A target tracking method and system based on parallel three-branch correlation filter

Abstract

The present invention relates to a target tracking method and system based on parallel three-branch correlation filters, the method comprising the following steps: obtaining the target position and target scale of an initial frame image, initializing three parallel correlation filters; for the t-th frame image, at the target position P t‑1 and target scale S t‑1 Extract position feature Z t,pos and scale feature Z t,scale The system fuses three parallel correlation filters to determine target position and scale. A confidence mechanism is used to determine whether to perform a sample update. If so, a new training sample set is generated using a Gaussian mixture model, the weights of the three correlation filters are updated, and the three parallel correlation filters are updated online. Otherwise, detection proceeds to the next frame. Compared to existing technologies, this method effectively addresses target tracking losses in complex ocean environments caused by wave and ship interference, light fluctuations, and rapid movement, demonstrating superior performance in terms of accuracy and success rate.

Description

Technical Field

[0001] The present invention relates to the technical field of target tracking, and in particular to a method and system for realizing target tracking in a complex ocean environment by utilizing a parallel three-branch correlation filter based on a confidence mechanism. Background Art

[0002] Marine object tracking is crucial for search and rescue operations in complex ocean environments. However, complex ocean environments pose significant challenges to tracking, including lighting variations, wave effects, and occlusion by other vessels. Traditional visual object trackers struggle to accurately and robustly track objects in complex ocean environments. Designing an efficient dynamic visual tracker that delivers accurate, real-time, and robust results is crucial in these challenging environments. Summary of the Invention

[0003] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a target tracking method and system based on parallel three-branch correlation filter, which can effectively solve the target tracking loss caused by wave ship interference, light changes, rapid movement and other reasons in complex marine environments, and show excellent performance in terms of accuracy and success rate, and has outstanding properties in terms of lighting changes, interference occlusion, rapid movement and scale changes.

[0004] The purpose of the present invention can be achieved by the following technical solutions:

[0005] A target tracking method based on a parallel three-branch correlation filter, which uses a parallel three-branch correlation filter based on a confidence mechanism to track a target, includes the following steps:

[0006] S1, obtain the target position P1 and target scale S1 of the initial frame image, establish a training sample set, initialize three parallel correlation filters, and determine the weight size of each correlation filter;

[0007] S2, for the t-th frame image, at the target position P t-1 and target scale S t-1 Extract position feature Z t，pos and scale feature Z t，scale , t>1;

[0008] S3, based on position feature Z t，pos The first response score y is calculated with three parallel correlation filters t，pos , based on the scale feature Z t，scale The second response score y is calculated with three parallel correlation filters t，scale ;

[0009] S4. Obtain target position P based on the first response score t, obtain the target scale S based on the second response score t ;

[0010] S5, determine whether to continue target tracking, if yes, execute step S4, otherwise, exit target tracking;

[0011] S6. Use the confidence mechanism to determine whether to update the sample. If yes, execute step S7; otherwise, execute step S2.

[0012] S7, for the t-th frame image, at the target position P t and target scale S t Extract the current position feature F t，pos and the current scale feature F t，scale ;

[0013] S8, based on the current position feature F t，pos and the current scale feature F t，scale , use the Gaussian mixture model to generate a new training sample set, update the weights of the three correlation filters, update the three parallel correlation filters online, and execute step S2.

[0014] Furthermore, the three correlation filters can be expressed as:

[0015]

[0016]

[0017]

[0018] Among them, ε(h ψt ) represents the updated model of the ψth correlation filter, h ψ l Indicates the ψth correlation filter that needs to be updated, ψ=1, 2, 3, k represents the number of samples generated by window sliding, p represents the number of samples in the training sample set, represents the weight of the ψth correlation filter, represents the target sample, gs k represents the joint expected response distribution, λ represents the regularization parameter, and ω represents the spatial regularization parameter.

[0019] Furthermore, the first response score y t，pos The calculation formula is:

[0020]

[0021] Among them, α and β are fusion factors. The solution is obtained by Fourier transforming equations (1) to (3). and and represents three parallel correlation filters, Represents the inverse Fourier transform.

[0022] Furthermore, the second response score y t，scale The calculation formula is:

[0023]

[0024] in, and B t-1，scale represents the numerator and denominator of the scale-dependent filter of the previous frame, λ is the regularization parameter, λ≥0, Represents the inverse Fourier transform.

[0025] Furthermore, generating a new training sample set is specifically as follows:

[0026] The current position feature F t，pos and the current scale feature F t，scale A training sample set is added and the Gaussian mixture model is used to detect the similarity between samples. If the similarity between samples in the training sample set is less than a preset threshold, the two closest samples are merged.

[0027] Furthermore, the update formula of the correlation filter weight is as follows:

[0028]

[0029] Among them, η ψ represents the learning rate of the ψth correlation filter, t represents the tth frame, represents the updated weight of the ψth correlation filter, represents the weight of the ψth correlation filter before updating, ψ = 1, 2, 3.

[0030] Furthermore, the confidence mechanism is used to determine whether to update the sample:

[0031] Get the response map of each correlation filter and weight it to get the response map R. Select n relevant responses from the response map R, where R(i) represents the response value at the i-th position and p(R(i)) is the probability of the response value appearing. Calculate the expected μ of the n response values:

[0032]

[0033] Calculate the difference value CR using the following formula:

[0034]

[0035] If the CR value is less than the set threshold, the sample is updated; otherwise, the sample is not updated.

[0036] A target tracking system based on a parallel three-branch correlation filter includes an image acquisition module and a tracking module. The image acquisition module is used to acquire a continuous image sequence and send it to the tracking module. The tracking module uses a parallel three-branch correlation filter based on a confidence mechanism to track the target, including the following steps:

[0037] S1, obtain the target position P1 and target scale S1 of the initial frame image, establish a training sample set, initialize three parallel correlation filters, and determine the weight size of each correlation filter;

[0038] S2, for the t-th frame image, at the target position P t-1 and target scale S t-1 Extract position feature Z t，pos and scale feature Z t，scale , t>1;

[0039] S3, based on position feature Z t，pos The first response score y is calculated with three parallel correlation filters t，pos , based on the scale feature Z t，scale The second response score y is calculated with three parallel correlation filters t，scale ;

[0040] S4. Obtain target position P based on the first response score t , obtain the target scale S based on the second response score t ;

[0041] S5, determine whether to continue target tracking, if yes, execute step S4, otherwise, exit target tracking;

[0042] S6. Use the confidence mechanism to determine whether to update the sample. If yes, execute step S7; otherwise, execute step S2.

[0043] S7, for the t-th frame image, at the target position P t and target scale S t Extract the current position feature F t，pos and the current scale feature F t，scale ;

[0044] S8, based on the current position feature F t，pos and the current scale feature F t，scale , use the Gaussian mixture model to generate a new training sample set, update the weights of the three correlation filters, update the three parallel correlation filters online, and execute step S2.

[0045] Furthermore, the three correlation filters can be expressed as:

[0046]

[0047]

[0048]

[0049] Among them, ε(h ψt ) represents the updated model of the ψth correlation filter, h ψ l Indicates the ψth correlation filter that needs to be updated, ψ=1, 2, 3, k represents the number of samples generated by window sliding, p represents the number of samples in the training sample set, represents the weight of the ψth correlation filter, represents the target sample, gs k represents the joint expected response distribution, λ represents the regularization parameter, and ω represents the spatial regularization parameter.

[0050] Furthermore, the confidence mechanism is used to determine whether to update the sample:

[0051] Get the response map of each correlation filter and weight it to get the response map R. Select n relevant responses from the response map R, where R(i) represents the response value at the i-th position and p(R(i)) is the probability of the response value appearing. Calculate the expected μ of the n response values:

[0052]

[0053] Calculate the difference value CR using the following formula:

[0054]

[0055] If the CR value is less than the set threshold, the sample is updated; otherwise, the sample is not updated.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] (1) The values of the three-branch parallel correlation filter are fused to detect the target position and scale. The center position of the object is accurately located through weighted fusion of the response map to make tracking more robust.

[0058] (2) Three correlation filters are trained online using different weights for the same sample, and the weight of each correlation filter is updated using a different learning rate, which effectively improves the robustness of the tracker to overcome the harsh conditions in the marine environment.

[0059] (3) A confidence mechanism is proposed to calculate the CR value in the response graph, analyze the tracking effect of the current frame, and judge whether the tracking of the current frame is reliable based on the CR value to determine whether to update the sample and the correlation filter, so as to train a more accurate correlation filter, prevent model drift, and achieve good tracking effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flow chart of the target tracking method;

[0061] Figure 2 This is a flowchart of the DCF framework;

[0062] Figure 3 This is a flowchart of the PCF framework;

[0063] Figure 4 Flowchart for target recognition in marine environment for UAVs. DETAILED DESCRIPTION

[0064] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operating process. Obviously, the described embodiment is only an embodiment of a part of the present invention, not all embodiments, and the protection scope of the present invention is not limited to the following embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0065] "One embodiment" or "embodiment" referred to herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present invention. In the description of the present invention, it should be understood that the terms "first", "second" and "third" in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units that are inherent to these processes, methods, products or devices.

[0066] This specification provides method operation steps such as embodiments or flowcharts, but may include more or fewer operation steps based on conventional or non-creative work. The order of steps listed in the embodiments is only one way of executing the order of many steps and does not represent the only execution order. When the actual system or server product is executed, it can be executed in sequence or in parallel (for example, in a parallel processor or multi-threaded processing environment) according to the embodiment or the method shown in the accompanying drawings, or the execution order of steps without timing restrictions can be adjusted.

[0067] Example 1:

[0068] A target tracking method based on parallel three-branch correlation filter, using a parallel three-branch correlation filter based on a confidence mechanism for target tracking, such as Figure 1 As shown, the following steps are included:

[0069] S1, obtain the target position P1 and target scale S1 of the initial frame image, establish a training sample set, initialize three parallel correlation filters, and determine the weight size of each correlation filter;

[0070] S2, for the t-th frame image, at the target position P t-1 and target scale S t-1 Extract position feature Z t，pos and scale feature Z t，scale , t>1;

[0071] S3, based on position feature Z t，pos The first response score y is calculated with three parallel correlation filters t，pos , based on the scale feature Z t，scale The second response score y is calculated with three parallel correlation filters t，scale ;

[0072] S4. Obtain target position P based on the first response score t , obtain the target scale S based on the second response score t ;

[0073] S5, determine whether to continue target tracking, if yes, execute step S4, otherwise, exit target tracking;

[0074] S6. Use the confidence mechanism to determine whether to update the sample. If yes, execute step S7; otherwise, execute step S2.

[0075] S7, for the t-th frame image, at the target position P t and target scale S t Extract the current position feature F t，pos and the current scale feature F t，scale ;

[0076] S8, based on the current position feature F t，pos and the current scale feature F t，scale , use the Gaussian mixture model to generate a new training sample set, update the weights of the three correlation filters, update the three parallel correlation filters online, and execute step S2.

[0077] The present invention is based on the improvement of the traditional filter. Figure 2 The framework of the classic Discriminative Correlation Filter (DCF) is explicitly described. The DCF tracker effectively trains a correlation filter model in the frequency domain by applying machine learning techniques to distinguish between targets and backgrounds. It updates the model online using features extracted from the detection results of the current frame. First, the correlation filter is trained on the first frame (Initial Frame) with an initial bounding box and the desired output. Second, it extracts features (Extra Features) from the current frame (Current Frame) and multiplies the features by a cosine window (Cosine Window) to emphasize the central area. The features are then transformed into the frequency domain using the Fast Fourier Transform (FFT). Third, a response map is obtained by multiplying the correlation filter (Correlation Filter) with the extracted features. The response map is converted to the time domain by applying the Inverse Fast Fourier Transform (IFFT). Finally, the position of the maximum value of the response map is taken as the center position of the target in the current frame to complete target detection. Then, we extract new features (Extra Features) from the detection results and put them into the training sample set, adding the desired output (Desired Output) to train and update the relevant filters online in the frequency domain. Repeat the above steps to continue object detection for subsequent frames (Sebsequent Frame).

[0078] The flowchart of the Polynomial Correlation Filters (PCF) framework is as follows: Figure 3As shown in the figure, in the initial frame, the PCF tracker uses shared samples and sharp correlation output to train two parallel correlation filters PCF1 and PCF2 online in the frequency domain. In the current frame, PCF1 and PCF2 are used to track the target respectively. PCF1 and PCF2 obtain response maps Response Map1 and Response Map2 respectively based on shared features. By weighting the response maps of PCF1 and PCF2 (WeightedFusion), the PCF tracker applies Newton's method (NM) to accurately detect the position of the target. Then, it uses the Gaussian mixture model (GMM) to generate a new shared sample set by adding new samples or merging the two closest samples. It uses the new sample set to update the two parallel correlation filters every six frames.

[0079] Inspired by the DCF and PCF, this paper proposes a parallel three-branch correlation filter for target tracking in complex ocean environments based on a confidence mechanism. Using three parallel correlation filters makes tracking more robust. However, using more branches can lead to a sharp drop in speed, making tracking less sensitive to time. Furthermore, too many branches can lead to overfitting.

[0080] Similar to the PCF tracker, in the initial video frame, this application uses the prior information of the target object (such as the selection box, target position, scale, etc.) to initialize three parallel correlation filters, then performs feature extraction, and then fuses the values of the three correlation filters to update the position, and then uses the scale filter to calculate the scale of the object.

[0081] The three correlation filters can be expressed as:

[0082]

[0083]

[0084]

[0085] Among them, ε(h ψt ) represents the updated model of the ψth correlation filter, h ψ l Indicates the ψth correlation filter that needs to be updated, ψ=1, 2, 3, k represents the number of samples generated by window sliding, p represents the number of samples in the training sample set, represents the weight of the ψth correlation filter, represents the target sample, gs k represents the joint expected response distribution, λ represents the regularization parameter, and ω represents the spatial regularization parameter.

[0086] The update formula of the correlation filter weight is as follows:

[0087]

[0088] Among them, η ψ represents the learning rate of the ψth correlation filter, t represents the tth frame, represents the updated weight of the ψth correlation filter, represents the weight of the ψth correlation filter before update, ψ = 1, 2, 3. As can be seen, the value of the weight is determined by the learning rate. This application designs a different learning rate for each correlation filter, so that their weights are different. During each online training and update, the same sample has different weights for different correlation filters, which can effectively improve the robustness of the tracker to overcome the harsh conditions in the marine environment.

[0089] First response score y t，pos The calculation formula is:

[0090]

[0091] Among them, α and β are fusion factors. The solution is obtained by Fourier transforming equations (1) to (3). and and represents three parallel correlation filters, Represents the inverse Fourier transform.

[0092] Get the target position P t When the first response score y t，pos , three different fusion factors are used for weighted fusion, and the Newton method can be used to determine the target position. This application designs the fusion factors of three correlation filters respectively, which can effectively improve the performance of object trackers in complex ocean environments.

[0093] In order to determine the scale of the target, the DCF tracker is used to extract the scale features under different scale factors, and then the second response score is calculated using the following formula to determine the scale of the target. The scale corresponding to the response peak can be used as the target scale. t，scale The calculation formula is:

[0094]

[0095] in, and B t-1，scale represents the numerator and denominator of the scale-dependent filter of the previous frame, λ is the regularization parameter, λ≥0, Represents the inverse Fourier transform.

[0096] Generate a new training sample set specifically as follows:

[0097] The current position feature F t，pos and the current scale feature F t，scale A training sample set is added and the Gaussian mixture model is used to detect the similarity between samples. If the similarity between samples in the training sample set is less than a preset threshold, the two closest samples are merged.

[0098] In most cases, the coordinates corresponding to the maximum position of the peak are usually used as the target position, which is unreliable. Based on this, this application proposes a new confidence indicator (confidence response) to analyze the peak distribution of the relevant response and evaluate the tracking results. The confidence mechanism is used to determine whether to perform a sample update. Specifically:

[0099] Get the response map of each correlation filter and weight it to get the response map R. Select n relevant responses from the response map R, where R(i) represents the response value at the i-th position and p(R(i)) is the probability of the response value appearing. Calculate the expected μ of the n response values:

[0100]

[0101] Calculate the difference value CR using the following formula:

[0102]

[0103] If the CR value is less than the set threshold, a sample update is performed; otherwise, no sample update is performed. The purpose of is to evaluate the numerical difference between the largest n responses on the correlation response graph. When the CR value is less than the set threshold, it is considered reliable and a sample update is performed. The weights are updated, and the three correlation filters are trained and updated online. Otherwise, the sample and three correlation filter updates are stopped.

[0104] In this embodiment, a parallel three-branch correlation filter is deployed on a DJI drone using an onboard computer: Manifold 2-G. Manifold is DJI's drone development platform that allows developers to develop their own applications and install them on DJI aircraft expansion brackets. Manifold 2 has higher computing power than the previous generation, and it can support complex calculations such as artificial intelligence and machine learning. After the transplantation environment work is prepared, the parallel three-branch correlation filter target tracking algorithm is run based on this hardware platform and software environment and tracking is performed. The overall process is as follows: Figure 4As shown, first use the camera to obtain image data, transmit data based on ROS, and display it on the screen. Then use the mouse to frame the target to be tracked on the image, process the transmitted data, and call the parallel three-branch correlation filter target tracking algorithm to track the target. This can prevent model drift and effectively reduce the impact of various interferences on the filter, achieving good tracking effects. It plays an excellent role in complex marine environments and can solve target tracking losses caused by wave and ship interference, light changes, rapid movement, etc. in complex marine environments.

[0105] The present application also provides a target tracking system based on a parallel three-branch correlation filter, comprising an image acquisition module and a tracking module. The image acquisition module is used to acquire a continuous image sequence and feed it into the tracking module. The tracking module uses a parallel three-branch correlation filter based on a confidence mechanism to track the target, comprising the following steps:

[0106] S1, obtain the target position P1 and target scale S1 of the initial frame image, establish a training sample set, initialize three parallel correlation filters, and determine the weight size of each correlation filter;

[0107] S2, for the t-th frame image, at the target position P t-1 and target scale S t-1 Extract position feature Z t，pos and scale feature Z t，scale , t>1;

[0108] S3, based on position feature Z t，pos The first response score y is calculated with three parallel correlation filters t，pos , based on the scale feature Z t，scale The second response score y is calculated with three parallel correlation filters t，scale ;

[0109] S4. Obtain target position P based on the first response score t , obtain the target scale S based on the second response score t ;

[0110] S5, determine whether to continue target tracking, if yes, execute step S4, otherwise, exit target tracking;

[0111] S6. Use the confidence mechanism to determine whether to update the sample. If yes, execute step S7; otherwise, execute step S2.

[0112] S7, for the t-th frame image, at the target position P t and target scale S t Extract the current position feature F t，pos and the current scale feature Ft，scale ;

[0113] S8, based on the current position feature F t，pos and the current scale feature F t，scale , use the Gaussian mixture model to generate a new training sample set, update the weights of the three correlation filters, update the three parallel correlation filters online, and execute step S2.

[0114] The specific content of target tracking by the three-branch correlation filter in the target tracking system has been explained above and will not be repeated here.

[0115] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A target tracking method based on parallel three-branch correlation filter, characterized in that: Target tracking is performed using a parallel three-branch correlation filter based on a confidence mechanism, which includes the following steps: S1, obtain the target position P1 and target scale S1 of the initial frame image, establish a training sample set, initialize three parallel correlation filters, and determine the weight size of each correlation filter; S2, for the t-th frame image, at the target position P t-1 and target scale S t-1 Extract position feature Z t,pos and scale feature Z t,scale , t>1; S3, based on position feature Z t,pos The first response score y is calculated with three parallel correlation filters t,pos , based on the scale feature Z t,scale The second response score y is calculated with three parallel correlation filters t,scale ; S4. Obtain target position P based on the first response score t , obtain the target scale S based on the second response score t ; S5, determine whether to continue target tracking, if yes, execute step S4, otherwise, exit target tracking; S6. Use the confidence mechanism to determine whether to update the sample. If yes, execute step S7; otherwise, execute step S2. S7, for the t-th frame image, at the target position P t and target scale S t Extract the current position feature F t,pos and the current scale feature F t,scale ; S8, based on the current position feature F t,pos and the current scale feature F t,scale , use the Gaussian mixture model to generate a new training sample set, update the weights of the three correlation filters, update the three parallel correlation filters online, and execute step S2; The three correlation filters can be expressed as: Among them, ε(h ψt ) represents the updated model of the ψth correlation filter, h ψ l Indicates the ψth correlation filter that needs to be updated, ψ=1, 2, 3, k represents the number of samples generated by window sliding, p represents the number of samples in the training sample set, represents the weight of the ψth correlation filter, represents the target sample, gs k represents the joint expected response distribution, λ represents the regularization parameter, and ω represents the spatial regularization parameter; First response score y t,pos The calculation formula is: Among them, α and β are fusion factors. The solution is obtained by Fourier transforming equations (1) to (3). and and represents three parallel correlation filters, Represents the inverse Fourier transform.

2. A target tracking method based on parallel three-branch correlation filter according to claim 1, characterized in that: Second response score y t,scale The calculation formula is: in, and B t-1,scale represents the numerator and denominator of the scale-dependent filter of the previous frame, λ is the regularization parameter, λ≥0, Represents the inverse Fourier transform.

3. A target tracking method based on parallel three-branch correlation filter according to claim 1, characterized in that, Generate a new training sample set specifically as follows: The current position feature F t,pos and the current scale feature F t,scale A training sample set is added and the Gaussian mixture model is used to detect the similarity between samples. If the similarity between samples in the training sample set is less than a preset threshold, the two closest samples are merged.

4. A target tracking method based on parallel three-branch correlation filter according to claim 1, characterized in that: The update formula of the correlation filter weight is as follows: Among them, η ψ represents the learning rate of the ψth correlation filter, t represents the tth frame, represents the updated weight of the ψth correlation filter, represents the weight of the ψth correlation filter before updating, ψ = 1, 2, 3.

5. A target tracking method based on parallel three-branch correlation filter according to claim 1, characterized in that: The confidence mechanism is used to determine whether to update the sample: Get the response map of each correlation filter and weight it to get the response map R. Select n relevant responses from the response map R, where R(i) represents the response value at the i-th position and p(R(i)) is the probability of the response value appearing. Calculate the expected μ of the n response values: Calculate the difference value CR using the following formula: If the CR value is less than the set threshold, the sample is updated; otherwise, the sample is not updated.

6. A target tracking system based on a parallel three-branch correlation filter, characterized in that: The system includes an image acquisition module and a tracking module. The image acquisition module is used to acquire a continuous image sequence and feed it into the tracking module. The tracking module uses a parallel three-branch correlation filter based on a confidence mechanism to track the target, including the following steps: S1, obtain the target position P1 and target scale S1 of the initial frame image, establish a training sample set, initialize three parallel correlation filters, and determine the weight size of each correlation filter; S2, for the t-th frame image, at the target position P t-1 and target scale S t-1 Extract position feature Z t,pos and scale feature Z t,scale , t>1; S3, based on position feature Z t,pos The first response score y is calculated with three parallel correlation filters t,pos , based on the scale feature Z t,scale The second response score y is calculated with three parallel correlation filters t,scale ; S4. Obtain target position P based on the first response score t , obtain the target scale S based on the second response score t ; S5, determine whether to continue target tracking, if yes, execute step S4, otherwise, exit target tracking; S6. Use the confidence mechanism to determine whether to update the sample. If yes, execute step S7; otherwise, execute step S2. S7, for the t-th frame image, at the target position P t and target scale S t Extract the current position feature F t,pos and the current scale feature F t,scale ; S8, based on the current position feature F t,pos and the current scale feature F t,scale , use the Gaussian mixture model to generate a new training sample set, update the weights of the three correlation filters, update the three parallel correlation filters online, and execute step S2; The three correlation filters can be expressed as: Among them, ε(h ψt ) represents the updated model of the ψth correlation filter, h ψ l Indicates the ψth correlation filter that needs to be updated, ψ=1, 2, 3, k represents the number of samples generated by window sliding, p represents the number of samples in the training sample set, represents the weight of the ψth correlation filter, represents the target sample, gs k represents the joint expected response distribution, λ represents the regularization parameter, and ω represents the spatial regularization parameter; First response score y t,pos The calculation formula is: Among them, α and β are fusion factors. The solution is obtained by Fourier transforming equations (1) to (3). and and represents three parallel correlation filters, Represents the inverse Fourier transform.

7. A target tracking system based on parallel three-branch correlation filter according to claim 6, characterized in that: The confidence mechanism is used to determine whether to update the sample: Get the response map of each correlation filter and weight it to get the response map R. Select n relevant responses from the response map R, where R(i) represents the response value at the i-th position and p(R(i)) is the probability of the response value appearing. Calculate the expected μ of the n response values: Calculate the difference value CR using the following formula: If the CR value is less than the set threshold, the sample is updated; otherwise, the sample is not updated.

Images