptive labels, construct the regression target.
7. The satellite video dim vehicle tracking method based on adaptive label update according to claim 6 is characterized in that: In step 32, an adaptive tag is constructed; the specific process is as follows: The adaptive label update mechanism is as follows: L=R k [ψ Δ1 ]-R k-1 [ψ2] (12) Among them, R k and R k-1 are the filter responses of the kth frame and the k-1th frame respectively; ψ Δ1 is the filter response R k The target tracking position offset, ψ Δ2 is the filter response R k-1 Target tracking position offset; [·] indicates a circular shift operation; R k [ψ Δ1 ] represents the filter response R k Cyclic displacement ψ Δ1 , R k1 [ψ Δ2 ] represents the filter response R k-1 Cyclic displacement ψ Δ2 , the shifted filter response R k [ψ Δ1 ] and R k-1 [ψ Δ2 ]’s peaks are aligned; Λ represents the difference between the filter responses after alignment; B refers to the clipping matrix; β is the penalty factor parameter; y is the Gaussian label; u is the adjusted adaptive label.
8. The satellite video dim vehicle tracking method based on adaptive label update according to claim 7 is characterized in that: In step 33, a regression target is constructed based on the basic tracking framework and the adaptive label; the specific process is as follows: The regression target is expressed as:
9. The satellite video dim vehicle tracking method based on adaptive label update according to claim 8, characterized in that: In step 4, the regression target constructed in step 3 is solved to obtain the tracking filter optimization; The specific process is: Step 41: Introduce g=f, and write formula (14) as an extended Lagrangian formula: in, represents the extended Lagrangian objective function; g represents an auxiliary variable; g d represents the dth feature in the auxiliary variable; s represents the Lagrange multiplier, s=[s 1 ...,s d ...,s] D ∈R K×D ,s 1 represents the first characteristic of the Lagrange multiplier, s d represents the d-th characteristic of the Lagrange multiplier, s D represents the Dth feature of the Lagrange multiplier, D represents the number of features, and K represents the number of samples; γ is the step size parameter; * represents the convolution operation, · represents the Hadamard product; Step 42: By introducing Formula (15) is simplified to: in, represents the extended Lagrangian objective function; h represents the Lagrangian multiplier; h d represents the d-th characteristic of the Lagrange multiplier; Step 43: Use the ADMM method to decompose formula (16) into three sub-problems.
10. The satellite video dim vehicle tracking method based on adaptive label update according to claim 9, characterized in that: In step 43, the ADMM method is used to decompose formula (16) into three sub-problems, as shown below: Among them, f j+1 represents the j+1th iteration solution of filter f; * represents the convolution operation; g j+1 represents the solution result of the auxiliary variable g at the j+1th iteration; h j+1 represents the j+1th iteration solution of the Lagrange multiplier h; Sub-problem f: right Perform Fourier transform to obtain the sub-problem f after Fourier transform; the expression is: in, represents the filter f after Fourier transformation; represents the sample characteristics after Fourier transformation, represents the Hadamard product, Represents the Fourier transformed The d-th feature of represents the adaptive label after Fourier transformation, represents the auxiliary variable after Fourier transformation, represents the Lagrange multiplier after Fourier transformation, represents the filter of the k1th frame after Fourier transformation; Formula (18) is independently decomposed into K sub-problems according to the number of elements. Each sub-problem is described as follows: in, Representation sample The vectorized vector of the i-th element of ; Representation filter The vectorization vector of the i-th element of Represents adaptive label The vectorization vector of the i-th element of Represents auxiliary variables The vectorized vector of the i-th element of represents the Lagrange multiplier The vectorization vector of the i-th element of Represents the k1th frame filter The vectorized vector of the i-th element of ; When the derivative of equation (19) is 0 The solution, The solution is: Where I represents the identity matrix; the superscript T represents the vector transpose; Convert formula (20) to: Sub-problem g: The solution of g is obtained by Taking the derivative, we get: Where w represents the spatial weight matrix, It is the diagonal matrix of the spatial weight matrix w, diag() represents the diagonalization of the matrix, and K represents the number of samples; Subproblem h: Based on the solved f and g, through h j+1 =h+f (j+1) -g (j+1) To update and obtain h.
Citation Information
Patent Citations
Retinex low-illumination color image enhancement method based on variation constraint
CN109872285A
Satellite video single target tracking method based on motion perception and correlation filtering
CN117830349A
Target tracking method, system, device and storage medium
US11821986B1
