A correlation filtering target tracking method, system and storage medium

Abstract

The present invention relates to a correlation filtering target tracking method, system, and storage medium, belonging to the field of computer vision technology. The method solves the problems of existing technologies in tracking targets in complex scenes, high computational cost, and poor real-time performance. The method includes: using the current correlation filtering model and the frequency domain features of the current frame image to perform a correlation operation to obtain a first response map and determine the current target position; using the current target position as a reference, segmenting the current frame image to obtain a new window image and extracting its frequency domain features; optimizing the objective function using the first response map and the frequency domain features of the new window image to obtain the correlation filtering model corresponding to the next frame image; using the correlation filtering model corresponding to the next frame image and its frequency domain features to perform a correlation operation to obtain a second response map, and obtaining the target position of the next frame image based on the second response map. Based on the frequency domain features of the image, the method can track targets in complex environments, with low computational cost and high real-time performance.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a correlation filtering target tracking method, system and storage medium. Background Art

[0002] The target tracking algorithm based on correlation filtering applies the correlation filtering technology in the communication field to the target tracking field, converting the time-consuming correlation calculations in the airspace to frequency domain calculations, which greatly improves the tracking speed and enables the target tracking algorithm to meet the real-time requirements of various applications. It has not only been widely used in the military field, such as drone tracking and precision guidance, but also plays an important role in the civilian field, such as intelligent video surveillance, human-computer interaction, and robot visual navigation.

[0003] In the prior art, early correlation filter-based tracking algorithms focused on improving target features and scale. Regarding target features, the earliest MOSSE tracking algorithm used a single grayscale feature, while the KCF algorithm added HOG features, and the SAMF algorithm combined HOG features with CN features. Regarding target scale, the SAMF algorithm scaled the target by seven scales, while the DSST algorithm combined scale-space and translation correlation filters. On the other hand, correlation filter-based target tracking algorithms also face the problem of boundary effects. This is because they primarily use cyclic shifts to obtain dense training samples and frequency domain computation for acceleration. These shifted image patches are implicitly generated based on the cyclic characteristics of correlation in the frequency domain and serve as negative samples for training the filter. All shifted image patches are subject to boundary effects and are not truly representative of negative samples in real scenes. To address this shortcoming, prior art methods mainly expand the search area through the following methods. Among them, the SRDCF algorithm introduces a background regularization term when training the correlation filter to enable learning within a larger search area. Unfortunately, this method is computationally expensive and reduces the tracking algorithm speed. The BACF algorithm uses a cropping matrix to densely extract image blocks from the background, expanding the search area at a lower computational cost.

[0004] The existing technology has at least the following flaws: First, the SRDCF algorithm introduces a background regularization term when training the correlation filter, resulting in high computational cost, reduced speed of the tracking algorithm, and poor real-time performance. However, the BACF algorithm expands the search area, which increases the number of samples, but also introduces more background noise when there is too much background information. The correlation filter is based on spatial domain features and is more likely to learn from the environment rather than the target. In this case, coupled with changes in the appearance of the object caused by complete or partial occlusion, changes in lighting, etc., it will make the detection process more prone to deviations, thereby greatly reducing the credibility of the target tracking results. Summary of the Invention

[0005] In view of the above analysis, the present invention aims to provide a correlation filtering target tracking method, system and storage medium to solve the problems that the existing technology is difficult to track targets in complex environments, and has high computational costs, poor real-time performance and low credibility of target tracking results.

[0006] In one aspect, the present invention provides a correlation filtering target tracking method, comprising:

[0007] The current correlation filter model is used to perform correlation operation with the frequency domain features of the current frame image to obtain the first response map, thereby determining the current target position;

[0008] Segmenting the current frame image to obtain a new window image based on the current target position, and extracting frequency domain features of the new window image;

[0009] Optimizing the objective function using the frequency domain features of the first response graph and the new window image to obtain a correlation filter model corresponding to the next frame image;

[0010] A correlation operation is performed using the correlation filter model corresponding to the next frame image and the frequency domain features of the next frame image to obtain a second response map, and the target position of the next frame image is obtained according to the second response map.

[0011] Furthermore, extracting the frequency domain features of the new window image includes:

[0012] Performing Rees transform on the new window image to obtain a corresponding frequency domain image matrix;

[0013] Performing an inverse Fourier transform on the frequency domain image matrix to obtain a spatial domain image matrix corresponding to the new window image;

[0014] Obtaining a monoscopic image matrix corresponding to the new window image according to the spatial domain image matrix and the grayscale matrix corresponding to the new window image;

[0015] The frequency domain features of the new window image are extracted from the monoscopic image matrix, where the frequency domain features include local amplitude, local direction and local phase.

[0016] Furthermore, the monoscopic image matrix is ​​expressed by the following formula:

[0017] I M =I-iR 1(X) -jR 2(X) ,

[0018] in,

[0019]

[0020] In the above formula, R1(X) 、R 2(X) Represent the two components in the single-image matrix, I represents the grayscale matrix corresponding to the new window image, A represents the local amplitude of the new window image, represents the local phase of the new window image, and θ represents the local direction of the new window image.

[0021] Furthermore, the local amplitude, local direction and local phase are respectively:

[0022]

[0023] Furthermore, optimizing the objective function using the frequency domain features of the first response graph and the new window image to obtain a correlation filter model corresponding to the next frame image specifically includes:

[0024] Establish an objective function, which is:

[0025]

[0026] Among them, y represents the ideal response map corresponding to the current frame image, y∈R N , D represents the total number of channels corresponding to the image, d represents the dth channel, Represents the frequency domain features of the d-th channel corresponding to the k-th frame new window image, Represents the d-th channel correlation filter model corresponding to the k-th frame new window image, B represents the clipping matrix, B∈R M×N , N represents the length of the vector corresponding to each frequency domain feature, M represents the number of central elements of the vector corresponding to each frequency domain feature, λ represents the spatial regularization parameter, * represents the correlation operation, [ψ p,q ] represents a shift operation that makes the peak value of the response graph of the current frame image coincide with the peak value of the response graph of the previous frame image;

[0027] An optimization algorithm is used to solve the objective function, and the correlation filtering model that minimizes the objective function value is used as the correlation filtering model corresponding to the next frame image.

[0028] Furthermore, the correlation filtering model corresponding to the next frame image is:

[0029]

[0030] in, λ represents the spatial regularization parameter, μ represents the penalty factor, I D represents a D×D two-dimensional matrix, F represents the discrete Fourier transform matrix, W k Represents the correlation filter model corresponding to the k-th frame image.

[0031] Furthermore, obtaining the target position of the next frame image according to the second response map includes: determining the position of the maximum response value of the second response map as the target position of the next frame image.

[0032] In another aspect, the present invention provides a correlation filtering target tracking system, comprising:

[0033] A target position determination module is used to perform a correlation operation using the current correlation filter model and the frequency domain features of the current frame image to obtain a first response map, thereby determining the current target position;

[0034] A frequency domain feature extraction module is used to segment the current frame image to obtain a new window image based on the current target position and extract frequency domain features of the new window image;

[0035] A correlation filter model updating module, configured to optimize an objective function using the frequency domain features of the first response graph and the new window image to obtain a correlation filter model corresponding to a next frame of image;

[0036] The target position determination module is further configured to perform correlation calculations on the frequency domain features of the next frame image using the correlation filter model corresponding to the next frame image to obtain a second response map, and obtain the target position of the next frame image according to the second response map.

[0037] Furthermore, the frequency domain feature extraction module is specifically used to:

[0038] Performing an inverse Fourier transform on the frequency domain image matrix to obtain a spatial domain image matrix corresponding to the new window image;

[0039] Obtaining a monoscopic image matrix corresponding to the new window image according to the spatial domain image matrix and the grayscale matrix corresponding to the new window image;

[0040] Extract the local amplitude, local direction and local phase of the new window image from the single image matrix, and then form the frequency domain features of the new window image

[0041] On the other hand, the present invention provides a storage medium for storing computer instructions, and a processor executes the computer program to implement the aforementioned correlation filtering target tracking method.

[0042] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0043] 1. The correlation filtering target tracking method and system proposed in the present invention creatively propose to update the correlation filtering model according to the frequency domain characteristics of the image, and perform target tracking based on the correlation filtering model, wherein the local phase contains most of the image information, and the higher the time domain and spatial domain frequencies, the smaller the role of the local amplitude in the signal, and the more prominent the role of the phase. Therefore, the image reconstructed based on the local phase will enhance and highlight edges, lines and other narrow structures, and will not be affected by external brightness and noise changes. Compared with spatial domain image features, it can more accurately describe image features. Therefore, target tracking based on the local phase can greatly improve the accuracy of target tracking.

[0044] 2. The correlation filtering target tracking method and system proposed in the present invention only describes the image through three features: local amplitude, local phase and local direction. It has low computational complexity and high computational cost. It avoids the problem of large computational complexity and high computational cost caused by describing the image through spatial domain features such as image grayscale features, HOG features and CN features in the existing technology, and can greatly improve the real-time performance of target tracking.

[0045] 3. The correlation filtering target tracking method and system proposed in the present invention suppress the distortion of the response map by adding frequency domain image features and response map distortion regularization terms, thereby improving the image feature description and improving the boundary effect, reducing the impact of background noise, and making it possible to quickly and accurately track targets based on images with scenes such as background clutter, rapid motion, motion blur, occlusion, and deformation.

[0046] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0048] Figure 1 Flowchart of the correlation filtering target tracking method according to an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of any original image in an embodiment of the present invention;

[0050] Figure 3 The embodiment of the present invention uses local amplitude to Figure 2 A schematic diagram representing any original image shown;

[0051] Figure 4 The embodiment of the present invention uses local phase Figure 2 A schematic diagram representing any original image shown;

[0052] Figure 5 The embodiment of the present invention uses local direction Figure 2 A schematic diagram representing any original image shown;

[0053] Figure 6 Schematic diagram of a correlation filtering target tracking system according to an embodiment of the present invention.

[0054] Reference numerals:

[0055] 110 - target position determination module; 120 - frequency domain feature extraction module; 130 - correlation filter model update module. DETAILED DESCRIPTION

[0056] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0057] Method Example

[0058] A specific embodiment of the present invention discloses a correlation filtering target tracking method. Figure 1 As shown, the method includes:

[0059] S110 , performing a correlation operation with the frequency domain features of the current frame image using the current correlation filter model to obtain a first response map corresponding to the current frame image, and using the position corresponding to the maximum response value in the first response map as the current target position.

[0060] S120: Segment the current frame image based on the current target position to obtain a new window image, and extract frequency domain features of the new window image. For example, if the target in the current frame image is a face with a size of 10 cm * 10 cm, the new window image is segmented from the current frame image at a certain ratio. Specifically, in the current frame image, a 12 cm * 12 cm window is selected as the new window image with the current target position as the center. This allows for accurate face extraction while reducing image size, thereby increasing computation speed.

[0061] S130 , optimizing the objective function using the first response graph and the frequency domain features of the new window image to obtain a correlation filter model corresponding to the next frame of image.

[0062] S140 , performing a correlation operation on the frequency domain features of the next frame of image using the correlation filter model corresponding to the next frame of image to obtain a second response map, and obtaining the target position of the next frame of image according to the second response map.

[0063] Preferably, the next frame image is sequentially used as the current frame image, and steps S110 to S140 are repeated, thereby achieving real-time tracking of the target.

[0064] In the existing technology, spatial domain image features, namely grayscale features, HOG features and CN features are used to describe images. However, the HOG features include hundreds of feature channels, which are computationally intensive and costly, resulting in poor real-time target tracking.

[0065] In order to overcome the above defects, considering that the frequency domain features of the image can enhance and highlight the edges, lines and other narrow structures of the image, and are not affected by external brightness and noise changes, compared with the spatial domain image features, they can more accurately describe the image features. Therefore, the present invention creatively proposes to use frequency domain features to describe the image.

[0066] Preferably, extracting the frequency domain features of the new window image includes:

[0067] S1201, perform Rees transform on the new window image to obtain the corresponding frequency domain image matrix, which is specifically in the form of:

[0068]

[0069] Where ω=[ω1,ω2], expressed as frequency domain variables, (R1(ω),R2(ω)) T Represents the frequency domain image matrix, I F The matrix representing the Fourier transform of the grayscale matrix corresponding to the new window image.

[0070] S1202, perform inverse Fourier transform on the frequency domain image matrix to obtain the spatial domain image matrix corresponding to the new window image, which is specifically in the form of:

[0071]

[0072] Where x = [x, y], represents the spatial variable, (R1(x), R2(x)) T Represents the spatial domain image matrix, and I represents the grayscale matrix corresponding to the new window image.

[0073] S1203 : Obtain a monoscopic image matrix corresponding to the new window image according to the spatial domain image matrix and the grayscale matrix corresponding to the new window image.

[0074] S1204. Extract frequency domain features of the new window image from the single-image matrix. The frequency domain features include local amplitude, local direction, and local phase. These three frequency domain features can accurately characterize the characteristics of the image, i.e., three feature channels. Therefore, the amount of calculation is small and the calculation cost is high, which can greatly improve the real-time performance and accuracy of target tracking.

[0075] Preferably, the monoscopic image matrix is ​​represented by the following formula:

[0076] I M =I-iR 1(X) -jR 2(X) ,

[0077] in,

[0078]

[0079] In the above formula, R 1(X) 、R 2(X) Represent the two components in the single-image matrix, I represents the grayscale matrix corresponding to the new window image, A represents the local amplitude of the new window image, represents the local phase of the new window image, and θ represents the local direction of the new window image.

[0080] Preferably, the local amplitude, local direction and local phase are respectively:

[0081]

[0082] For example, the local amplitude, local phase, and local direction are used to analyze any original image (such as Figure 2 The corresponding representation images are respectively as shown in Figures 3 to 5 As shown. Figure 3 It can be seen from the figure that the local amplitude can accurately characterize the image, but when the frequency in the spatial domain and frequency domain is higher, the local amplitude’s characterization of the image features weakens and cannot accurately describe the image information. Figure 4 It can be seen from the above that the local phase can contain most of the information of the image and accurately describe the image. Moreover, the representation of the image information by the local phase is not affected by the changes in frequency and image brightness, and the representation is stable, thus improving the accuracy and reliability of target tracking. Figure 5 It can be seen from the figure that the local direction has a relatively weak effect on the representation of image information, but it can still roughly describe the image information. Therefore, the combination of local amplitude and local phase can accurately describe the image information.

[0083] Preferably, the objective function is optimized using the frequency domain features of the first response graph and the new window image to obtain a correlation filter model corresponding to the next frame image, specifically including:

[0084] S1301, establish an objective function, which is:

[0085]

[0086] Among them, y represents the ideal response map corresponding to the current frame image, y∈R N , D represents the total number of channels corresponding to the image, that is, the three feature channels corresponding to the three frequency domain features, d represents the dth channel, Represents the frequency domain features of the d-th channel corresponding to the k-th frame new window image, Represents the d-th channel correlation filter model corresponding to the k-th frame new window image, B represents the clipping matrix, B∈R M×N , N represents the length of the vector corresponding to each frequency domain feature, M represents the number of central elements of the vector corresponding to each frequency domain feature, λ represents the spatial regularization parameter, * represents the correlation operation, [ψ p,q ] represents a shift operation that makes the peak value of the response graph of the current frame image coincide with the peak value of the response graph of the previous frame image.

[0087] Preferably, the ideal response map refers to an ideal response map determined based on the first response map and the current target position. Specifically, the response value corresponding to the current target position in the first response map is set to 1, and the response values ​​corresponding to other positions are set to 0, thereby obtaining an ideal response map. According to the ideal response map, the current position information can be integrated into the relevant filtering model corresponding to the next frame image, thereby improving the accuracy of target tracking.

[0088] S1302: Using an optimization algorithm to solve the objective function, and using the correlation filter model that minimizes the objective function value as the correlation filter model corresponding to the next frame image.

[0089] Specifically, the objective function is converted into augmented Lagrangian form:

[0090]

[0091] in, Indicates M k-1 [ψ p,q ] discrete Fourier transform, M k-1 represents the first response image corresponding to the k-1th frame image, ^ represents discrete Fourier transform, represents the Lagrange auxiliary vector, that is, the Lagrange vector in the Fourier domain; X k for General representation of .

[0092] Applying the ADMM method, the objective function solution can be transformed into the solution of the following two sub-problems:

[0093]

[0094] Then, Decompose into N sub-problems to solve:

[0095]

[0096] in, μ and γ represent penalty factors, and n = [1, 2, 3...N].

[0097] The Lagrange auxiliary variables are updated as follows:

[0098]

[0099] During target tracking, by updating And Lagrangian auxiliary variables are used to update the relevant filter model and solve The relevant filtering model corresponding to the next frame image is obtained as follows:

[0100]

[0101] in, λ represents the spatial regularization parameter, μ represents the penalty factor, I D represents a D×D two-dimensional matrix, F represents the discrete Fourier transform matrix, W k Represents the correlation filter model corresponding to the k-th frame image.

[0102] Preferably, obtaining the target position of the next frame image according to the second response map includes: determining the position of the maximum response value of the second response map, and the position is the target position of the next frame image.

[0103] System Example

[0104] Since the correlation filtering target tracking system provided by the embodiment of the present invention is based on the same principle as the aforementioned correlation filtering target tracking method, the repetitions may refer to the aforementioned method embodiment and will not be repeated here.

[0105] Specifically, another embodiment of the present invention discloses a correlation filtering target tracking system. Figure 6 As shown, the system includes:

[0106] The target position determination module 110 is configured to perform a correlation operation using the current correlation filter model and the frequency domain features of the current frame image to obtain a first response map, thereby determining the current target position.

[0107] The frequency domain feature extraction module 120 is used to segment the current frame image to obtain a new window image based on the current target position, and extract the frequency domain features of the new window image.

[0108] The correlation filter model updating module 130 is used to optimize the objective function by using the first response image and the frequency domain characteristics of the new window image to obtain the correlation filter model corresponding to the next frame image.

[0109] The target position determination module 110 is further configured to perform a correlation operation using the correlation filter model corresponding to the next frame image and the frequency domain features of the next frame image to obtain a second response map, and obtain the target position of the next frame image according to the second response map.

[0110] Preferably, the frequency domain feature extraction module 120 is specifically used to:

[0111] Perform inverse Fourier transform on the frequency domain image matrix to obtain the spatial domain image matrix corresponding to the new window image.

[0112] The monoscopic image matrix corresponding to the new window image is obtained according to the spatial domain image matrix and the grayscale matrix corresponding to the new window image.

[0113] The local amplitude, local direction and local phase of the new window image are extracted from the monoscopic image matrix to form the frequency domain features of the new window image.

[0114] Another embodiment of the present invention discloses a storage medium for storing computer instructions, and a processor executing a computer program to implement the aforementioned correlation filtering target tracking method. The storage medium can be a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0115] Compared with the prior art, the correlation filtering target tracking method and system disclosed in the embodiments of the present invention first creatively propose to update the correlation filtering model according to the frequency domain features of the image and perform target tracking based on the correlation filtering model, wherein the local phase contains most of the image information, and the higher the time domain and spatial domain frequency, the smaller the role of the local amplitude in the signal, and the more prominent the role of the phase. Therefore, the image reconstructed based on the local phase will enhance and highlight edges, lines and other narrow structures, and is not affected by external brightness and noise changes. Compared with spatial domain image features, it can more accurately describe image features. Therefore, target tracking based on local phase can greatly improve the accuracy of target tracking. Secondly, the correlation filtering target tracking method and system disclosed in the embodiments of the present invention only describes the image through three features: local amplitude, local phase and local direction. The computational complexity is small and the computational cost is high. It avoids the problem of large computational complexity and high computational cost caused by describing the image through spatial domain features such as image grayscale features, HOG features and CN features in the prior art, and can greatly improve the real-time performance of target tracking. In addition, the correlation filtering target tracking method and system disclosed in the embodiments of the present invention suppress response map distortion by adding frequency domain image features and response map distortion regularization terms, thereby improving image feature description and improving boundary effects, and reducing the impact of background noise. This allows for rapid and accurate target tracking based on images with scenes such as background clutter, rapid motion, motion blur, occlusion, and deformation.

[0116] The above description is only a preferred specific 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 thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A correlation filtering target tracking method, characterized in that: include: The current correlation filter model is used to perform correlation operation with the frequency domain features of the current frame image to obtain the first response map, thereby determining the current target position; Segmenting the current frame image to obtain a new window image based on the current target position, and extracting frequency domain features of the new window image; Optimizing the objective function using the frequency domain features of the first response graph and the new window image to obtain a correlation filter model corresponding to the next frame image; specifically comprising: Establish an objective function, which is: Among them, y represents the ideal response map corresponding to the current frame image, y∈R N , D represents the total number of channels corresponding to the image, d represents the dth channel, Represents the frequency domain features of the d-th channel corresponding to the k-th frame new window image, Represents the d-th channel correlation filter model corresponding to the k-th frame new window image, B represents the clipping matrix, B∈R M×N , N represents the length of the vector corresponding to each frequency domain feature, M represents the number of central elements of the vector corresponding to each frequency domain feature, λ represents the spatial regularization parameter, * represents the correlation operation, [ψ p,q ] represents a shift operation that makes the peak value of the response graph of the current frame image coincide with the peak value of the response graph of the previous frame image; An optimization algorithm is used to solve the objective function, and the correlation filter model that minimizes the objective function value is used as the correlation filter model corresponding to the next frame image; The relevant filtering model corresponding to the next frame image is: in, λ represents the spatial regularization parameter, μ represents the penalty factor, I D represents a D×D two-dimensional matrix, F represents the discrete Fourier transform matrix, W k Represents the correlation filter model corresponding to the k-th frame image; A correlation operation is performed using the correlation filter model corresponding to the next frame image and the frequency domain features of the next frame image to obtain a second response map, and the target position of the next frame image is obtained according to the second response map.

2. The correlation filtering target tracking method according to claim 1, characterized in that: Extracting the frequency domain features of the new window image includes: Performing Rees transform on the new window image to obtain a corresponding frequency domain image matrix; Performing an inverse Fourier transform on the frequency domain image matrix to obtain a spatial domain image matrix corresponding to the new window image; Obtaining a monoscopic image matrix corresponding to the new window image according to the spatial domain image matrix and the grayscale matrix corresponding to the new window image; The frequency domain features of the new window image are extracted from the monoscopic image matrix, where the frequency domain features include local amplitude, local direction and local phase.

3. The correlation filtering target tracking method according to claim 2, wherein: The monoscopic image matrix is ​​expressed by the following formula: I M =I-iR 1(X) -jR 2(X) , in, In the above formula, R 1(X) 、R 2(X) Represent the two components in the single-image matrix, I represents the grayscale matrix corresponding to the new window image, A represents the local amplitude of the new window image, represents the local phase of the new window image, and θ represents the local direction of the new window image.

4. The correlation filtering target tracking method according to claim 3, wherein: The local amplitude, local direction and local phase are:

5. The correlation filtering target tracking method according to any one of claims 2 to 4, characterized in that: Obtaining the target position of the next frame image according to the second response map includes: determining the position of the maximum response value of the second response map, which is the target position of the next frame image.

6. A correlation filtering target tracking system, characterized in that: include: A target position determination module is used to perform a correlation operation using the current correlation filter model and the frequency domain features of the current frame image to obtain a first response map, thereby determining the current target position; A frequency domain feature extraction module is used to segment the current frame image to obtain a new window image based on the current target position and extract frequency domain features of the new window image; A correlation filter model updating module is used to optimize the objective function using the frequency domain features of the first response graph and the new window image to obtain a correlation filter model corresponding to the next frame image; specifically comprising: Establish an objective function, which is: Among them, y represents the ideal response map corresponding to the current frame image, y∈R N , D represents the total number of channels corresponding to the image, d represents the dth channel, Represents the frequency domain features of the d-th channel corresponding to the k-th frame new window image, Represents the d-th channel correlation filter model corresponding to the k-th frame new window image, B represents the clipping matrix, B∈R M×N , N represents the length of the vector corresponding to each frequency domain feature, M represents the number of central elements of the vector corresponding to each frequency domain feature, λ represents the spatial regularization parameter, * represents the correlation operation, [ψ p,q ] represents a shift operation that makes the peak value of the response graph of the current frame image coincide with the peak value of the response graph of the previous frame image; An optimization algorithm is used to solve the objective function, and the correlation filter model that minimizes the objective function value is used as the correlation filter model corresponding to the next frame image; The relevant filtering model corresponding to the next frame image is: in, λ represents the spatial regularization parameter, μ represents the penalty factor, I D represents a D×D two-dimensional matrix, F represents the discrete Fourier transform matrix, W k Represents the correlation filter model corresponding to the k-th frame image; The target position determination module is further configured to perform correlation calculations on the frequency domain features of the next frame image using the correlation filter model corresponding to the next frame image to obtain a second response map, and obtain the target position of the next frame image according to the second response map.

7. The correlation filtering target tracking system according to claim 6, characterized in that: The frequency domain feature extraction module is specifically used for: Performing Rees transform on the new window image to obtain a corresponding frequency domain image matrix; Performing an inverse Fourier transform on the frequency domain image matrix to obtain a spatial domain image matrix corresponding to the new window image; Obtaining a monoscopic image matrix corresponding to the new window image according to the spatial domain image matrix and the grayscale matrix corresponding to the new window image; The local amplitude, local direction and local phase of the new window image are extracted from the monoscopic image matrix to form the frequency domain features of the new window image.

8. A storage medium for storing computer instructions, characterized in that: The processor executes the computer program to implement the correlation filtering target tracking method described in any one of claims 1 to 5.

Images