A target tracking protection method based on low-frequency perturbation attack retrieval

Abstract

The application discloses a target tracking protection method based on low-frequency disturbance attack retrieval, relates to the technical field of visual target tracking, and comprises the following steps: acquiring an input video sequence, determining a subframe detection area center image of the video sequence, performing low-frequency disturbance detection through the subframe detection area center image, and obtaining a low-frequency sampling result; inputting the low-frequency sampling result into a target tracking model for processing, and obtaining a corrected target tracking result.The method of the application runs through the complete chain from detection to response in the whole defense strategy, ensures that the judgment can be quickly made and corresponding measures can be taken when encountering a confrontational attack, and effectively protects the safety and stability of the visual target tracking system.

Description

Technical Field

[0001] The present invention relates to the technical field of visual target tracking, and in particular to a target tracking protection method based on low-frequency disturbance attack retrieval. Background Art

[0002] Visual object tracking is a key area in computer vision. It locates the target in each frame by continuously inferring key frames in a video sequence to provide the target's trajectory and activity area. This technology has been significantly improved in practical applications in recent years and is widely used in fields such as autonomous driving, biological behavior monitoring, and aerial visual target tracking. In recent years, visual object tracking algorithms based on deep learning have made great progress. However, this has also exposed the system's vulnerability to adversarial attacks. To improve the robustness of the system, researchers have proposed a variety of defense mechanisms, including adversarial training, input preprocessing, and model architecture optimization. However, in the face of complex and diverse attack methods, the effectiveness of these defense measures is still limited.

[0003] Adversarial attacks have become a serious problem in existing technologies. Adversarial attacks generate carefully designed perturbation signals, causing target trackers to misjudge the target's true location, thereby affecting tracking accuracy and stability. Such attacks can have serious consequences in practical applications such as autonomous driving, security monitoring, and biological behavior monitoring, significantly reducing safety and reliability.

[0004] The importance of defense systems lies not only in improving the security and reliability of individual applications, but also in providing a comprehensive and reliable technical framework that enables visual object tracking technology to operate stably in a variety of complex environments and face potential attacks. By implementing a target tracking protection method based on low-frequency perturbation attack retrieval, we can effectively address the challenges of adversarial attacks and provide the necessary security and technical support for a wide range of applications.

[0005] In real-world applications, autonomous driving systems must operate reliably in complex environments, including extreme weather, complex road conditions, and unexpected situations. Attackers can generate carefully crafted disturbance signals to cause the visual target tracking model's predictions to significantly deviate from the actual target, leading to system misjudgments and potentially posing safety risks. Accurately calibrating the target's position is crucial in real-time tracking systems. In the autonomous driving sector, vehicles rely on highly accurate visual object tracking technology for reliable navigation and obstacle avoidance in complex road environments. In security monitoring systems, surveillance cameras and intelligent detection systems are widely used to safeguard public spaces and critical facilities. Adversarial attacks can generate interference signals, preventing monitoring systems from correctly identifying and tracking intruders, leading to security vulnerabilities. Many defense methods are unstable and unreliable in the face of rapidly changing environments or complex interference noise, compromising the system's defense capabilities and failing to guarantee stable operation in various complex scenarios.

[0006] Furthermore, existing defense methods often require high-performance hardware when faced with large-scale data and complex tasks. Applications requiring real-time processing, particularly in low-resource environments such as edge devices and embedded systems, require extensive forward and backward propagation computations, significantly increasing system response time and resource requirements. The high computational power requirements of these methods directly impact system deployment and operational efficiency. This creates significant limitations for applications requiring high real-time performance, such as autonomous driving and security monitoring, making it difficult to guarantee the efficiency and effectiveness of existing defense methods in practical deployments. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes a target tracking protection method based on low-frequency perturbation attack retrieval, which ensures that judgments can be made quickly and corresponding measures can be taken when encountering adversarial attacks, thereby effectively protecting the security and stability of the visual target tracking system.

[0008] On the one hand, to achieve the above-mentioned object, the present invention provides a target tracking and protection method based on low-frequency disturbance attack retrieval, comprising:

[0009] Obtaining an input video sequence, determining a central image of a sub-frame detection area of ​​the video sequence, performing low-frequency disturbance detection using the central image of the sub-frame detection area, and obtaining a low-frequency sampling result;

[0010] Inputting the low-frequency sampling result into the target tracking model for processing to obtain a corrected target tracking result;

[0011] The target tracking model quantifies the low-frequency sampling results through a linear regression method, and dynamically adjusts parameters to correct the input frame.

[0012] Preferably, obtaining the low-frequency sampling result includes:

[0013] Converting the central image of the sub-frame detection area into a grayscale image, dividing the grayscale image into a number of identical small image blocks, performing discrete cosine transform on each of the small image blocks, and converting the pixel values ​​in the spatial domain into frequency domain representation;

[0014] The coefficients of the same part in each of the small blocks of images are selected to extract low-frequency components, and all low-frequency components are integrated to obtain the low-frequency sampling result.

[0015] Preferably, the process of inputting the low-frequency sampling result into the target tracking model for processing is:

[0016] Regression analysis is performed on the low-frequency sampling results to determine the nature and scope of the attack disturbance and quantify the characteristics of the disturbance signal.

[0017] Preferably, before the step of performing linear regression analysis on the low-frequency sampling results, the method includes:

[0018] Calculating the mean and standard deviation based on the low-frequency sampling results;

[0019] A reference model is constructed, a threshold is set for the reference model, the mean and standard deviation are compared with the threshold, and it is determined whether there is an abnormal disturbance signal. If there is an abnormal disturbance signal, a regression linear analysis is performed; if there is no abnormal disturbance signal, the low-frequency sampling result is input into the target tracking model as the actual image to be predicted.

[0020] Preferably, the process of performing the regression linear analysis is:

[0021] Linear regression is introduced to analyze the multi-directional detection disturbance characteristics, and a disturbance characteristic curve is established to quantify the abnormal disturbance signal; wherein, the disturbance characteristic curve is related to the offset, disturbance slope and disturbance error.

[0022] Preferably, the process of inputting the low-frequency sampling result into the target tracking model for processing further includes:

[0023] The low-frequency domain repair signal is iteratively searched through the low-frequency sampling result, specifically:

[0024] Generate a two-dimensional signal based on the low-frequency sampling result, and project the two-dimensional signal to a sub-frame detection area to obtain a test frame;

[0025] The time-domain-based IoU score is calculated through the test frame to evaluate the impact of the current video frame on the target tracking model, and the sample with the highest time-domain-based IoU score is selected as the actual sample to be predicted.

[0026] Preferably, the fixed number of times is set to have the best efficiency based on the computing capability of the device.

[0027] On the other hand, to achieve the above-mentioned purpose, the present invention also provides a target tracking and protection system based on low-frequency disturbance attack retrieval, comprising:

[0028] Low-frequency sampling result acquisition module: used to obtain an input video sequence, determine the central image of the sub-frame detection area of ​​the video sequence, perform low-frequency disturbance detection based on the central image of the sub-frame detection area, and obtain a low-frequency sampling result;

[0029] Target tracking model processing module: used for inputting the low-frequency sampling result into the target tracking model for processing to obtain a corrected target tracking result;

[0030] Wherein, the target tracking model processing module includes:

[0031] A disturbance detection unit, used for analyzing low-frequency coefficients and detecting low-frequency disturbance signals;

[0032] A regression analysis unit, used for performing regression analysis on the detected disturbance signal;

[0033] The model adjustment unit is used to adjust and correct the target tracking model according to the regression analysis results.

[0034] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the target tracking and protection method based on low-frequency disturbance attack retrieval.

[0035] The present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the target tracking and protection method based on low-frequency disturbance attack retrieval are implemented.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] The present invention analyzes low-frequency disturbance signals in input video sequences through discrete cosine transform (DCT), which can effectively identify and defend against potential adversarial attacks, and has significant advantages in improving the stability and robustness of the tracking system. Through regression analysis and dynamic adjustment, it can adapt to different types of disturbance attacks and effectively improve the robustness of the model, significantly enhancing the reliability and security of visual target tracking in practical applications.

[0038] The present invention constructs a comprehensive and highly integrated defense framework through a series of steps including input data preprocessing, discrete cosine transform, low-frequency disturbance detection, disturbance regression analysis and model adjustment. The entire defense strategy runs through the complete chain from detection to response, ensuring that judgments can be made quickly and corresponding measures can be taken when encountering adversarial attacks, effectively protecting the security and stability of the visual target tracking system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0040] Figure 1 This is a flow chart of a target tracking and protection method based on low-frequency disturbance attack retrieval according to an embodiment of the present invention;

[0041] Figure 2 This is a flowchart of iteratively searching a low-frequency domain repair signal based on a low-frequency sampling space according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0043] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0044] The present invention proposes a target tracking protection method based on low-frequency disturbance attack retrieval, such as Figure 1 ,include:

[0045] Obtaining an input video sequence, determining a central image of a sub-frame detection area of ​​the video sequence, performing low-frequency disturbance detection using the central image of the sub-frame detection area, and obtaining a low-frequency sampling result;

[0046] The low-frequency sampling results are input into the target tracking model for processing to obtain the corrected target tracking results;

[0047] Among them, the target tracking model quantifies the low-frequency sampling results through the linear regression method, and dynamically adjusts the parameters to correct the input frame.

[0048] Specifically, in the image processing process, the input video frame is first denoised and preprocessed to ensure the quality and consistency of the data. Subsequently, the input image is converted to the frequency domain space through discrete cosine transform (DCT), and the low-frequency components are extracted in the frequency domain space to generate orthogonal vector space and sample perturbations. This method can effectively reduce the misleading of high-frequency components on images and video frames, and ensure the directionality and effectiveness of generated patches. In order to improve the efficiency and resource optimization of the generation process, this embodiment adopts a sampling reconstruction strategy. The key sampling points are selected through a greedy algorithm, and combined with the reconstruction algorithm, the iterative process is optimized under the background of ensuring the need for repair effect.

[0049] In attack scenarios, an adjustment signal is generated through iterative query and optimization. Leveraging query data, the signal's direction and amplitude are continuously optimized, improving the accuracy and effectiveness of the defense. Finally, the resulting repaired video frame is further reconstructed and used in the visual object tracking process.

[0050] Furthermore, low-frequency sampling results are obtained, including:

[0051] Convert the central image of the sub-frame detection area into a grayscale image, and divide the grayscale image into several identical small blocks of images. Perform discrete cosine transform on each small block of images to convert the pixel values ​​in the spatial domain into representations in the frequency domain.

[0052] The coefficients of the same part in each small block of the image are selected, the low-frequency components are extracted, and all the low-frequency components are integrated to obtain the low-frequency sampling result.

[0053] Specifically, for an input video sequence, T, W, H, and C represent the sequence, width, height, and number of color channels of the video frames, respectively. During the actual tracking process, the first frame is used to provide target features. Detection and defense begin with the second frame. The video sequence is sequentially input into the target tracking model. For a system that can relocalize after an object is lost, the target tracking model can skip frames and reset itself as the target is lost during tracking.

[0054] Input video sequence It is a four-element two-dimensional vector that represents the clockwise position of the frame coordinate system where the predicted target is located, starting from the upper left corner. Based on the center of the area, the image patch area is obtained, with both length and width being Calculate the center coordinates of the image patch area (x c ,y c ) to obtain the center image of the sub-frame detection area.

[0055] In this embodiment, the calculation for generating the low-frequency sampling space is performed based on the image of the determined area, specifically:

[0056] (1-1) Read the color pixel information in the area and convert it into a grayscale image. For a traditional three-channel RGB color image, perform channel-by-channel conversion Gray = 0.299·R + 0.587·G + 0.114·B, where R represents the red component of a single pixel and takes the value [0, 1], G represents the green component of a single pixel and takes the value [0, 1], and B represents the blue component of a single pixel and takes the value [0, 1].

[0057] (1-2) Divide the grayscale image into 8*8 small blocks of equal size;

[0058] (1-3) Each small image block is subjected to discrete cosine transform to convert the pixel value in the spatial domain into a representation in the frequency domain:

[0059]

[0060] Where cos() is the corresponding trigonometric function; F(u, ν) is a two-dimensional matrix used to describe the frequency domain coefficients after two-dimensional discrete cosine transform; M and N are the number of rows and columns of the image, respectively; α(u) and α(ν) are normalization coefficients, defined as:

[0061]

[0062] Where u is the frequency component in the horizontal direction of the two-dimensional DCT transform, ranging from 0 to N-1, and v is the frequency component in the vertical direction of the two-dimensional DCT transform, ranging from 0 to N-1;

[0063] (1-4) After the transform calculation, most of the information in the small image block is concentrated in the low-frequency part. The low-frequency components are extracted by selecting several coefficients in the upper left corner of the small image block. In this embodiment, only the 2×2 or 4×4 area in the upper left corner is selected from the 8×8 DCT coefficient matrix;

[0064] (1-5) All extracted low-frequency components are integrated to generate a new image or feature space, which represents the low-frequency sampling results of the image.

[0065] Furthermore, the process of inputting the low-frequency sampling results into the target tracking model for processing is as follows:

[0066] Linear regression analysis is performed on the low-frequency sampling results to determine the nature and scope of the attack disturbance and quantify the characteristics of the disturbance signal.

[0067] Furthermore, the step of performing linear regression analysis on the low-frequency sampling results includes:

[0068] Calculate the mean and standard deviation based on the low-frequency sampling results;

[0069] Set a threshold, compare the mean and standard deviation with the threshold, and determine whether there is an abnormal disturbance signal. If there is an abnormal disturbance signal, perform regression linear analysis. If not, input the low-frequency sampling result as the actual image to be predicted into the target tracking model.

[0070] Specifically, based on the low-frequency sampling results, a low-frequency coefficient matrix L=F(u, v)|0≤u, ν<K, K represents the value range of the low-frequency coefficient, and the mean and standard deviation are calculated;

[0071] The method to calculate the standard deviation is:

[0072] Where, σ L is the standard deviation, m is the number of rows of data, n is the number of columns of data, i and j are the indexes of specific rows and columns respectively, L ij is the value of the two-dimensional data in row i and column j, μ L is the mean of two-dimensional data. The formula for calculating the mean of two-dimensional data is:

[0073] A reference model R is constructed to implement detection based on the threshold τ: ΔL = max(|LR|), where ΔL is used to represent the model difference coefficient. If ΔL > τ, an abnormal disturbance signal exists and regression analysis is performed. If no abnormal disturbance signal exists, the low-frequency sampling result is used as the actual image to be predicted and input into the target tracking model.

[0074] Furthermore, the process of performing regression linear analysis is as follows:

[0075] Linear regression is introduced to analyze the multi-directional detection disturbance characteristics and a disturbance characteristic curve is established. The disturbance characteristic curve is related to the offset, disturbance slope and disturbance error.

[0076] Specifically, linear regression is introduced to analyze the multi-directional detection disturbance characteristics and the disturbance characteristic curve L is established. perturb :

[0077] L perturb ≈β0+β1x+∈;

[0078] Where β0 is the offset, β1 is the disturbance slope, and ∈ is the disturbance error.

[0079] Apply the linear regression results to the target tracking model to analyze the impact of the disturbance signal on the target tracking model response. The specific process is as follows:

[0080] Calculate the residuals and calculate the mean. The method for calculating the residuals is:

[0081] ∈ i =L perturb，i -(β0+β1x i );

[0082] Where,∈ i Represents the difference between the actual observation value and the predicted value, L perturb，i is the actual observation value of the i-th data point, x i are the input variables used for prediction.

[0083] The mean of the residuals should be close to zero. If the mean of the residuals deviates significantly from zero, there may be an abnormal disturbance signal.

[0084] If there are abnormal disturbance signals, try to search for repair patches to improve the accuracy of the target tracking model, such as Figure 2 ,include:

[0085] Randomly select a direction from the low-frequency sampling result, generate a signal and superimpose it on the sampling area to obtain a test frame. For the corresponding pixels on the signal and the sampling area, the color values ​​of the three RGB channels are superimposed.

[0086] Compute a Gaussian signal with a maximum range based on its distribution function and variance:

[0087] The distribution function is

[0088] The variance is

[0089] In the formula, π is the circumference of the circle, x is the horizontal coordinate in the normal distribution, and μ is the center position of the normal distribution, which is also the average value of the data. It represents the square of the standard deviation and is used to measure the degree of dispersion of the data.

[0090] The above distribution function generates corresponding full-channel noise for each pixel in the image patch area.

[0091] The generated Gaussian signal is appended to the original frame as a candidate test frame.

[0092] The image space generated by sampling the image patch area is used to generate an orthogonal patch area, and a block discrete cosine transform is performed. The low-frequency coefficients are extracted from the transformed frequency domain signal and analyzed. Specifically:

[0093] (2-1) Select the direction vector q from the low-frequency sampling result, and consider the negative vector q'=-q;

[0094] (2-2) According to the input vector x=[x1, x2, ..., x n ], generate a sequence Q = [x1, x 1+s , x 1+2s ,...,x 1+ks ], this sequence contains all possible alternative repair signals;

[0095] (2-3) Divide the generated sequence Q into L groups evenly according to the stride, and randomly select a signal from each group, selecting a total of L' signals (L' = L);

[0096] (2-4) The above selected signals are linearly combined according to the corresponding weighting coefficients to generate a linear combination signal v = α1DCT1 + α2DCT2 + ... + α n DCT n , where DCT represents the orthogonal discrete cosine transform vector and the coefficient α n is the corresponding weight when generating the signal, and n is used to represent the number of DCT basis functions of the signal;

[0097] (2-5) Convert the time domain signal into the frequency domain signal:

[0098]

[0099] Where DCT(υ) is the frequency domain signal after transformation, v j is the j-th sample of the input time domain signal sequence Q, where j is the sequence number and takes the value [0, N-1]; k is the index of the transformed frequency domain component, and N is the sample length of the input time domain signal sequence Q;

[0100] (2-6) According to and Convolution is performed with the grayscale image to obtain the horizontal gradient image g y And the vertical gradient image g x , and calculate the tangent direction

[0101] Calculate the direction orthogonal to the tangent direction as the auxiliary direction;

[0102] (2-7) A two-dimensional signal is generated according to the tangential direction and the auxiliary direction. This process is based on the accumulation of low-frequency components δ = v + β1u1 + β2u2 + ... + β k u k ;

[0103] Where δ is the two-dimensional signal accumulated in the tangential direction and the auxiliary direction, v is the low-frequency part of the signal, and β k is the weight coefficient in the kth signal direction, u k is the signal component on the kth signal;

[0104] Based on the low-frequency sampling space, iteratively search for a low-frequency domain repair signal to generate a test frame, and calculate a time-domain-based intersection over union score through the test frame, in each iteration, by evaluating the influence of the current video frame on the target tracking model, select the signal patch that maximizes the reduction of the disturbance effect, update the current frame, and select the sample with the highest tracking score as the actual to-be-predicted sample based on a fixed number of queries.

[0105] Further, based on the low-frequency sampling space, iteratively search for a low-frequency domain repair signal, specifically:

[0106] (3-1) Obtain a two-dimensional signal δ using the selected direction and superimpose it on the sampling region to generate a test frame I +q , I -q :

[0107]

[0108] In the formula, is the video frame after the k-1th generated repair signal, used for superposition to generate the current signal, δ q represents a two-dimensional signal generated by the direction vector q, δ q′ represents a two-dimensional signal generated by the direction vector q';

[0109] For each test frame, generate a detection target region through the target tracking model

[0110] Calculate a time-domain-based intersection over union score

[0111] In the formula, λ is a weighting coefficient, with a value range of [0, 1], used to balance the weights of the current frame and the previous frame; is the predicted bounding box of the original frame; is the predicted bounding box of the test frame; is the predicted bounding box of the previous original frame;

[0112] (3-2) Add the intersection over union score to the set ψ k = ψ k ∪ S IOU In each iteration, by evaluating the influence of the current video frame on the target tracking model, select the signal patch that maximizes the reduction of the disturbance effect to perform the optimal greedy strategy;

[0113] (3-3) Update the current frame: I' tk = I' tk-1 + ψ k · η;

[0114] In the formula, η is the patch update step size, and I' tk is the video frame after the signal patch repair generated by the kth iteration;

[0115] (3-4) Update the iteration count to maintain computational efficiency:

[0116] Based on a fixed number of queries, the sample with the highest tracking score is selected as the actual sample to be predicted and input into the target tracking model. In this embodiment, the fixed number of queries is based on the computing power of the device to select the most efficient setting. Each time the target tracking model is input and the intersection-over-union score S is calculated IOU , select the single test frame with the highest score for reconstruction. Specifically, Complete reconstruction; I' n is the pixel value of the reconstructed image.

[0117] Specifically, the reconstruction function The build process consists of the following steps:

[0118] (4-1) Find the corresponding source coordinates in the original image from each target coordinate in the single test frame with the highest score:

[0119]

[0120] Where x and y represent the horizontal and vertical coordinates of the image, src and dst represent the target positions before and after reconstruction, w and h represent the length of the reconstruction, and x src is the horizontal coordinate in the source image, z dst To transform the image horizontal coordinate, w dst is the converted image width, w src is the source image width, y src is the ordinate of the source image, y dst To transform the image ordinate, h dst To convert the image height, h src is the source image height;

[0121] (4-2) Based on the four directions of the two-dimensional coordinate axis, find the four neighboring pixels in the source image that are closest to the calculated position;

[0122] (4-3) Calculate the weight of each neighbor pixel. These weights are calculated based on the relative distance between the target pixel position and the neighbor pixel position. The distance is proportional to the weight.

[0123] (4-4) Use the calculated weights to perform weighted average calculation on the values ​​of the four neighboring pixels;

[0124] (4-5) For pixels whose calculated positions exceed the range of the source image, round them down to avoid out-of-bounds access and obtain the reconstructed image I′ n ; Finally, I′ n The target tracking model is input as the actual image to be predicted; the actual defense system does not rely on any working mechanism or calculation rules of the specific target tracking model.

[0125] This embodiment further provides a target tracking and protection system based on low-frequency disturbance attack retrieval, including:

[0126] Low-frequency sampling result acquisition module: used to obtain an input video sequence, determine the central image of the sub-frame detection area of ​​the video sequence, perform low-frequency disturbance detection based on the central image of the sub-frame detection area, and obtain a low-frequency sampling result;

[0127] Target tracking model processing module: used for inputting the low-frequency sampling result into the target tracking model for processing to obtain a corrected target tracking result;

[0128] Wherein, the target tracking model processing module includes:

[0129] A disturbance detection unit, used for analyzing low-frequency coefficients and detecting low-frequency disturbance signals;

[0130] A regression analysis unit, used for performing regression analysis on the detected disturbance signal;

[0131] The model adjustment unit is used to adjust and correct the target tracking model according to the regression analysis results.

[0132] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of a target tracking and protection method based on low-frequency disturbance attack retrieval.

[0133] This embodiment also provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the target tracking and protection method based on low-frequency disturbance attack retrieval are implemented.

[0134] This embodiment proposes a method for protecting target tracking from low-frequency perturbation attacks. This method uses discrete cosine transform (DCT) analysis of low-frequency perturbation signals in input video sequences to effectively identify and defend against potential adversarial attacks. Experiments have shown that this method significantly improves the stability and robustness of tracking systems.

[0135] First, this method excels in applications requiring high security and real-time response, such as autonomous driving and security monitoring. Experimental data demonstrates that the system can effectively detect over 90% of low-frequency perturbation attacks and implement defensive measures with early warning. In a 12,000-frame video test, the system's false detection rate remained below 2%, while its missed detection rate was less than 5%. This high detection rate and low false detection rate ensure the system maintains stable tracking performance even in complex environments. Second, through regression analysis and dynamic adjustment, the system can adapt to different types of perturbation attacks and effectively improve the robustness of the model. When facing adversarial attacks with intensities between 0.05 and 0.1, tracking accuracy drops by less than 10%, compared to a drop of over 30% in the absence of defenses. This demonstrates effective defense against various attack intensities, significantly improving the reliability and security of visual object tracking systems in practical applications.

[0136] Furthermore, the computational overhead during detection and defense is low, with the average processing time for single-frame detection being less than 10 milliseconds, a 30% improvement in efficiency compared to traditional methods. This low computational resource requirement makes this defense method suitable not only for high-performance computing environments but also for resource-constrained embedded systems, further expanding its practical application.

[0137] This defense approach builds a comprehensive and highly integrated defense framework through a series of steps, including input data preprocessing, discrete cosine transform, low-frequency perturbation detection, perturbation regression analysis, and model tuning. This entire defense strategy spans the entire detection-to-response chain, ensuring rapid judgment and appropriate action in the event of an adversarial attack, effectively protecting the security and stability of the visual object tracking system.

[0138] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A target tracking and protection method based on low-frequency disturbance attack retrieval, characterized in that: include: Obtaining an input video sequence, determining a central image of a sub-frame detection area of ​​the video sequence, performing low-frequency disturbance detection using the central image of the sub-frame detection area, and obtaining a low-frequency sampling result; Inputting the low-frequency sampling result into the target tracking model for processing to obtain a corrected target tracking result; The target tracking model quantifies the low-frequency sampling results by a linear regression method and dynamically adjusts parameters to correct the input frame; Obtaining the low-frequency sampling result includes: Converting the central image of the sub-frame detection area into a grayscale image, dividing the grayscale image into a number of identical small image blocks, performing discrete cosine transform on each of the small image blocks, and converting the pixel values ​​in the spatial domain into frequency domain representation; Selecting coefficients of the same part in each of the small blocks of images, extracting low-frequency components, and integrating all low-frequency components to obtain the low-frequency sampling result; The process of inputting the low-frequency sampling results into the target tracking model for processing is as follows: Performing regression analysis on the low-frequency sampling results to determine the nature and scope of the attack disturbance and quantify the characteristics of the disturbance signal; Before performing linear regression analysis on the low-frequency sampling results, the method includes: Calculating the mean and standard deviation based on the low-frequency sampling results; Constructing a reference model, setting a threshold for the reference model, comparing the mean and standard deviation with the threshold to determine whether an abnormal disturbance signal exists, performing a linear regression analysis if an abnormal disturbance signal exists, and inputting the low-frequency sampling result as the actual image to be predicted into the target tracking model if no abnormal disturbance signal exists; The process of performing the regression linear analysis is as follows: Linear regression is introduced to analyze the multi-directional detection disturbance characteristics, and a disturbance characteristic curve is established to quantify the abnormal disturbance signal; wherein, the disturbance characteristic curve is related to the offset, disturbance slope and disturbance error.

2. The target tracking and protection method based on low-frequency disturbance attack retrieval according to claim 1 is characterized in that: The process of inputting the low-frequency sampling result into the target tracking model for processing further includes: The low-frequency domain repair signal is iteratively searched through the low-frequency sampling result, specifically: Generate a two-dimensional signal based on the low-frequency sampling result, and project the two-dimensional signal to a sub-frame detection area to obtain a test frame; The time-domain-based intersection-over-union (IoU) score is calculated through the test frame to evaluate the impact of the current video frame on the target tracking model, select the signal patch that maximizes the reduction of the disturbance effect, update the current frame, and select the sample with the highest time-domain IoU score based on a fixed number of queries as the actual sample to be predicted.

3. The target tracking and protection method based on low-frequency disturbance attack retrieval according to claim 2 is characterized in that: The fixed number of times is set to the most efficient setting based on the computing power of the device.

4. A target tracking and protection system based on low-frequency disturbance attack retrieval, characterized in that: include: Low-frequency sampling result acquisition module: used to obtain an input video sequence, determine the central image of the sub-frame detection area of ​​the video sequence, perform low-frequency disturbance detection based on the central image of the sub-frame detection area, and obtain a low-frequency sampling result; Target tracking model processing module: used for inputting the low-frequency sampling result into the target tracking model for processing to obtain a corrected target tracking result; Wherein, the target tracking model processing module includes: A disturbance detection unit, used for analyzing low-frequency coefficients and detecting low-frequency disturbance signals; A regression analysis unit, used for performing regression analysis on the detected disturbance signal; A model adjustment unit, used to adjust and correct the target tracking model according to the regression analysis results; Wherein, obtaining the low-frequency sampling result includes: Converting the central image of the sub-frame detection area into a grayscale image, dividing the grayscale image into a number of identical small image blocks, performing discrete cosine transform on each of the small image blocks, and converting the pixel values ​​in the spatial domain into frequency domain representation; Selecting coefficients of the same part in each of the small blocks of images, extracting low-frequency components, and integrating all low-frequency components to obtain the low-frequency sampling result; The process of inputting the low-frequency sampling results into the target tracking model for processing is as follows: Performing regression analysis on the low-frequency sampling results to determine the nature and scope of the attack disturbance and quantify the characteristics of the disturbance signal; Before performing linear regression analysis on the low-frequency sampling results, the method includes: Calculating the mean and standard deviation based on the low-frequency sampling results; Constructing a reference model, setting a threshold for the reference model, comparing the mean and standard deviation with the threshold to determine whether an abnormal disturbance signal exists, performing a linear regression analysis if an abnormal disturbance signal exists, and inputting the low-frequency sampling result as the actual image to be predicted into the target tracking model if no abnormal disturbance signal exists; The process of performing the regression linear analysis is: Linear regression is introduced to analyze the multi-directional detection disturbance characteristics, and a disturbance characteristic curve is established to quantify the abnormal disturbance signal; wherein, the disturbance characteristic curve is related to the offset, disturbance slope and disturbance error.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the method according to any one of claims 1 to 3 are implemented.

Images