A correlation filtering target tracking method based on manifold background perception

Abstract

The invention discloses a correlation filtering target tracking method based on manifold background perception, which relates to the field of computer vision. An initial filter template is constructed by extracting the appearance features of a target; a double exponential distribution is established to construct an initial manifold search area; the manifold search area is adjusted according to the target movement speed and movement direction; the current manifold search area is determined according to the unchanged area of ​​the manifold search area; background information is extracted in the manifold search area to construct a correlation filtering algorithm model of manifold background perception; an augmented Lagrange multiplier method and an alternating direction multiplier method are adopted to solve the correlation filtering algorithm model of manifold background perception; an adaptive strategy is adopted to update the correlation filtering algorithm model of manifold background perception; a target tracking result is output; and the method has a dynamic search mechanism, can effectively cover the high-probability spatial range of target movement, solve the problem of target loss, effectively solve the problem of target drift in complex scenes, and improve the accuracy and speed of target tracking.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a correlation filtering target tracking method based on manifold background perception. Background Art

[0002] Target tracking is a research hotspot in the field of computer vision. Its essence is the process of analyzing and identifying the position and state of the target of interest in continuous video images based on prior information. Target tracking has important research and application value in the fields of video surveillance, modern military, human-computer interaction, and intelligent visual navigation. Due to factors such as low image resolution, target scale changes, and complex background in the tracking process, it is easy to lose the target. How to solve these problems and improve the accuracy of target tracking is a research hotspot and difficulty in the field of target tracking.

[0003] In 2010, Bolme et al. proposed the Minimum Output Sum of Squares Error (MOSSE) tracking algorithm. This algorithm converts the image from the time domain to the frequency domain through convolution calculation, converts the convolution operation into a dot product to improve the tracking speed, and judges the correlation between the initial position and the candidate position through the response value of the image to improve the accuracy of the filter. However, this algorithm uses a single-channel grayscale image to build a target appearance model. Since this model is too simple, the tracking effect is poor and it is impossible to accurately track the target. In 2012, Henriques et al. proposed the Kernel Cyclic Structure (CSK) algorithm. This algorithm calculates the correlation between two adjacent frames and maximizes the correlation between the initial and candidate positions. The response value is used as the predicted target center to improve the tracking effect, but the algorithm has a fixed target scale and is not robust to target scale changes; in 2014, Henriques et al. improved the CSK algorithm and proposed the kernel correlation filter algorithm (KCF), which enhances the performance of the tracker by constructing a circulant matrix to expand the number of negative samples and optimize the algorithm, but when the target moves quickly, the tracking effect is not obvious; in 2015, Danelljan et al. proposed a spatial regularization-based correlation filter algorithm (SRDCF), which can effectively suppress the response of the background area, expand the search range, and improve the tracking effect of the target in a complex background. In 2017, Galoogahi et al. proposed a background-aware correlation filter (BACF) algorithm. When the foreground and background of the target change, the algorithm expands the sampling area of ​​the circulant matrix and increases the number of samples. At the same time, the effective sample area is extracted on each sample to improve the sample quality, which greatly improves the effect and speed of the tracker. Although the background-aware correlation filter (BACF) algorithm increases the number of samples and ensures the sample quality, the algorithm is prone to target drift in complex scenes such as similar target and background information, occlusion, and fast motion. Therefore, how to effectively extract background information in the direction of target motion is crucial. Based on the background-aware correlation filter algorithm, this paper proposes a background-aware correlation filter target tracking algorithm based on manifold search. The algorithm uses a double exponential distribution to construct a manifold search area, and can dynamically adjust the search range and search angle of the manifold search area according to the target's motion speed and direction. It effectively extracts background samples for filter training and updates the filter template, which is of great significance to improving the performance of the background-aware correlation filter tracking algorithm. It has good accuracy, real-time and robustness for target tracking under complex conditions such as similar background, occlusion, fast motion, and motion blur. Summary of the invention

[0004] In order to solve the shortcomings of the prior art, the present invention proposes a correlation filtering target tracking method based on manifold background perception, which is divided into the following steps:

[0005] S1: Extract the appearance features of the target and construct the initial filter template;

[0006] S2: Establish a double exponential distribution and construct the initial manifold search area;

[0007] S3: Adjust the manifold search area according to the target movement speed and direction;

[0008] S4: Determine the current manifold search area according to the unchanged area of ​​the manifold search area;

[0009] S5: extracting background information in the manifold search area and constructing a manifold background perception correlation filtering algorithm model;

[0010] S6: The augmented Lagrange multiplier method and the alternating direction multiplier method are used to solve the correlation filtering algorithm model for manifold background perception;

[0011] S7: Adopting adaptive strategy to update the correlation filtering algorithm model of manifold background perception;

[0012] S8: Output target tracking results;

[0013] The specific process of S1 is: selecting a target to be tracked, extracting the appearance features of the target, establishing an initial target feature model, training a filter with initial background information to obtain an initial filter template;

[0014] The specific process of S2 is as follows: The mathematical model of the manifold search area is shown in formula (1):

[0015] y=±ae x (a>0) (1)

[0016] Among them, e is a natural constant, a is a coefficient, which is used to describe the slope of the curve. Two exponential distributions are connected back to back to form the background perception space range of the moving target, which is called the popular search area. The target position is set as the origin, and the function y = ±ae x The area enclosed by x=x1, x=x2 (x1<x2) is used as the initial search area to establish a plane rectangular coordinate system and construct an initial manifold search area;

[0017] The specific process of S3 is as follows: when the target moves to point Q(m,n) at time t, the target moves m units horizontally from the initial position (m is a positive number when the target moves to the right, and m is a negative number when the target moves to the left), and moves n units vertically from the initial position (n is a positive number when the target moves upward, and n is a negative number when the target moves downward), that is, the function is translated and its relative position changes. Then, when the target moves from the initial position to point Q, formula (1) can be transformed into:

[0018] y=±ae (x-m) +n (2)

[0019] The target motion speed is Angle x is the horizontal coordinate of the initial position of the target. If the target's moving speed increases, the target's moving direction is not easy to change. Therefore, the search range can be expanded in the direction of the target's moving direction and the search range can be reduced in the direction perpendicular to the target's moving direction, that is, a is reduced to a. i Similarly, if the target speed decreases, the target's direction of movement is easy to change. Therefore, the search range can be reduced in the direction of the target's movement and expanded in the direction perpendicular to the target's movement, that is, a is increased to a i ; Among them, a i Defined as:

[0020]

[0021] Where, v∈(0,+∞);

[0022] One of the search points C(p,q) is rotated counterclockwise by angle ω to C(p',q'), where p' and q' are:

[0023]

[0024] Where p is the horizontal coordinate of the search point, q is the vertical coordinate of the search point, p' is the horizontal coordinate of the search point after rotation, and q' is the vertical coordinate of the search point after rotation;

[0025] The specific process of S4 is as follows: the area of ​​the initial manifold search area S0:

[0026]

[0027] Among them, x1 is the horizontal coordinate of the left boundary of the initial manifold search area, and x2 is the horizontal coordinate of the right boundary of the initial manifold search area. When the target moves to point Q(m,n), x1'=x1+m. The area S of the manifold search area when the target moves to point Q is obtained by using definite integral. t :

[0028]

[0029] It can be seen from formula (6) that the manifold search area will change with the target motion state. In order to effectively control the computational complexity and make the computational amount of the manifold search not affected by the target motion state, the manifold search area when the target is at point Q is always equal to the manifold search area when the target is at the initial position, that is, S t =S0, the coordinates of the right boundary of the current manifold search area can be obtained by formula (5) and (6), as shown in formula (7):

[0030]

[0031] The specific process of S5 is as follows: The manifold search region of the moving target at the Q coordinate position can be re-determined using formulas (4) and (7). Finally, the background information of the search points is extracted to construct a correlation filtering algorithm model for manifold background perception.

[0032]

[0033] In formula (8), T represents the size of the picture, Y is the regression target, Y(u) is the u-th element of Y, h ∈ R D , h represents the training filter, h K is a multi-channel filter, H is the transpose, P is a two-dimensional matrix of D×T, used to extract D elements from the signal x, D is the target sample size. Usually, D << T, and P is a constant matrix that can be calculated. x k ∈R T (k = 1, 2...K), represents the extracted feature samples, Δτ v is the cyclic shift operation, λ is the regularization coefficient; γ is the regularization coefficient, is the target background information extracted from the manifold search region determined by formula (7), is the number of background blocks extracted using the manifold search algorithm, is the weight coefficient;

[0034] The specific process of S6 is as follows: Formula (8) is transformed to be calculated in the frequency domain, and the formula is:

[0035]

[0036] In formula (9), is the auxiliary variable, is a K×T-dimensional column vector, which is the concatenation of K vector channels g k The symbol ^ is the Fourier transform of the signal, is the Fourier transform of Y, is the target state at time t; F is a T×T orthogonal matrix, which is the mapping of the basis vectors of the T-dimensional vector signal in the Fourier domain, is the Kronecker product, I K is the K-order identity matrix, h = [h1 H ,..., h k H H is a K×D-dimensional column vector, which is the concatenation of K vector channels of h k cascade;

[0037] To solve formula (9), the augmented Lagrangian multiplier method is used to add a constraint term to the optimization function, that is: ​

[0038]

[0039] In the formula, is the Lagrangian vector of KT×1 in the Fourier domain, μ is the penalty factor; the alternating solution algorithm ADMM method is used to iteratively solve formula (10), transforming the target problem into two sub-problems g * and h * Solve, since both sub-problems are convex smooth differentiable functions, the global optimal solution can be obtained;

[0040] 1) Sub-problem 1: Solve h * , formula (10) is transformed into a sub-problem with a closed formula solution:

[0041]

[0042] Taking partial derivative of h in formula (11) we get:

[0043]

[0044] In formula (12), Let formula (12) equal to zero and solve it to get

[0045]

[0046] 2) Sub-problem 2: Solve for g * , formula (10) is transformed into a sub-problem with a closed formula solution:

[0047]

[0048] In formula (14), solve for g * To achieve real-time tracking, we need to solve g * The problem is divided into T independent objective functions for solution. Therefore, formula (14) can be decomposed into T linear sub-problems of K×K and solved separately. Depends only on and The K value is, conj(.) is the conjugate operation; according to the Pasval theorem, formula (14) is optimized to obtain

[0049]

[0050] In formula (15), and right Find the partial derivative

[0051]

[0052] Let formula (16) equal to zero to obtain

[0053]

[0054] The time complexity of formula (17) is O(TK 3 ), according to the Sherman-Morrison theorem, we get

[0055]

[0056] Substituting formula (18) into formula (17) we get

[0057]

[0058] In formula (19), yes The value after derivation is, b is an intermediate variable; compared with the time complexity of formula (17) O(TK 3 ), calculated using formula (19) The time complexity is O(TK), so real-time tracking can be achieved;

[0059] After solving the subproblem, solve for ξ and update the Lagrange multiplier vector to:

[0060]

[0061] In the formula, and Is to solve and The result of ADMM iteration u+1 times, where The parameter μ is μ u+1 =min(μ max ,θμ (u) )renew;

[0062] The specific process of S7 is: using an adaptive strategy to update the algorithm model, that is:

[0063]

[0064] In formula (21), η is the online adaptation rate, Instead of the formula (19) Used to solve

[0065] The specific process of S8 is: according to the adaptive strategy described in S7, the relevant filtering algorithm model of manifold background perception is updated, the filter template locates the target of the next frame, tracks the target, and outputs the tracking result.

[0066] Beneficial technical effects

[0067] 1. A correlation filtering target tracking method based on manifold background perception is proposed, which has a dynamic search mechanism and can effectively cover the high-probability spatial range of target motion and solve the problem of target loss;

[0068] 2. Use dynamic search mechanism for background perception, which can flexibly adjust the search range and angle according to the target's motion state, making the extracted background information richer and more accurate, enhancing the filter's ability to distinguish between the target and the background, and effectively solving the problem of target drift in complex scenes;

[0069] 3. Since the amount of manifold search data is small and the alternating solution algorithm is used to decompose the global problem into multiple sub-problems for rapid solution, the overall computational complexity of this method is low. While effectively controlling the amount of calculation, it also improves the accuracy and speed of target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A flow chart of a correlation filtering target tracking method based on manifold background perception provided by an embodiment of the present invention;

[0071] Figure 2 A schematic diagram of the probability of a manifold search area provided by an embodiment of the present invention;

[0072] Figure 3 A schematic diagram of a search point provided by an embodiment of the present invention;

[0073] Figure 4 A schematic diagram of the algorithm tracking effect provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0074] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments;

[0075] The present invention provides a correlation filtering target tracking method based on manifold background perception, such as Figure 1 As shown, the following steps are included:

[0076] S1: Select the target to be tracked, extract the appearance features of the target, establish the initial target feature model, and train the filter with the initial background information to obtain the initial filter template

[0077] S2: The mathematical model of the manifold search area is: y = ±ae x (a>0) (1)

[0078] In the formula, e is a natural constant and a is a coefficient, which is used to describe the slope of the curve. Two exponential distributions are connected back to back to form the background perception space range of the moving target, which is called the popular search area. The target position is set as the origin, and the function y = ±ae x The area enclosed by x=x1, x=x2 (x1<x2) is used as the initial search area to establish a plane rectangular coordinate system and construct the initial manifold search area

[0079] The target motion between any two adjacent frames is random, and the probability of the position of the random motion follows the normal distribution. This paper takes the normal distribution as an example to prove the effectiveness of the manifold search area, where the normal distribution function is:

[0080]

[0081] like Figure 2 As shown in Figure 2, assuming that the initial position of the target is the origin, the initial velocity of the target is a unit velocity of 1 pixel / s, and the moving direction is along the positive direction of the x-axis, then the probability of the target appearing in the next frame approximately follows the standard normal distribution, then μ = 0, σ = 1, and the probability distribution diagram is shown in Figure 2. Figure 2 The shaded part in the figure shows that (the normal distribution graph in the figure is rotated and placed in the manifold area;

[0082] Based on formula (2), the probability of the target appearing in the manifold perception area can be obtained as:

[0083]

[0084] In formula (3), e s and -e s All are functions of s. The integral of formula (3) is transformed into a double integral polar coordinate formula for solution. At this time,

[0085]

[0086] From formula (4), we can see that when s takes different values, the results of formula (3) are all close to 1, indicating that the probability of the target falling into the manifold search area is close to 1. It can be proved that the manifold perception method basically covers the high-probability spatial range of the target movement. At the same time, experiments are carried out using different numbers of moving particles. The positions of the experimental particles are randomly distributed, and the horizontal speed is also randomly distributed. It is verified by experiments that the shape of the curve of the vertical speed of the particles, whether it is randomly distributed or Gaussian random distribution, tends to Gaussian distribution. It can be seen that the initial distribution formula of the moving particle speed is not important, as long as it is random, so the speed factor can be ignored in the proof process. Therefore, it is theoretically feasible and effective to use formula (1) as the mathematical model of the manifold perception range.

[0087] Assume that the coordinates of the target position in the initial frame are the origin (0,0), and set the function y = ±ae x The area enclosed by x=x1, x=x2 (x1<x2) is used as the initial search area. When the target is at the initial position (0,0), the function x=x1 is symmetric with the function x=x2 about the y axis. In order to effectively extract the background sample, the function is performed in the search area:

[0088]

[0089]

[0090] The intersection of formula (5) and (6) As search points, where the intersection A total of indivual;

[0091] When the target is at the initial position (0,0), assuming a=1, x1=-2, x2=2, Δ1=0.8, Δ2=0.2, draw as well as The function graph of the search point diagram is as follows Figure 3 As shown, the dotted line in the figure is the function The graph of the function The image of the function in the figure is used as the search point (point C(p,q) in the figure is the search point If , there are 66 search points in total to extract background samples for filter training;

[0092] S3: When the target moves to point Q(m,n) at time t, the target moves m units horizontally from the initial position (m is a positive number when the target moves to the right, and m is a negative number when the target moves to the left), and moves n units vertically from the initial position (n is a positive number when the target moves upward, and n is a negative number when the target moves downward), that is, the function is translated and its relative position changes. Then, when the target moves from the initial position to point Q, formula (1) can be transformed into:

[0093] y=±ae (x-m) +n (7)

[0094] Assume that the target moving speed is Angle If the target's moving speed increases, the target's moving direction is not easy to change. Therefore, the search range can be expanded in the direction of the target's moving direction and the search range can be reduced in the direction perpendicular to the target's moving direction, that is, a is reduced to a iSimilarly, if the target speed decreases, the target's direction of movement is easy to change. Therefore, the search range can be reduced in the direction of the target's movement and expanded in the direction perpendicular to the target's movement, that is, a is increased to a i ; Among them, a i Defined as:

[0095]

[0096] One of the search points C(p,q) is rotated counterclockwise by angle ω to C(p',q'), where p' and q' are:

[0097]

[0098] In formula (8), v∈(0,+∞);

[0099] In order to verify the effectiveness of the manifold background-aware correlation filter tracking algorithm, this paper selects the video sequence images of the puppy movement in the dataset OTB100 for example testing. The tracking target is the whole body area of ​​the puppy. When the puppy runs sideways, its movement speed is fast and the angle changes dramatically. The puppy's body has scale expansion and shape changes, and its body color is similar to the background color. These interference factors make it difficult to track the puppy. The example test results of the manifold background-aware correlation filter tracking algorithm proposed in this paper are as follows: Figure 4 As shown;

[0100] Figure 4 The test results show that during the target's movement, the manifold background perception algorithm can dynamically adjust the scope and angle of the search area according to the target's movement speed and direction, cover the high-probability range of the target's movement, effectively perceive the target's background information, overcome the interference of complex factors on the dog's tracking, and accurately and stably track the target.

[0101] S4: The area of ​​the initial manifold search region S0:

[0102]

[0103] When the target moves to point Q(m,n), x1'=x1+m. The area S of the manifold search area when the target moves to point Q is obtained by using definite integral. t :

[0104]

[0105] It can be seen from formula (11) that the manifold search area will change with the target motion state. In order to effectively control the computational complexity and make the computational amount of the manifold search not affected by the target motion state, the manifold search area when the target is at point Q is always equal to the manifold search area when the target is at the initial position, that is, St = S0, which can be obtained from formulas (10) and (11)

[0106]

[0107] S5: Using formula (9) and formula (12), the manifold search region of the moving target at the Q coordinate position can be re-determined. Finally, the background information of the search points is extracted to construct a correlation filtering algorithm model for manifold background perception

[0108]

[0109] In formula (13), T represents the size of the picture, Y is the regression target, Y(u) is the u-th element of Y, h ∈ R D , h represents the training filter, h K is a multi-channel filter, H is the transpose, P is a two-dimensional matrix of D×T, used to extract D elements from the signal x, D is the size of the target sample. Usually, D << T, and P is a constant matrix that can be calculated, x k ∈ R T (k = 1, 2...K), representing the extracted feature samples, Δτ v is the cyclic shift operation, λ is the regularization coefficient; γ is the regularization coefficient, is the target background information extracted from the manifold search region determined by formula (12), is the number of background blocks extracted using the manifold search algorithm, is the weight coefficient and is related to the number of extracted background blocks. The more background blocks are selected, is closer to 1;

[0110] S6: Convert formula (13) to the frequency domain for calculation. The formula is:

[0111]

[0112] In formula (14), is the auxiliary variable, is a K×T dimensional column vector, which is the concatenation of K vector channels g k , the symbol ^ is the Fourier transform of the signal, is the Fourier transform of Y, is the target state at time t; F is a T×T orthogonal matrix, which is the mapping of the basis vectors of the T-dimensional vector signal in the Fourier domain, is the Kronecker product, I K is the K-order identity matrix, h = [h1 H ,..., h k H ​H is a K×D-dimensional column vector, h of K vector channels k cascade;

[0113] To solve formula (14), the augmented Lagrange multiplier method is used to add constraints to the optimization function, that is:

[0114]

[0115] In the formula, is the Lagrangian vector of KT×1 in the Fourier domain, μ is the penalty factor; the alternating solution algorithm ADMM method is used to iteratively solve formula (15), transforming the target problem into two sub-problems g * and h * Solve, since both sub-problems are convex smooth differentiable functions, the global optimal solution can be obtained;

[0116] 1) Sub-problem 1: Solve h * ; If in formula (15) λ, μ, γ, P are all known quantities, then formula (15) is transformed into a sub-problem with a closed formula solution;

[0117]

[0118] Taking the partial derivative of h in formula (16) we get:

[0119]

[0120] In formula (17), Let formula (17) equal to zero and solve

[0121]

[0122] 2) Sub-problem 2: Solve for g * , formula (15) is transformed into a sub-problem with a closed formula solution;

[0123]

[0124] In formula (19), solve for g * To achieve real-time tracking, we need to solve g * The problem is divided into T independent objective functions for solution. Therefore, formula (19) can be decomposed into T linear sub-problems of K×K and solved separately. Depends only on and The K value is, conj(.) is the conjugate operation; according to the Pasval theorem, formula (19) is optimized to obtain

[0125]

[0126] In formula (20), and right Find the partial derivative

[0127]

[0128] Let formula (21) equal to zero to obtain

[0129]

[0130] The time complexity of formula (22) is O(TK 3 ), according to the Sherman-Morrison theorem, we get

[0131]

[0132] Substituting formula (23) into formula (22) yields

[0133]

[0134] In formula (24), yes The value after derivation is, b is the intermediate variable;

[0135] Compared with the time complexity of formula (22) O(TK 3 ), calculated using formula (24) The time complexity is O(TK), so real-time tracking can be achieved;

[0136] After solving the subproblem, solve for ξ and update the Lagrange multiplier vector to:

[0137]

[0138] In the formula, and Is to solve and The result of ADMM iteration u+1 times, where The parameter μ is μ u+1 =min(μ max ,θμ (u) )renew;

[0139] S7: Adopt an adaptive strategy to update the algorithm model, namely:

[0140]

[0141] In formula (26), η is the online adaptation rate, Instead of the formula (24) Used to solve

[0142] S8: The programming software used in this paper is GNU Octave, the processor is Inteli7-6700CPU, the memory is 16GB; the operating system is Windows 10; the experimental parameters of this paper are set as follows: the number of HOG feature channels K = 31, the regularization factor λ = 0.01, γ = 1, the number of iterations L = 2, the penalty factor μ = 1, and the parameter μ uses μ i+1 =min(μ max ,θμ (i) ) update, μ max =10 3 ,θ=10; number of blocks

[0143] According to the adaptive strategy described in S7, the relevant filtering algorithm model of manifold background perception is updated, the filter template locates the target of the next frame, tracks the target, and outputs the tracking result;

[0144] The success rate and precision of the manifold search algorithm used in this paper reached 0.774 and 0.829 respectively, both higher than the comparison algorithm.

Claims

1. A correlation filtering target tracking method based on manifold background perception, characterized in that: It is divided into the following steps: S1: Extract the appearance features of the target and construct the initial filter template; S2: Establish a double exponential distribution and construct the initial manifold search area; The mathematical model of the manifold search area is shown in formula (1): y=±ae x (a>0) (1) Among them, e is a natural constant, a is a coefficient, which is used to describe the slope of the curve. Two exponential distributions are connected back to back to form the background perception space range of the moving target, which is called the manifold search area. The target position is set as the origin, and the function y = ±ae x The area enclosed by x=x1, x=x2 (x1<x2) is used as the initial search area to establish a plane rectangular coordinate system and construct an initial manifold search area; S3: Adjust the manifold search area according to the target movement speed and direction; The process of adjusting the manifold search area according to the target movement speed and movement direction is as follows: when the target moves to the Q(m,n) point at time t, the target moves m units horizontally from the initial position, where m is a positive number when the target moves to the right and m is a negative number when the target moves to the left, and moves n units vertically from the initial position, where n is a positive number when the target moves upward and n is a negative number when the target moves downward, that is, the function is translated and its relative position changes. Then, when the target moves from the initial position to the Q point, formula (1) can be transformed into formula (2): y=±ae (x-m) +n (2) The target motion speed is Angle x is the horizontal coordinate of the initial position of the target. If the target's moving speed increases, the target's moving direction is not easy to change. Therefore, the search range can be expanded in the direction of the target's moving direction and the search range can be reduced in the direction perpendicular to the target's moving direction, that is, a is reduced to a. i Similarly, if the target speed decreases, the target's direction of movement is easy to change. Therefore, the search range can be reduced in the direction of the target's movement and expanded in the direction perpendicular to the target's movement, that is, a is increased to a i ; Among them, a i Defined as: Where v∈(0,+∞), one of the search points C(p,q) is rotated counterclockwise by angle ω to C(p',q'), where p' and q' are: Where p is the horizontal coordinate of the search point, q is the vertical coordinate of the search point, p' is the horizontal coordinate of the search point after rotation, and q' is the vertical coordinate of the search point after rotation; S4: Determine the current manifold search area according to the unchanged area of ​​the manifold search area; S5: extracting background information in the manifold search area and constructing a manifold background perception correlation filtering algorithm model; S6: The augmented Lagrange multiplier method and the alternating direction multiplier method are used to solve the correlation filtering algorithm model for manifold background perception; S7: Adopting adaptive strategy to update the correlation filtering algorithm model of manifold background perception; S8: Output target tracking results.

2. The correlation filtering target tracking method based on manifold background perception as claimed in claim 1, characterized in that: The specific process of S1 is: selecting a target to be tracked, extracting the appearance features of the target, establishing an initial target feature model, and training a filter with initial background information to obtain an initial filter template.

3. The correlation filtering target tracking method based on manifold background perception as claimed in claim 1, characterized in that: The specific process of S4 is as follows: the area of ​​the initial manifold search area S0: Among them, x1 is the horizontal coordinate of the left boundary of the initial manifold search area, and x2 is the horizontal coordinate of the right boundary of the initial manifold search area. When the target moves to point Q(m,n), x1'=x1+m. The area S of the manifold search area when the target moves to point Q is obtained by using definite integral. t : It can be seen from formula (6) that the manifold search area will change with the target motion state. In order to effectively control the computational complexity and make the computational amount of the manifold search not affected by the target motion state, the manifold search area when the target is at point Q is always equal to the manifold search area when the target is at the initial position, that is, S t =S0, the coordinates of the right boundary of the current manifold search area can be obtained by formula (5) and (6), as shown in formula (7):

4. The correlation filtering target tracking method based on manifold background perception as claimed in claim 1, characterized in that: The specific process of S5 is as follows: the manifold search area of ​​the moving target at the Q coordinate position can be re-determined by using formula (4) and formula (7), and finally the background information of the search point is extracted to construct a manifold background perception correlation filtering algorithm model: where T represents the size of the image, Y is the regression target, Y(u) is the u-th element of Y, h ∈ R D , h represents the training filter, h K is a multi-channel filter, H is the transpose, P is a two-dimensional matrix of D×T, used to extract D elements from the signal x, D is the size of the target sample, D << T, and P is a constant matrix that can be calculated, x k ∈R T (k = 1, 2... K), represents the extracted feature samples, Δτ v is the cyclic shift operation, λ is the regularization coefficient; γ is the regularization coefficient, is the target background information extracted from the manifold search area determined by formula (7), is the number of background blocks extracted using the manifold search algorithm, is the weight coefficient.

5. The correlation filtering target tracking method based on manifold background perception as claimed in claim 1, characterized in that: The specific process of S6 is: converting formula (8) to the frequency domain for calculation, the formula is: in, is an auxiliary variable, is a K×T-dimensional column vector, which is K vector channels g k Cascade, symbol ^ is the Fourier transform of the signal, is the Fourier transform of Y, is the target state at time t; F is a T×T orthogonal matrix, which is the mapping of the T-dimensional vector signal to the basis vector in the Fourier domain. is the Kronecker product, I K is the K-order identity matrix, h=[h1 H ,...,h k H ] H is a K×D-dimensional column vector, h of K vector channels k cascade; To solve formula (9), the augmented Lagrange multiplier method is used to add constraints to the optimization function, that is: in, is the Lagrangian vector of KT×1 in the Fourier domain, μ is the penalty factor; the alternating solution algorithm ADMM method is used to iteratively solve formula (10), transforming the target problem into two sub-problems g * and h * Solve, since both sub-problems are convex smooth differentiable functions, the global optimal solution can be obtained; 1) Sub-problem 1: Solve h * , formula (10) is transformed into a sub-problem with a closed formula solution; Taking partial derivative of h in formula (11) we get: In formula (12), Let formula (12) equal to zero and solve it to get 2) Sub-problem 2: Solve for g * , formula (10) is transformed into a sub-problem with a closed formula solution; Among them, solving for g * To achieve real-time tracking, we need to solve g * The problem is divided into T independent objective functions for solution. Therefore, formula (14) can be decomposed into T linear sub-problems of K×K and solved separately. Depends only on and The K value is, conj(.) is the conjugate operation; according to the Pasval theorem, formula (14) is optimized to obtain: In formula (15), and right Find the partial derivative Let formula (16) equal to zero to obtain The time complexity of formula (17) is O(TK 3 ), according to the Sherman-Morrison theorem, we get Substituting formula (18) into formula (17) we obtain in, yes The value after derivation is, b is an intermediate variable; compared with the time complexity of formula (17) O(TK 3 ), calculated using formula (19) The time complexity is O(TK), so real-time tracking can be achieved; After solving the subproblem, solve for ξ and update the Lagrange multiplier vector to: In the formula, and Is to solve and The result of ADMM iteration u+1 times, where The parameter μ is μ u+1 =min(μ max ,θμ (u) )renew.

6. The correlation filtering target tracking method based on manifold background perception as claimed in claim 1, characterized in that: The specific process of S7 is: using an adaptive strategy to update the algorithm model, that is: In formula (21), η is the online adaptation rate, Instead of the formula (19) Used to solve 7. The correlation filtering target tracking method based on manifold background perception as claimed in claim 1, characterized in that: The specific process of S8 is: according to the adaptive strategy described in step 7, the relevant filtering algorithm model of manifold background perception is updated, the filter template locates the target of the next frame, tracks the target, and outputs the tracking result.

Images