A small target tracking method for infrared moving platform based on complementary correlation filter

Abstract

The present invention discloses a method for tracking small targets on an infrared moving platform based on a complementary correlation filter, which belongs to the field of intelligent processing of remote sensing video targets. The method comprises the following steps: S1, calculating a position response map of a target based on a complementary correlation filter, and judging whether occlusion occurs. If the target is not occluded, adopting the position with the largest response value as the position of the target; otherwise, performing subsequent steps; S2, adopting a multi-strategy combined image matching method to align two frames of images; S3, adopting an inter-frame difference method to detect moving targets and obtain a moving target set; S4, adopting a Kalman filter to predict the motion state of the target, and adopting the target in the moving target set that is closest to the position predicted by the Kalman filter as the final target; the complementary correlation filter proposed by the present invention has strong small target tracking and anti-occlusion capabilities under a moving platform, and can effectively improve the accuracy of infrared small target tracking.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing video intelligent processing, and in particular relates to a small target tracking method of an infrared moving platform based on a complementary correlation filter. Background Art

[0002] With the widespread use of mobile platforms such as drones, missiles, and ships in both military and civilian sectors, accurate and stable tracking of small infrared targets in complex, moving environments has become a key technology. Small infrared targets are typically small and have weak signals. This is particularly true on moving platforms, where the combined effects of platform motion and environmental interference make target tracking more complex and challenging. Target tracking technology based on correlation filters, with its high computational efficiency and strong anti-interference capabilities, has become a potential solution to this problem. However, practical applications still face numerous technical challenges.

[0003] First, in complex, dynamic scenes, small targets are often obscured by other objects in the environment. This occlusion can prevent tracking algorithms from acquiring valid information about the target, leading to target loss. To address this issue, designing effective occlusion handling mechanisms that ensure robust tracking of small targets even in the presence of partial occlusion remains a core challenge in the development of this technology.

[0004] Secondly, small infrared targets typically appear as small point objects. Especially when observed from long distances, their feature information is extremely limited and susceptible to noise interference. Combined with the inherent thermal and sensor noise of infrared imaging systems, the signal contrast between small targets and the background is very low, and the target features appear extremely weak. This weak feature makes it difficult for traditional feature extraction and matching algorithms to operate stably in complex environments. Therefore, it is necessary to optimize the design of the tracker to fully utilize feature information from various sources and enhance the trackability of small targets.

[0005] Finally, the platform's own motion can introduce additional disturbances, especially at high speeds or in non-stationary conditions. Image jitter and blur caused by platform motion can significantly impact the accuracy and stability of the tracking algorithm. The unpredictability of platform motion can also complicate the target's trajectory, increasing the difficulty of tracking. Summary of the Invention

[0006] In order to solve the problems of weak target features, frequent occlusion and rapid background motion in the tracking of small targets on infrared moving platforms, the present invention provides a method for tracking small targets on infrared moving platforms based on complementary correlation filters. By using key technologies such as correlation filters, image matching, inter-frame difference method and Kalman filtering, stable tracking of small targets on infrared moving platforms is achieved.

[0007] In order to solve the technical problem, the technical solution of the present invention is:

[0008] A method for tracking small targets on an infrared moving platform based on a complementary correlation filter, the method comprising:

[0009] S1: Use the complementary correlation filter to calculate the target position response map and determine whether occlusion occurs. If the target is not occluded, the position with the largest response value is used as the target position;

[0010] S2: If the target is occluded, a multi-strategy image matching method is used to align the previous and next frame images and correct the background motion;

[0011] S3: Moving target detection, using the inter-frame difference method to detect moving targets and generate a candidate set of moving targets;

[0012] S4: Use the Kalman filter to predict the motion state of the target, and take the target closest to the predicted position in the candidate set of moving targets as the final confirmed target position.

[0013] Furthermore, the step S1 specifically includes:

[0014] First, the target position response map is calculated based on the complementary correlation filter;

[0015] Based on the gradient histogram features and color attribute features, two response maps are generated, and then the two response maps are linearly fused to obtain the final fused response map. The specific position of the target is determined based on the fused response map. The correlation filter tracking algorithm uses an image block x with a pixel size of P×Q for training; all cyclic shifts of this image block x are p,q (p,q)∈{0,1,…P-1}×{0,1,…Q-1} is used as a training sample, and the label is a Gaussian label r p,q ; The goal is to find an optimal w that satisfies the following objective function:

[0016]

[0017] Among them, φ represents a nonlinear mapping and λ represents a regularization term. Solving the above equation using fast Fourier transform yields:

[0018]

[0019]

[0020] in, represents the Fourier transform, Represents the inverse Fourier transform; in the next frame, an image block z of the same size as x is cut out from the new image; the response graph of z is calculated as:

[0021]

[0022] Where ⊙ represents the Hadamard product; in the next frame, the position of the target is given by f template The position of the maximum value is determined; in order to make the tracker have a certain memory capacity, the online update mechanism is defined as follows:

[0023]

[0024] Where η is a learning rate parameter, t is the sequence number of the current frame, and two related filter models are trained based on the gradient histogram features and color attribute features to generate the gradient histogram feature response map and the color attribute feature response map, respectively. These two response maps are then linearly weighted fused to obtain the final fused response map. The response map can be used to effectively determine whether the target is occluded. Specifically, the maximum value of the response map is used for judgment. If the maximum value is less than a certain threshold, the target is considered to be occluded.

[0025] Furthermore, the step S2 specifically includes:

[0026] After the target is occluded, the inter-frame difference method and the Kalman filter are used to determine the position of the moving target. A multi-strategy joint image matching motion parameter estimation is used to extract the inter-frame translation parameters of the scene motion based on feature matching and phase correlation. This method combines ORB feature matching and image frequency domain phase correlation methods to improve robustness and stability.

[0027] First, the original video frame is converted into a single-channel grayscale image, and a nR-ORB algorithm is used to detect feature points P = {p1, p2, ..., p n}, each feature point p i With location information; generate a descriptor D for each feature point i , capturing local image features; the nR-ORB algorithm simplifies the operation used in the original ORB algorithm to process significant image rotation, that is, canceling the calculation of the main direction of the feature points in the ORB algorithm, making the reference directions of the feature point descriptors consistent, and canceling the rotation operation of the feature descriptor;

[0028] Calculate the distance between the descriptors of the feature points of the two frames I1 and I2 Finding the best matching pair The RANSAC algorithm is used to eliminate mismatches. The initial translation (Δx1, Δy1) is estimated through the correspondence between feature points and applied to the video frame i1 to obtain the transformed frame i1′:

[0029] i1′(x,y)=I1(x+Δx1,y+Δy1)

[0030] Perform Fourier transform on the transformed frame I1′ and the reference frame I2:

[0031]

[0032]

[0033] Compute the product of the complex conjugates of F1′ and F2:

[0034]

[0035] Perform inverse Fourier transform on the product result to obtain the related image:

[0036]

[0037] Find the maximum position (x0, y0) in the correlation image R(x, y), which corresponds to the fine translation (Δx2, Δy2) between the two images; superimpose the fine translation with the preliminary translation to obtain the final inter-frame motion vector:

[0038] Δx=Δx1+Δx2

[0039] Δy=Δy1+Δy2。

[0040] Furthermore, the step S3 specifically includes:

[0041] First, the corresponding pixel values ​​of the current frame are subtracted from the previous frame to obtain a differential image; the non-zero area in the differential image usually indicates the part with motion; then, the differential image is thresholded to convert larger pixel changes to white, indicating the motion area, and smaller or no changes to black, indicating the background; finally, morphological operations are performed on the binary image to extract the continuous moving target area.

[0042] Furthermore, the step S4 specifically includes:

[0043] In terms of target motion estimation, the Kalman filter is used to predict the position and velocity of the moving target; the Kalman filter is for discrete-time linear dynamic systems. A discrete-time linear dynamic system is defined as follows:

[0044] x t =F t-1 x t-1 +G t-1 w t-1

[0045] z t =H t x t +v t

[0046] Among them, x t represents the state variable of the system at time t, w t-1 represents the state noise vector, whose covariance matrix is ​​Q; F t-1 represents the state transition matrix, G t-1 represents the noise distribution matrix; z t represents the observation vector, H t represents the observation matrix, v t represents the observation noise vector, whose covariance matrix is ​​R;

[0047] Let x t =[p x , p y , v x , v y ] T is the state vector of the target, where p x , p y is the target position, v x , v y is the target speed; let z t is the position calculated by the correlation filtering algorithm; considering that the time between two frames is very short and assuming that the target is in a uniform motion state, the state matrix and observation matrix are defined as:

[0048]

[0049]

[0050] After obtaining the target observation vector z t Finally, the state vector of the target is estimated based on the Kalman filter; the Kalman filter algorithm is divided into three steps: prediction process, Kalman gain calculation, and update process;

[0051] First is the prediction process:

[0052]

[0053]

[0054] in represents the optimal state vector at time t-1, represents the predicted state vector at time t;

[0055] The second step is the Kalman gain calculation process:

[0056]

[0057] The third step is the update process:

[0058]

[0059]

[0060] When the target is occluded, the position of the next frame is directly calculated using the prediction process; the inter-frame difference method can be used to obtain a set of all possible moving targets in the image, and the position of the target can be predicted by the Kalman filter, so it is necessary to further screen the targets in the moving target set; the target in the moving target set that is closest to the position predicted by the Kalman filter is used as the final moving target.

[0061] Compared with the prior art, the advantages of the present invention are:

[0062] The present invention provides a highly robust and stable infrared small target tracking method by combining complementary correlation filters, image registration, inter-frame difference, and Kalman filtering. The specific beneficial effects are as follows:

[0063] Enhanced anti-occlusion capability: The fusion response map generated by the complementary correlation filter is used to accurately determine the target position. Even in the case of temporary occlusion, the Kalman filter prediction and image registration strategy can still effectively track the target and avoid target loss.

[0064] Efficient background interference suppression: Through multi-strategy image registration (combined with ORB feature matching and phase correlation), the influence of background motion between frames can be effectively eliminated, making the tracking system more adaptable to the interference caused by dynamic background in complex scenes.

[0065] Accurate target position prediction: Kalman filtering is used to predict and update the target's position and velocity in real time, which can quickly screen out the most likely target position from the candidate targets provided by the inter-frame difference method, thereby improving the continuity and accuracy of target tracking.

[0066] Real-time and stability: This solution uses fast Fourier transform (FFT) to improve computational efficiency and combines gradient histogram and color feature response map to achieve fast detection, which helps to achieve high real-time tracking in an environment with limited hardware resources.

[0067] The present invention can achieve high precision, strong robustness and good real-time response effect in infrared small target tracking, and is particularly suitable for dynamic scenes with complex backgrounds and short-term occlusions. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 , the main flow chart of the infrared moving platform small target tracking method based on complementary correlation filter of the present invention. DETAILED DESCRIPTION

[0069] The specific implementation of the present invention is described below in conjunction with examples:

[0070] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions under which the present invention can be implemented. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the efficacy and purpose that can be achieved by the present invention.

[0071] At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" quoted in this specification are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments to their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0072] Example 1:

[0073] The following are the input and output (data flow) of each step in the present invention:

[0074] Step 1: Target detection and occlusion judgment

[0075] Input: current frame image

[0076] process:

[0077] Generating position response maps based on complementary correlation filters.

[0078] The gradient histogram features and color attribute features are fused to obtain the fused response map.

[0079] Whether occlusion occurs is determined based on the maximum value of the response graph.

[0080] Output:

[0081] If not occluded: the maximum position of the response map (i.e. the target position).

[0082] If blocked: go to the next step for image registration.

[0083] Step 2: Image Registration

[0084] Input: current frame image, previous frame image

[0085] process:

[0086] The translation parameters between frames are estimated using ORB feature matching and phase correlation method.

[0087] Transform the current frame image to eliminate the influence of background motion.

[0088] Output:

[0089] The current frame image after translation.

[0090] Inter-frame translation vector, used for image calibration.

[0091] Step 3: Moving object detection

[0092] Input: The current frame image and the previous frame image after translation

[0093] process:

[0094] The inter-frame difference method is used to detect motion regions.

[0095] The difference image is binarized and morphologically operated to extract the candidate region of the moving target.

[0096] Output:

[0097] Moving target candidate set.

[0098] Step 4: Kalman filter prediction and target confirmation

[0099] Input: moving target candidate set, target position and speed at the last moment

[0100] process:

[0101] Use the Kalman filter to predict the next moment's position and velocity of the target.

[0102] The target positions in the moving target candidate set are compared with the positions predicted by the Kalman filter, and the target closest to the predicted position is selected as the final target.

[0103] Output:

[0104] Confirm the final target position (update the target position and speed, and continue tracking for the next frame).

[0105] Example 2:

[0106] This embodiment is a specific development of a method for tracking small infrared targets based on complementary correlation filters provided by the present invention. The method flow is as follows: Figure 1 As shown, the following steps are included:

[0107] S1. Calculate the target's position response map based on the complementary correlation filter and determine whether occlusion occurs. If the target is not occluded, use the position with the largest response value as the target's position; otherwise, proceed to the subsequent steps.

[0108] First, the target position response map is calculated based on the complementary correlation filter. Specifically, two response maps are generated based on the gradient histogram features and color attribute features, and then the two response maps are linearly fused to obtain the final fused response map. The specific position of the target is determined based on the fused response map. The correlation filter tracking algorithm uses an image block x with a pixel size of P×Q for training. All cyclic shifts x of this image block are p，q (p, q)∈{0, 1, ...P-1}×{0, 1, ...Q-1} is used as a training sample, and the label is a Gaussian label r p，q The goal is to find an optimal w that satisfies the following objective function:

[0109]

[0110] Where φ represents a nonlinear mapping and λ represents a regularization term. Solving the above equation using the fast Fourier transform yields:

[0111]

[0112]

[0113] in represents the Fourier transform, Represents the inverse Fourier transform. In the next frame, an image block z of the same size as x is cropped from the new image. The response graph of z is calculated as:

[0114]

[0115] Where ⊙ represents the Hadamard product. In the next frame, the position of the target is given by f template In order to make the tracker have a certain memory capacity, the online update mechanism is defined as follows:

[0116]

[0117] Where η is a learning rate parameter, t is the sequence number of the current frame, and two correlation filter models are trained based on the gradient histogram features and color attribute features to generate a gradient histogram feature response map and a color attribute feature response map, respectively. These two response maps are then linearly weighted fused to obtain a final fused response map. The response map can effectively determine whether the target is occluded. Specifically, the present invention uses the maximum value of the response map to make the judgment. If the maximum value is less than a certain threshold, the target is considered to be occluded.

[0118] S2, using a multi-strategy combined image matching method to align the two frames of images;

[0119] After the target is occluded, the inter-frame difference method and the Kalman filter are used to jointly determine the position of the moving target. However, since the background of the image will move, it will greatly affect the accuracy of the inter-frame difference method and the Kalman filter. Therefore, it is necessary to use image matching to match the two frames before and after. In the technical solution of the present invention, a multi-strategy joint image matching motion parameter estimation is adopted to jointly extract the inter-frame translation parameters of the scene motion based on feature matching and phase correlation. This method combines ORB (Oriented FAST and Rotated BRIEF) feature matching and image frequency domain phase correlation technology to improve robustness and stability.

[0120] First, the original video frame is converted into a single-channel grayscale image, and a nR-ORB (nonRotation-ORB) algorithm is used to detect feature points P = {p1, p2, ..., p n}, each feature point p i With location information. Generate a descriptor D for each feature point i , capturing local image features. The nR-ORB algorithm simplifies the operations used in the original ORB algorithm to handle significant image rotation. That is, it cancels the calculation of the main direction of the feature points in the ORB algorithm, keeps the base directions of the feature point descriptors consistent, and cancels the rotation operation of the feature descriptors.

[0121] Calculate the distance between the descriptors of the feature points of the two frames I1 and I2 Finding the best matching pair The RANSAC algorithm is used to eliminate mismatches. The initial translation (Δx1, Δy1) is estimated through the feature point correspondence and applied to the video frame I1 to obtain the transformed frame I1′:

[0122] I1′(x, y)=I1(x+Δx1, y+Δy1)

[0123] Furthermore, the transformed frame I1′ and the reference frame I2 are subjected to Fourier transform:

[0124]

[0125]

[0126] Compute the product of the complex conjugates of F1′ and F2:

[0127]

[0128] Perform inverse Fourier transform on the product result to obtain the related image:

[0129]

[0130] Find the maximum position (x0, y0) in the correlation image R(x, y), which corresponds to the fine translation (Δx2, Δy2) between the two images. The fine translation is superimposed on the preliminary translation to obtain the final inter-frame motion vector:

[0131] Δx=Δx1+Δx2

[0132] Δy=Δy1+Δy2

[0133] S3, using the inter-frame difference method to detect moving targets and obtain a moving target set;

[0134] The inter-frame differencing method is a simple and effective image processing technique for moving target detection. Its basic principle is to detect changing regions in the scene by performing pixel-level differencing operations on consecutive frame images, thereby identifying moving targets. The specific steps are as follows: First, the corresponding pixel values ​​of the current frame are subtracted from the previous frame to obtain a difference image. Non-zero regions in this difference image generally indicate areas of motion. Second, the difference image is thresholded, converting larger pixel changes to white (indicating motion) and smaller or no changes to black (indicating background). Finally, morphological operations are performed on the binary image to further extract the continuous moving target regions.

[0135] S4. Use the Kalman filter to predict the motion state of the target, and take the target in the moving target set that is closest to the position predicted by the Kalman filter as the final target.

[0136] In terms of target motion estimation, the Kalman filter is used to predict the position and velocity of the moving target. The Kalman filter is designed for discrete-time linear dynamic systems. A discrete-time linear dynamic system is defined as follows:

[0137] x t =F t-1 x t-1 +G t-1 w t-1

[0138] z t =H t x t +v t

[0139] Among them, x t represents the state variable of the system at time t, w t-1 represents the state noise vector, whose covariance matrix is ​​Q. t-1 represents the state transition matrix, G t-1 represents the noise distribution matrix; z t represents the observation vector, H trepresents the observation matrix, v t represents the observation noise vector, whose covariance matrix is ​​R.

[0140] Assume x t =[p x , p y , v x , v y ] T is the state vector of the target, where p x , p y is the target position, v x , v y is the target speed. Let z t is the observation vector, which is the position calculated by the correlation filtering algorithm in this invention. Since the time between two frames is very short, it can be assumed that the target is in a uniform motion state, so the state matrix and observation matrix can be defined as:

[0141]

[0142]

[0143] After obtaining the target observation vector z t After that, we can estimate the state vector of the target based on Kalman filtering. The Kalman filter algorithm is divided into three steps: prediction process, Kalman gain calculation, and update process. The first is the prediction process:

[0144]

[0145]

[0146] in represents the optimal state vector at time t-1, represents the predicted state vector at time t.

[0147] The second step is the Kalman gain calculation process:

[0148]

[0149] The third step is the update process:

[0150]

[0151]

[0152] When a target is occluded, its observation vector is unreliable. In this case, the next frame's position is directly calculated using a prediction process. Using inter-frame differencing, we can obtain a set of all possible moving targets in the image. Using Kalman filtering, we can predict the target's position. Therefore, we need to further filter the targets in the moving target set. This invention proposes selecting the target in the moving target set that is closest to the Kalman filter-predicted position as the final moving target.

[0153] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

[0154] Many other changes and modifications can be made without departing from the spirit and scope of the present invention. It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. A method for tracking small targets on an infrared moving platform based on complementary correlation filters, characterized in that: The method comprises: S1: Use the complementary correlation filter to calculate the target position response map and determine whether occlusion occurs. If the target is not occluded, the position with the largest response value is used as the target position; S2: If the target is occluded, a multi-strategy image matching method is used to align the previous and next frame images and correct the background motion; The step S2 specifically includes: After the target is occluded, the inter-frame difference method and the Kalman filter are used to determine the position of the moving target. A multi-strategy joint image matching motion parameter estimation is used to extract the inter-frame translation parameters of the scene motion based on feature matching and phase correlation. This method combines ORB feature matching and image frequency domain phase correlation methods to improve robustness and stability. First, the original video frame is converted into a single-channel grayscale image, and a nR-ORB algorithm is used to detect feature points P = {p1, p2, ..., p n }, each feature point p i With location information; generate a descriptor D for each feature point i , capturing local image features; the nR-ORB algorithm simplifies the operation used in the original ORB algorithm to process significant image rotation, that is, canceling the calculation of the main direction of the feature points in the ORB algorithm, making the reference directions of the feature point descriptors consistent, and canceling the rotation operation of the feature descriptor; Calculate the distance between the descriptors of the feature points of the two frames I1 and I2 Finding the best matching pair The RANSAC algorithm is used to eliminate mismatches. The initial translation (Δx1, Δy1) is estimated through the correspondence between feature points and applied to the video frame I1 to obtain the transformed frame I′1: I1′(x,y)=I1(x+Δx1,y+Δy1) Perform Fourier transform on the transformed frame I1′ and the reference frame I2: Compute the product of the complex conjugates of F1′ and F2: Perform inverse Fourier transform on the product result to obtain the related image: Find the maximum position (x0, y0) in the correlation image R(x, y), which corresponds to the fine translation (Δx2, Δy2) between the two images; superimpose the fine translation with the preliminary translation to obtain the final inter-frame motion vector: Δx=Δx1+Δx2 Δy=Δy1+Δy2; S3: Moving target detection, using the inter-frame difference method to detect moving targets and generate a candidate set of moving targets; S4: Use the Kalman filter to predict the motion state of the target, and take the target closest to the predicted position in the candidate set of moving targets as the final confirmed target position.

2. The method for tracking small targets on an infrared moving platform based on a complementary correlation filter according to claim 1, characterized in that: The step S1 specifically includes: First, the target position response map is calculated based on the complementary correlation filter; Based on the gradient histogram features and color attribute features, two response maps are generated, and then the two response maps are linearly fused to obtain the final fused response map. The specific position of the target is determined based on the fused response map. The correlation filter tracking algorithm uses an image block x with a pixel size of P×Q for training; all cyclic shifts of this image block x are p,q (p,q)∈{0,1,…P-1}×{0,1,…Q-1} is used as a training sample, and the label is a Gaussian label r p,q ; The goal is to find an optimal w that satisfies the following objective function: Among them, φ represents a nonlinear mapping and λ represents a regularization term. Solving the above equation using fast Fourier transform yields: in, represents the Fourier transform, Represents the inverse Fourier transform; in the next frame, an image block z of the same size as x is cut out from the new image; the response graph of z is calculated as: Where ⊙ represents the Hadamard product; in the next frame, the position of the target is given by f template The position of the maximum value is determined; in order to make the tracker have a certain memory capacity, the online update mechanism is defined as follows: Where η is a learning rate parameter, t is the sequence number of the current frame, and two related filter models are trained based on the gradient histogram features and color attribute features to generate the gradient histogram feature response map and the color attribute feature response map, respectively. These two response maps are then linearly weighted fused to obtain the final fused response map. The response map can be used to effectively determine whether the target is occluded. Specifically, the maximum value of the response map is used for judgment. If the maximum value is less than a certain threshold, the target is considered to be occluded.

3. The method for tracking small targets on an infrared moving platform based on complementary correlation filters according to claim 1, characterized in that: The step S4 specifically includes: In terms of target motion estimation, the Kalman filter is used to predict the position and velocity of the moving target; the Kalman filter is for discrete-time linear dynamic systems. A discrete-time linear dynamic system is defined as follows: x t =F t-1 x t-1 +G t-1 w t-1 z t =H t x t +v t Among them, x t represents the state variable of the system at time t, w t-1 represents the state noise vector, whose covariance matrix is ​​Q; F t-1 represents the state transition matrix, G t-1 represents the noise distribution matrix; z t represents the observation vector, H t represents the observation matrix, v t represents the observation noise vector, whose covariance matrix is ​​R; Let x t =[p x ,p y ,v x ,v y ] T is the state vector of the target, where p x ,p y is the target position, v x ,v y is the target speed; let z t is the position calculated by the correlation filtering algorithm; considering that the time between two frames is very short and assuming that the target is in a uniform motion state, the state matrix and observation matrix are defined as: After obtaining the target observation vector z t Finally, the state vector of the target is estimated based on the Kalman filter; the Kalman filter algorithm is divided into three steps: prediction process, Kalman gain calculation, and update process; First is the prediction process: in represents the optimal state vector at time t-1, represents the predicted state vector at time t; The second step is the Kalman gain calculation process: The third step is the update process: When the target is occluded, the position of the next frame is directly calculated using the prediction process; the inter-frame difference method can be used to obtain a set of all possible moving targets in the image, and the position of the target can be predicted by the Kalman filter, so it is necessary to further screen the targets in the moving target set; the target in the moving target set that is closest to the position predicted by the Kalman filter is used as the final moving target.

Images