A target tracking method for a high-speed rotating platform

Abstract

The application relates to a target tracking method for a high-speed rotating platform, which comprises the following steps: acquiring an initial frame image for target tracking; acquiring a next frame image, performing image rotation according to a rotation angle existing in the initial frame image, and obtaining a rotated image eliminating the rotation angle; taking a target point coordinate of a previous frame image as a center, adopting a correlation filtering model to perform target detection in the rotated image, and obtaining a target point coordinate in the rotated image; converting the target point coordinate in the rotated image to a target point coordinate in a pre-rotated image, and taking the target point coordinate in the pre-rotated image as a center to obtain a search region; updating the correlation filtering model, and performing target tracking on a next frame image; and further performing target tracking on all subsequent frame images. The application realizes stable tracking in the high-speed rotating process of the image.

Description

Technical Field

[0001] This invention belongs to the field of image processing and target tracking technology, and specifically relates to a target tracking method for a high-speed rotating platform. Background Technology

[0002] In recent years, the need for automatic tracking of moving targets in complex backgrounds has become increasingly urgent. Meanwhile, high-speed rotating platforms, as mounting platforms for photoelectric tracking devices, involve large-angle rotations during their photoelectric imaging process. The drastic rotation of the target in the image often leads to target tracking loss. Existing target tracking algorithms cannot meet these application conditions, significantly impacting the performance of guidance systems. Therefore, existing technologies urgently need research and improvement. Summary of the Invention

[0003] Based on the above analysis, the present invention aims to disclose a target tracking method for a high-speed rotating platform, which solves the problem of stable tracking during high-speed image rotation.

[0004] This invention discloses a target tracking method for a high-speed rotating platform, comprising:

[0005] Step S1: Obtain an initial frame image for target tracking; the initial frame image and subsequent frame images have a rotation angle;

[0006] Step S2: Obtain the next frame image, and rotate the image according to the rotation angle between the next frame image and the initial frame image to obtain the rotated image after eliminating the rotation angle;

[0007] Step S3: Using the target point coordinates of the previous frame image as the center, perform target detection in the rotated image using a correlation filtering model to obtain the target point coordinates in the rotated image;

[0008] Step S4: Transform the coordinates of the target point in the rotated image to the image before rotation to obtain the target point coordinates in the image before rotation; and calculate the line-of-sight angular velocity of the rotating platform as the tracking output signal based on the target point coordinates.

[0009] Step S5: Using the target point coordinates of the rotated image as the center, extract the search area and update the correlation filter model. Then, follow the method of steps S2-S4 to perform target tracking on the next frame of the image.

[0010] Repeat steps S2-S5 until target tracking is performed on all subsequent frame images.

[0011] Further, in step S2, the image is rotated according to the rotation angle relative to the initial frame image, and the pixel coordinates (x', y') of the rotated image after eliminating the rotation angle are obtained as follows: Where (x,y) are the pixel coordinates of the image before rotation; (xr ,y r ) represents the coordinates of the image rotation center, which is any point in the image before rotation; ξ is the image rotation angle.

[0012] Furthermore, step S3 includes:

[0013] 1) Interpolate the rotated image to obtain a grayscale image that includes the grayscale values ​​of each pixel in the image;

[0014] 2) Using the target point coordinates of the previous frame as the center, crop an image of a set size from the grayscale image, normalize it, and then extract the Hog features to obtain the feature matrix Z;

[0015] 3) The target coordinates in the rotated image are obtained by performing target detection in an image of a set size using a correlation filtering model.

[0016] Furthermore, in target detection using a correlation filtering model, the target point coordinates are obtained by calculating the filter response matrix R and searching for peak values ​​in the filter response matrix R.

[0017] in, F -1 This is the inverse Fourier transform; This is the Fourier transform of the correlation filter model A; λ is the Fourier transform of the Gaussian kernel function; λ is the regularization parameter. Fourier transform of the cross-correlation function; correlation function σ is the width parameter of the Gaussian kernel function; X is the template data of the correlation filter model; The conjugate of the Fourier transform of the template data; This is the Fourier transform of the characteristic matrix Z.

[0018] Furthermore, before the correlation filter model is updated, the initial value of the template data X of the correlation filter model is the feature matrix obtained by extracting Hog features after normalizing an image of a set size cropped from the grayscale image of the initial frame image with the target point coordinates of the initial frame image as the center.

[0019] The update formula for the template data X of the correlation filter model is X new = (1-α)*X old +α*X update Where α is the update rate; X old This is the template data from the previous correlation filter; X update The Hog feature matrix is ​​calculated by cropping the search region centered on the target point coordinates of the rotated image of the latest tracked frame.

[0020] Furthermore, before the correlation filter model is updated, the correlation filter model... The Fourier transform of the Gaussian kernel function; λ is the regularization parameter; K XX The autocorrelation function of the initial values ​​of the template data X;

[0021] The update formula for the correlation filter model is A. new = (1-α)*A old +α*A update Where α is the update rate; A old The correlation filter model for the previous correlation filter; A update To utilize the correlation filtering model obtained from the updated template data,

[0022] Furthermore, step S4 includes:

[0023] 1) Transform the coordinates of the target point in the rotated image to the original image to obtain the target point coordinates (X, Y, X) in the original image. T ,Y T );

[0024] 2) According to the formula The line-of-sight angular velocity of the rotating platform is calculated as the tracking output signal;

[0025] Among them, V x and V y Let X be the line-of-sight angular velocity in the X and Y directions. c and Y c R represents the coordinates of the center point of the image in the X and Y directions before rotation. x and R y These are the equivalent values ​​of the line-of-sight angular velocity in the X and Y directions.

[0026] Furthermore, the coordinates (X') of the target point in the rotated image are... T ,Y' T Transform the image to the original image to obtain the target point coordinates (X, Y, F, Z) in the original image. T ,Y T The formula for ) is:

[0027] Where (x,y) are the coordinates of a point in the image before rotation; (x',y') are the pixel coordinates of the image after rotation; (x r ,y r ) represents the coordinates of the image rotation center, which is any point in the image before rotation; ξ is the image rotation angle.

[0028] Furthermore, extracting Hog features from the image to form the feature matrix Z includes:

[0029] 1) Calculate the gradients in the X and Y directions, including the gradient magnitude and angle;

[0030] 2) Divide the data into cells and construct a gradient direction histogram for each cell;

[0031] 3) Normalized gradient histogram: Connect the normalized gradient histogram vectors together to obtain the Hog feature, forming the feature matrix Z.

[0032] Furthermore, a soft mapping strategy is adopted. When constructing the gradient direction histogram, for pixels located at the intersection of multiple cells, the gradient of each pixel is distributed to the adjacent cells by linear interpolation.

[0033] When normalizing the gradient histogram, each cell is normalized. Using the cell as the basic unit, 2*2 cells are integrated into 1 block, and the normalization factor of each block is calculated using the L2 norm of the undirected gradient histogram vector. The elements of each gradient histogram vector are normalized using the normalization factor, and the data is truncated using a set threshold. The average of the four values ​​obtained after normalization of each element is taken as the final result after normalization of the current element.

[0034] This invention can achieve at least one of the following beneficial effects:

[0035] 1) The method has a wide range of applications and can be universally applied to photoelectric tracking guidance in high-speed rotating platforms;

[0036] 2) The algorithm has low computational complexity and good real-time performance, and can meet the real-time computing requirements on commonly used embedded hardware platforms;

[0037] 3) It has good adaptability and can fully adapt to high-speed rotating platforms with a speed of not less than 100 degrees / second. Attached Figure Description

[0038] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0039] Figure 1 This is a flowchart of the target tracking method in an embodiment of the present invention. Detailed Implementation

[0040] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0041] One embodiment of the present invention discloses a target tracking method for a high-speed rotating platform, such as... Figure 1 As shown, it includes the following steps:

[0042] Step S1: Obtain an initial frame image for target tracking; the initial frame image and subsequent frame images have a rotation angle;

[0043] Step S2: Obtain the next frame image, and rotate the image according to the rotation angle between the next frame image and the initial frame image to obtain the rotated image after eliminating the rotation angle;

[0044] Step S3: Using the target point coordinates of the previous frame image as the center, perform target detection in the rotated image using a correlation filtering model to obtain the target point coordinates in the rotated image;

[0045] Step S4: Transform the coordinates of the target point in the rotated image to the image before rotation to obtain the target point coordinates in the image before rotation; and calculate the line-of-sight angular velocity of the rotating platform as the tracking output signal based on the target point coordinates.

[0046] Step S5: Using the target point coordinates of the rotated image as the center, extract the search area and update the correlation filter model. Then, follow the method of steps S2-S4 to perform target tracking on the next frame of the image.

[0047] Repeat steps S2-S5 until target tracking is performed on all subsequent frame images.

[0048] Specifically, since the target tracking is performed using a high-speed rotating platform in this embodiment, there is a certain image rotation angle between each frame due to the choice of platform.

[0049] Specifically, the calculation of the rotation angle in this embodiment includes:

[0050] 1) Based on the attitude data of the high-speed rotating platform, establish a coordinate system from point p in the world coordinate system. w (x w ,y w ,z w (to point p in the imaging coordinate system) m (x m ,y m ,z m The transformation matrix T:

[0051]

[0052] In the formula, φ, θ, and γ represent the imaging heading angle, pitch angle, and roll angle, respectively. -1 The transformation matrix from the imager coordinate system to the geodetic coordinate system is:

[0053] 2) Find the inverse matrix T of the transformation matrix T. -1 Obtain the transformation matrix from the imager coordinate system to the geodetic coordinate system;

[0054]

[0055] 3) Obtain the image rotation angle ξ:

[0056]

[0057] Specifically, in step S2, the image is rotated based on the rotation angle relative to the initial frame image, and the pixel coordinates (x', y') of the rotated image after eliminating the rotation angle are obtained as follows: Where (x,y) are the pixel coordinates of the image before rotation; (x r ,y r ) represents the coordinates of the image rotation center, which is any point in the image before rotation; ξ is the image rotation angle.

[0058] Specifically, step S3 includes:

[0059] Step S301: Interpolate the rotated image to obtain a grayscale image including the grayscale values ​​of each pixel in the image;

[0060] After rotation, the image coordinates are generally floating-point numbers, while the pixel coordinates are generally integers. Bilinear interpolation is used during image rotation, employing linear interpolation in two directions using the gray values ​​of the four surrounding neighboring points to obtain the gray value of the point to be sampled. In other words, the gray value of the point to be sampled is calculated by determining the corresponding weights based on the distance between the point to be sampled and its neighboring points. Here, (x, y) coordinates represent the pixel position, and f(x, y) represents the pixel's gray value. Its mathematical expression is:

[0061] f(i+u,j+v)=(1-u)(1-v)f(i,j)+(1-u)vf(i,j+1)+u(1-v)f(i+1,j)+uvf(i+1,j+1);

[0062] In the above formula, u and v take values ​​in the range of (0,1), representing the difference between the x and y coordinates of the sampling point and the top left corner of the four adjacent points; i and j represent the x and y coordinates of the sampling point and the top left corner of the four adjacent points.

[0063] Step S302: Using the target point coordinates of the previous frame image as the center, crop an image of a set size from the grayscale image, normalize it, and then extract the Hog features to obtain the feature matrix Z.

[0064] Step S303: Use a correlation filtering model to perform target detection in an image of a set size to obtain the coordinates of the target point in the rotated image.

[0065] Target capture of the initial frame image can be achieved through target edge extraction and target recognition, or by learning and training a classifier, or by utilizing existing target recognition techniques such as trained neural networks. No single existing target capture technique affects the inventiveness of the target tracking system of this invention.

[0066] Specifically, in step S302, the rotated image is I, and the initial size of the target is (w, h). An image I' of size (4w, 4h) is cropped from image I, centered on the target point coordinates of the previous frame. Image I' is then normalized to a 64*64 image I'. The steps for extracting Hog features from I' are as follows:

[0067] 1) Calculate the gradients in the X and Y directions, including the gradient magnitude and angle;

[0068] Gradient values ​​in the x and y directions:

[0069]

[0070] Gradient magnitude and gradient angle:

[0071]

[0072] In the above formula, f(x+1,y), f(x-1,y), f(x,y+1), and f(x,y-1) represent the grayscale values ​​of the corresponding pixels, respectively, and G... x (x,y) represents the gradient value in the x-direction at coordinates (x,y), G y (x,y) represents the gradient value in the y-direction at coordinates (x,y), G(x,y) is the gradient magnitude, and θ(x,y) is the gradient angle.

[0073] 2) Divide the data into cells and construct a gradient direction histogram for each cell;

[0074] Specifically, a soft mapping strategy is adopted. When constructing the gradient direction histogram, for pixels located at the intersection of multiple cells, the gradient of each pixel is distributed to the adjacent cells by linear interpolation.

[0075] More specifically, P1 to P4 represent the positions of four pixels, corresponding to the four cases of linear interpolation:

[0076] P1 is shared by 4 cells;

[0077] P2 is shared by the two cells above and below it;

[0078] P3 is shared by the left and right cells;

[0079] P4 is mapped to only one cell.

[0080] Mapping relationship of point P1:

[0081] Assuming the coordinates of P1 are (x, y), the formula for mapping the gradient histogram vector of point P1 to its four adjacent cells is:

[0082]

[0083] In the above formula, cellsize represents the number of pixels contained in each cell. and These represent rounding down a and b respectively, h p1 h represents the gradient histogram vector of point P1. A h B h C h D These represent the gradient histogram vectors of the four cells adjacent to point P1; preferably, the cellsize is 4.

[0084] The mapping formula between points P2 and P3 is:

[0085]

[0086] In the above formula, a and b are calculated in the same way as in the previous formula, h p2 h represents the gradient histogram vector of point P2. p3 h represents the gradient histogram vector of point P3. E h F h represents the gradient histogram vectors of the two cells adjacent to point P2, respectively. G h H These represent the gradient histogram vectors of the two cells adjacent to point P3 on the left and right, respectively.

[0087] The mapping formula for point P4 is:

[0088] h I +=h P4 ;

[0089] In the above formula, h p4 h represents the gradient histogram vector of point P4. I The gradient histogram vector represents the cell containing point P4.

[0090] 3) Normalized gradient histogram: Connect the normalized gradient histogram vectors together to obtain the Hog feature matrix.

[0091] Specifically, when normalizing the gradient histogram, each cell is normalized. Using the cell as the basic unit, 2*2 cells are integrated into 1 block, and the normalization factor of each block is calculated using the L2 norm of the undirected gradient histogram vector.

[0092] Cell unit E (coordinates (i,j)) spatially belongs to four blocks: B1, B2, B3, and B4. The normalization factor for each block is:

[0093]

[0094] In the above formula, δ and γ take values ​​of 1 or -1. The gradient histogram vector representing cell unit E, N δ,γ (i,j) is the block normalization factor at position (δ,γ) corresponding to cell unit E.

[0095] Preferably, when both δ and γ are 1, N δ,γ (i,j) is the normalization factor for cell unit E in the lower right block. The gradient histogram vector representing the cell unit to the right of cell unit E. The gradient histogram vector representing the cell unit below cell unit E. This represents the gradient histogram vector of the lower right cell unit of cell unit E.

[0096] The elements of each gradient histogram vector are normalized using a normalization factor, the data is truncated using a set threshold, and the average of the four normalized values ​​of each element is taken as the final normalized result of the current element.

[0097] Specifically, four normalization factors are used to normalize the elements of each gradient histogram vector, with a threshold of 0.2 used for truncation. The average of the four normalized values ​​for each element is then taken as the final normalized result for the current element, as shown in the formula below:

[0098]

[0099] In the above formula, This is the gradient histogram vector of cell units E in the k-th quadrant, with coordinates (i,j) after taking the mean and gradient direction divided by the gradient direction. N represents the gradient histogram vector of cell unit E with coordinates (i,j) before averaging; -1,-1 (i,j), N 1,-1 N -1,1 N 1,1 These represent the normalization factors for cell unit E in the top left, top right, bottom left, and bottom right blocks, respectively; T0.2 (x) represents threshold truncation of the variable x in parentheses, and average() represents averaging multiple values ​​in parentheses;

[0100] T 0.2 In (x), if x is less than or equal to 0.2, then the value is x; otherwise, the value is 0.2.

[0101] Furthermore, the undirected gradients normalized by different normalization factors are summed to obtain four gradient energies, which are used as a texture distribution feature of a cell, as shown in the following formula:

[0102]

[0103] In the above formula, nbins represents the number of quadrants divided by all gradient directions. This represents the sum of gradient values ​​in all directions; preferably, nbins is 6.

[0104] By concatenating the normalized histogram vectors mentioned above, we obtain the Hog feature matrix.

[0105] Specifically, in step S303, a correlation filtering model is used to perform target detection in an image of a set size using the extracted feature matrix; the target point coordinates (X') are obtained by searching for peak values ​​in the filter response matrix R. T ,Y' T ).

[0106] in, F -1 This is the inverse Fourier transform; This is the Fourier transform of the correlation filter model A; λ is the Fourier transform of the Gaussian kernel function; λ is the regularization parameter. The Fourier transform of the cross-correlation function is given; the regularization parameter λ is set to 0.1.

[0107] Related functions σ is the width parameter of the Gaussian kernel function; preferably, the σ of the Gaussian kernel function is set to 0.125. X is the template data of the correlation filter model. The conjugate of the Fourier transform of the template data; This is the Fourier transform of the characteristic matrix Z.

[0108] Specifically, step S4 includes:

[0109] 1) Transform the coordinates of the target point in the rotated image to the original image to obtain the target point coordinates (X, Y, X) in the original image. T ,Y T );

[0110] The coordinates (X') of the target point in the rotated image T ,Y T Transform the image to the original image to obtain the target point coordinates (X, Y) in the original image. T ,Y T The formula for ) is:

[0111] Where (x,y) are the coordinates of a point in the image before rotation; (x',y') are the pixel coordinates of the image after rotation; (x r ,y r ) represents the coordinates of the image rotation center, which is any point in the image before rotation; ξ is the image rotation angle.

[0112] 2) According to the formula The line-of-sight angular velocity of the rotating platform is calculated as the tracking output signal;

[0113] Among them, V x and V y Let X be the line-of-sight angular velocity in the X and Y directions. c and Y c R represents the coordinates of the center point of the image in the X and Y directions. x and R y These are the equivalent values ​​of the line-of-sight angular velocity in the X and Y directions.

[0114] Specifically, the update of the correlation filter model in step S5 includes template data update and correlation model update.

[0115] Template data updates include:

[0116] Before the correlation filter model is updated, the initial value of the template data X of the correlation filter model is obtained by extracting the Hog features from the grayscale image of the initial frame image with the target point coordinates as the center, normalizing the image, and then extracting the target point coordinates.

[0117] The update formula for the template data X of the correlation filter model is X new = (1-α)*X old +α*X update Where α is the update rate; X old This is the template data from the previous correlation filter; X update The Hog feature matrix is ​​calculated by cropping the search region centered on the target point coordinates of the rotated image of the latest tracked frame.

[0118] The relevant model updates include:

[0119] Before the correlation filter model was updated, the correlation filter model was... The Fourier transform of the Gaussian kernel function; λ is the regularization parameter; K XX The autocorrelation function of the initial values ​​of the template data X;

[0120] The update formula for the correlation filter model is A. new = (1-α)*A old +α*A update Where α is the update rate; A old The correlation filter model for the previous correlation filter; A update To obtain the relevant filtering model using the updated template data,

[0121] α is the update rate, and its value is set to 0.012.

[0122] For example, when performing correlation filtering on the image following the initial frame, the correlation filtering model is in its pre-update state, and the template data X of the correlation filtering model is its initial value; the correlation filtering model

[0123] After tracking the initial frame image and the subsequent frame image, when tracking the last frame image, the template data and correlation model are updated. At this time, the template data of the correlation filter model is updated, making the template data X... new = (1-α)*X old +α*X update ; where X old X is the template data from the previous correlation filter. update The Hog feature matrix is ​​calculated by cropping the search region centered on the target point coordinates of the rotated image following the initial frame image. Correlation filtering model A new = (1-α)*A old +α*A update Where α is the update rate; A old The correlation filter model for the previous correlation filter; A update To utilize the updated template data X new The correlation filtering model is obtained.

[0124] Following the above update process, target tracking is performed on all subsequent frame images using the updated correlation filter model.

[0125] In summary, this invention achieves stable tracking during high-speed image rotation.

[0126] Furthermore, the method of this invention has a wide range of applications and can be universally applied to photoelectric tracking guidance in high-speed rotating platforms; the algorithm has low computational complexity and good real-time performance, and can meet the real-time computing requirements on commonly used embedded hardware platforms; it has good adaptability and can fully adapt to high-speed rotating platforms with a speed of not less than 100 degrees / second.

[0127] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A target tracking method for a high-speed rotating platform, characterized in that, include: Step S1: Obtain an initial frame image for target tracking; the initial frame image and subsequent frame images have a rotation angle; Step S2: Obtain the next frame image, and rotate the image according to the rotation angle between the next frame image and the initial frame image to obtain the rotated image after eliminating the rotation angle; Step S3: Using the target point coordinates of the previous frame image as the center, perform target detection in the rotated image using a correlation filtering model to obtain the target point coordinates in the rotated image; In target detection using a correlation filtering model, the filter response matrix is ​​calculated. ; In the filter response matrix The target point coordinates are obtained by searching for the peak value. in, ; This is the inverse Fourier transform; For the correlation filter model Fourier transform; ; The Fourier transform of the Gaussian kernel function; For regularization parameters; For related functions Fourier transform; the correlation function ; is the width parameter of the Gaussian kernel function; Template data for the correlation filtering model; The conjugate of the Fourier transform of the template data; Characteristic matrix Fourier transform; Step S4: Transform the coordinates of the target point in the rotated image to the image before rotation to obtain the target point coordinates in the image before rotation; and calculate the line-of-sight angular velocity of the rotating platform as the tracking output signal based on the target point coordinates. Step S5: Using the target point coordinates of the rotated image as the center, extract the search area and update the correlation filter model. Then, follow the method of steps S2-S4 to perform target tracking on the next frame of the image. Repeat steps S2-S5 until target tracking is performed on all subsequent frame images.

2. The target tracking method according to claim 1, characterized in that, In step S2, the image is rotated according to the rotation angle relative to the initial frame image to obtain the pixel coordinates of the rotated image after the rotation angle is eliminated. for: ;in, These are the pixel coordinates of the image before rotation; Let be the coordinates of the image rotation center, and be any point in the image before rotation; This represents the image rotation angle.

3. The target tracking method according to claim 1, characterized in that, Step S3 includes: 1) Interpolate the rotated image to obtain a grayscale image that includes the grayscale values ​​of each pixel in the image; 2) Using the target point coordinates of the previous frame as the center, crop a grayscale image of a set size, normalize it, and then extract the Hog features to obtain the feature matrix. ; 3) The target coordinates in the rotated image are obtained by performing target detection in an image of a set size using a correlation filtering model.

4. The target tracking method according to claim 3, characterized in that, Template data of the correlation filter model before the model update The initial value is the feature matrix obtained by extracting Hog features after normalizing an image of a set size from the grayscale image of the initial frame image, with the target point coordinates of the initial frame image as the center. Correlation filter model template data The update formula is ;in, For update rate; This is the template data from the previous correlation filter; The Hog feature matrix is ​​calculated by cropping the search region centered on the target point coordinates of the rotated image of the latest tracked frame.

5. The target tracking method according to claim 4, characterized in that, Before the correlation filter model was updated, the correlation filter model was... ; The Fourier transform of the Gaussian kernel function; For regularization parameters; Template data The autocorrelation function of the initial values; The update formula for the correlation filter model is: ;in, For update rate; The correlation filter model from the previous correlation filter; To utilize the correlation filtering model obtained from the updated template data, .

6. The target tracking method according to claim 3, characterized in that, Step S4 includes: 1) Transform the coordinates of the target point in the rotated image to the original image to obtain the target point coordinates in the original image. ; 2) According to the formula ; Calculate the line-of-sight angular velocity of the rotating platform as the tracking output signal; in, and The line-of-sight angular velocities are the X and Y directions. and These are the coordinates of the center point of the image in the X and Y directions before rotation. and These are the equivalent values ​​of the line-of-sight angular velocity in the X and Y directions.

7. The target tracking method according to claim 6, characterized in that, The coordinates of the target point in the rotated image By converting to the image before rotation, we can obtain the coordinates of the target point in the image before rotation. The formula is: ;in, The coordinates of the points in the image before rotation: These are the pixel coordinates of the rotated image; Let be the coordinates of the image rotation center, and be any point in the image before rotation; This represents the image rotation angle.

8. The target tracking method according to claim 3, characterized in that, Extracting Hog features from an image to form the feature matrix Z includes: 1) Calculate the gradients in the X and Y directions, including the gradient magnitude and angle; 2) Divide the data into cells and construct a gradient direction histogram for each cell; 3) Normalized gradient histogram: Connect the normalized gradient histogram vectors together to obtain the Hog features, forming the feature matrix Z.

9. The target tracking method according to claim 8, characterized in that, A soft mapping strategy is adopted. When constructing the gradient direction histogram, for pixels located at the intersection of multiple cells, the gradient of each pixel is distributed to the adjacent cells by linear interpolation. When normalizing the gradient histogram, each cell is normalized. Using the cell as the basic unit, 2*2 cells are integrated into 1 block, and the normalization factor of each block is calculated using the L2 norm of the undirected gradient histogram vector. The elements of each gradient histogram vector are normalized using the normalization factor, and the data is truncated using a set threshold. The average of the four values ​​obtained after normalization of each element is taken as the final result after normalization of the current element.

Images