A correlation filter tracking method against occlusion and similar target interference

Abstract

The application provides an anti-occlusion and anti-similar target interference correlation filtering tracking method, which can realize more accurate tracking in the case of occlusion and effectively solve the problem of similar target interference after target reappearance under the premise of ensuring high running speed; compared with the method of only using Kalman filter prediction, the application uses a fast NCC algorithm for detection in a larger area instead of only relying on the result of Kalman prediction, so that the tracking is more accurate in the case of occlusion; meanwhile, the fast NCC algorithm used in the application can ensure fast operation speed when realizing large range detection; finally, based on the advantage that NCC has good distinguishing ability for similar targets, the application introduces an NCC algorithm and a KCF algorithm fusing motion information to simultaneously perform detection after target reappearance, so that compared with the traditional KCF combined with Kalman algorithm, the application can better solve the problem of similar target interference after target reappearance.

Description

Technical Field

[0001] This invention belongs to the field of image target tracking, and in particular relates to a correlation filtering tracking method that is resistant to occlusion and interference from similar targets. Background Technology

[0002] Image target tracking has two main research directions: generative methods and discriminative methods.

[0003] The former method models the features of the tracked object based on the selected tracking region in the first frame, and then detects and finds the region whose features are most similar to those of the tracked object in subsequent frames. However, due to the difficulty in modeling sufficient and effective object features, generative methods have poor robustness and are difficult to apply to complex scenes with many changes and disturbances. Specifically, when disturbances such as changes in lighting, object deformation, and occlusion occur in subsequent frames, the target's appearance features will change significantly. The target features at this point will differ considerably from those established solely based on the first frame, leading to tracking failure.

[0004] Discriminative methods track targets by training an online classifier that can distinguish between foreground and background. Specifically, positive and negative samples are sampled near the target, a discriminative template is trained, and this template is used to locate the target in the next frame. This location is then used for a new round of sampling and training. The introduction of a large number of negative samples fully utilizes the background information around the object, giving the discriminative model a stronger ability to distinguish between foreground and background, thus providing better robustness to scenes with changes in lighting and deformation. Among these methods, correlation filtering, which uses cyclic sampling, has a high running speed and has become a major research direction in target tracking.

[0005] KCF is a classic correlation filter-based tracking method that is widely used.

[0006] The main steps of the KCF algorithm are as follows:

[0007] 1) Select the target region to be tracked in the first frame and calculate the template;

[0008] 2) Use the template in the next frame to detect and obtain the target's position;

[0009] 3) If the current frame is the last frame, then end; otherwise, recalculate the template at the new position and return to step 2).

[0010] Since KCF updates the template in each new frame, it is an online learner and has good robustness to changes in the appearance of the target during tracking.

[0011] KCF, as an improved version of the traditional correlation filter, has better adaptability to nonlinear changes, such as illumination changes, rotational stretching deformation, etc., due to the introduction of kernel functions, FHOG feature representation, and online learning mechanisms. However, it still cannot effectively handle occlusion problems.

[0012] There are two key reasons why it cannot handle occlusion issues: 1. The online learning mechanism cannot identify whether occlusion occurs. Therefore, when occlusion occurs, it learns the characteristics of the occluding object, resulting in an incorrect tracking trajectory. 2. When the occlusion disappears, it cannot accurately find the new location of the target after the target reappears and restart the tracking process. This is because the learned identification template has been contaminated by the occluding object, and interference from similar objects sometimes occurs during the reappearance stage.

[0013] To address the issues of predicting and retrieving target locations when they disappear, existing improved methods primarily use Kalman filters to model target motion and then predict the target location in occluded frames. However, targets rarely move at a standard linear constant velocity, making it difficult for Kalman filters to accurately predict their reappearance location. Furthermore, the default search region of the correlation filter is the same size as the learned template; when the predicted location differs significantly from the actual reappearance location (which frequently occurs during long-term occlusion), it is difficult to obtain an accurate new location using only the traditional KCF detection process. Additionally, since the robustness of the correlation filter to illumination changes relies mainly on FHOG feature extraction, artificially increasing the search region drastically increases the FHOG feature extraction time, thus affecting tracking speed. Summary of the Invention

[0014] To address the problem of interference from similar targets, this invention provides a correlation filtering tracking method that resists occlusion and interference from similar targets. This method can achieve more accurate tracking when occluded and effectively solves the problem of interference from similar targets after the target reappears.

[0015] A correlation filtering tracking method that is resistant to occlusion and interference from similar targets includes the following steps:

[0016] S1: Obtain the WPSR response value corresponding to the current frame image and determine whether the WPSR response value is greater than the set threshold t1. If yes, the target is not occluded in the current frame image. Use the KCF algorithm to detect the position of the target in the current frame image. If no, the target is occluded in the current frame image. Proceed to step S2.

[0017] S2: Use a Kalman filter to obtain the predicted position of the target in the current frame image, and obtain the NCC response map corresponding to the sub-map formed by the set neighborhood centered on the predicted position;

[0018] S3: Obtain a motion information weighted map based on the motion information corresponding to the nearest uncrowded image to the current frame image, and multiply the motion information weighted map with the NCC response map corresponding to the current frame image to obtain an improved NCC response map;

[0019] S4: Determine whether the maximum response value in the improved NCC response map exceeds the set threshold t2. If yes, proceed to step S5. If no, expand the set neighborhood and reselect the next frame image to execute steps S2 to S4.

[0020] S5: Obtain the KCF template corresponding to the nearest uncrowded image to the current frame image, and use the KCF template to perform KCF detection sequentially at the first M positions selected from the improved NCC response map that exceed the set threshold t2, to obtain M KCF detection response values, and then proceed to step S6, where M is at most 3;

[0021] S6: Determine whether there is at least one KCF detection response value greater than the set threshold t3. If yes, it is considered that the position of the target in the current frame image has been detected, and the position of the maximum value among the KCF detection response values ​​greater than the set threshold t3 is taken as the position of the target in the current frame image. If no, the next frame image is selected again and steps S2 to S4 are executed.

[0022] Furthermore, the calculation method for the WPSR response value corresponding to the current frame image is as follows:

[0023]

[0024] Where P is the maximum value in the KCF response graph corresponding to the current frame image, and μ and σ represent the mean and standard deviation of the sidelobe region, respectively.

[0025] Furthermore, in step S3, a motion information weighted map w is obtained based on the motion information corresponding to the nearest uncrowded image to the current frame image. x,y Specifically:

[0026]

[0027] Where, σ x σ y , respectively, are the standard deviations of the target's position along the x-axis and y-axis. If the target remains occluded, the standard deviation will gradually increase; x and y are the distances of the target along the x-axis and y-axis from the predicted position obtained by the Kalman filter in this frame image, respectively; W is the width of the set neighborhood; and H is the length of the set neighborhood.

[0028] Furthermore, the method for obtaining the KCF template corresponding to the nearest uncluttered image to the current frame image in step S5 is as follows:

[0029] F t =(1-η-η) max )F t-1 +η max F max +ηF t ′

[0030] Where t is the frame number of the nearest uncrowded image to the current frame, and F t For the optimized KCF template of the t-th frame image that is the nearest uncrowded image to the current frame image, F t-1 F is the optimized KCF template of the frame preceding the nearest t-th frame image that is not occluded to the current frame image. t ' is the original KCF template of the nearest uncrowded frame t to the current frame image, and η is the set baseline template learning rate. max F is the template learning rate corresponding to the image with the highest WPSR response value in the historical frame images; max An optimized KCF template for the image with the highest WPSR response value in historical frame images.

[0031] Furthermore, the method for expanding the defined neighborhood in step S4 is as follows:

[0032] Assume that the length of the neighborhood before expansion is k*h and the width is k*w, where h is the set minimum detection unit length, w is the set minimum detection unit width, and k represents the number of times in step S4 the judgment result is that the maximum response value in the improved NCC response map does not exceed the set threshold t2.

[0033] Each time k increases by 1, it is multiplied by the length and width of the smallest detection unit to obtain the expanded set neighborhood.

[0034] Furthermore, when the defined neighborhood is expanded to the size of the entire graph, the defined neighborhood is downsampled by a factor of two, and steps S2 to S4 are executed using the downsampled defined neighborhood.

[0035] Furthermore, when the judgment result in step S6 is yes, the Kalman filter is restarted first, and then the target's motion information and position information are updated.

[0036] Furthermore, if the judgment result in step S1 is negative, the Kalman filter update step is paused.

[0037] Furthermore, the NCC response map corresponding to the sub-image in step S2 is obtained by performing NCC operations on the sub-image using the three NCC templates with the highest WPSR response values ​​in the NCC template queue, and then filtering the three response maps formed by the channel maximum value. Each template in the NCC template queue is the target region of the five images with the highest WPSR response values ​​in the previous frame.

[0038] The NCC template queue is decayed before being used to obtain the NCC response graph corresponding to the subgraph. The decay method of the NCC template queue is as follows:

[0039]

[0040] Where i is the template number in the NCC template queue, with a value ranging from 1 to 5. This represents the value of the i-th NCC template in the NCC template queue at time T, where T represents the frame number of the current frame image, T-1 represents the frame number of the previous frame image, and β is the set attenuation factor.

[0041] Beneficial effects:

[0042] 1. This invention provides a correlation filtering tracking method that resists occlusion and interference from similar targets. It is a correlation filtering tracking method combining a fast NCC algorithm and a Kalman filter algorithm. This method can achieve more accurate tracking under occlusion while maintaining high operating speed, and effectively solves the problem of interference from similar targets after the target reappears. Compared to methods that only use Kalman filter prediction, this invention uses a fast NCC algorithm for detection over a larger area, rather than relying solely on Kalman prediction results, thus enabling more accurate tracking under occlusion. Simultaneously, the fast NCC algorithm used in this invention maintains a fast computation speed when performing large-scale detection. Finally, based on the advantage of NCC's good ability to distinguish similar targets, this invention introduces an NCC algorithm incorporating motion information and a KCF algorithm for simultaneous target detection. Therefore, compared to the traditional KCF combined with Kalman algorithm, it can better solve the problem of interference from similar targets after the target reappears.

[0043] 2. This invention provides a correlation filtering tracking method that resists occlusion and interference from similar targets. Based on the assumption that the difference between the center peak value when occlusion occurs and when occlusion does not occur is usually large, the WPSR response value used in this invention places more emphasis on the magnitude of the peak response compared to the traditional PSR response value. Therefore, this invention can improve the accuracy of occlusion judgment by comparing the magnitude between WPSR and a threshold.

[0044] 3. This invention provides a correlation filtering tracking method that resists occlusion and interference from similar targets. It uses an improved KCF template update mechanism, which can record the optimal template and better integrate historical information. This further reduces the impact of mislearning the features of the occluded object when entering a partially occluded state due to the insignificant decrease in WPSR value.

[0045] 4. This invention provides a correlation filtering tracking method that resists occlusion and interference from similar targets. While performing NCC detection, a motion information weighting mechanism is proposed, which can find a more accurate restart position. The solution to interference from similar targets is mainly due to the relatively accurate image matching capability of NCC and the fact that the normalization process of NCC can effectively alleviate the impact of illumination changes on the determination of interference from similar targets.

[0046] 5. This invention provides a correlation filtering tracking method that is resistant to occlusion and interference from similar targets. By leveraging the powerful foreground and background resolution capabilities of the KCF algorithm, secondary detection is performed at the possible recurrence locations detected by the NCC algorithm, which helps to further refine the location of the target's recurrence and improve the accuracy of target recurrence judgment. Attached Figure Description

[0047] Figure 1 The flowchart shows a correlation filtering tracking method for resisting occlusion and interference from similar targets provided by the present invention. Detailed Implementation

[0048] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0049] like Figure 1 As shown, a correlation filtering tracking method that resists occlusion and interference from similar targets includes the following steps:

[0050] S1: Obtain the WPSR response value corresponding to the current frame image and determine whether the WPSR response value is greater than the set threshold t1. If yes, the target is not occluded in the current frame image. Use the KCF algorithm to detect the position of the target in the current frame image. If no, the target is occluded in the current frame image. Proceed to step S2.

[0051] It should be noted that before obtaining the WPSR response value corresponding to the current frame image, the tracking target is selected, and the Kalman filter and tracker are initialized. WPSR stands for Weighted Peak Sidelobe Ratio, and the WPSR response value corresponding to the current frame image is calculated as follows:

[0052]

[0053] Where P is the maximum value in the KCF response map corresponding to the current frame image, and μ and σ represent the mean and standard deviation of the sidelobe region, respectively. Compared with the traditional PSR response value, the WPSR response value places more emphasis on the magnitude of the peak response, because the difference between the center peak value when occlusion occurs and when occlusion does not usually occur is large. Therefore, this invention compares the WPSR with the threshold and determines whether occlusion exists based on the comparison result, which can improve the accuracy of occlusion judgment.

[0054] It should be noted that the KCF algorithm is used to detect the target's position in the current frame image as follows:

[0055] The KCF tracking process is then performed. The KCF algorithm solves for the function f(z) by establishing a regularized least squares loss function as shown in the following formula.

[0056] f(z) = w T z

[0057] The output of this function is the similarity score between the template and the input region. By using a kernel function, the original image space can be transformed into a high-dimensional kernel space, enabling nonlinear modeling and increasing the robustness of tracking.

[0058] After applying the kernel trick, f(z) can be expressed as:

[0059]

[0060] Constructing a regularized least-squares loss function:

[0061]

[0062] Finally, we can obtain the closed-form solution of the template w for the linear expression:

[0063]

[0064]

[0065] S2: Use a Kalman filter to obtain the predicted position of the target in the current frame image, and obtain the NCC response map corresponding to the sub-map formed by the set neighborhood centered on the predicted position;

[0066] The NCC response map corresponding to the sub-image in step S2 is obtained by performing NCC operation on the sub-image using the three NCC templates with the highest WPSR response values ​​in the NCC template queue and filtering the three response maps formed by the channel maximum value. Each template in the NCC template queue is the target region of the five images with the highest WPSR response values ​​in the previous frame.

[0067] The NCC template queue is decayed before being used to obtain the NCC response graph corresponding to the subgraph. The decay method of the NCC template queue is as follows:

[0068]

[0069] Where i is the template number in the NCC template queue, with a value ranging from 1 to 5. This represents the value of the i-th NCC template in the NCC template queue at time T, where T represents the frame number of the current frame image, T-1 represents the frame number of the previous frame image, and β is the set attenuation factor.

[0070] It should be noted that the NCC template queue is updated by attenuation in order to prevent NCC templates with high WPSR response values ​​that are too old from continuously occupying the highest position in the queue. This would cause the target appearance information that is too old to be used for matching in the current occlusion frame. If the target appearance changes during the occlusion phase, the matching success rate will be reduced. Therefore, the memory attenuation mechanism is added.

[0071] Furthermore, the method for obtaining the predicted position of the target in the current frame image using a Kalman filter is as follows:

[0072]

[0073] Where, x t-1|t-1 The state vector at time t-1 includes position and velocity in the x and y directions. t|t-1 F represents the predicted state vector at time t. F is the state transition matrix. P is the covariance matrix, and Q is the system noise covariance matrix.

[0074] S3: Obtain a motion information weighted map based on the motion information corresponding to the nearest uncrowded image to the current frame image, and multiply the motion information weighted map with the NCC response map corresponding to the current frame image to obtain an improved NCC response map;

[0075] In step S3, the motion-weighted image w corresponding to the nearest uncrowded image to the current frame image. x,y The method for obtaining it is as follows:

[0076]

[0077] Where, σ x σ y , respectively, are the standard deviations of the target's position along the x-axis and y-axis. If the target remains occluded, the standard deviations will gradually increase; x and y are the distances of the target from the predicted position along the x-axis and y-axis, respectively; W is the width of the set neighborhood; and H is the length of the set neighborhood.

[0078] S4: Determine whether the maximum response value in the improved NCC response map exceeds the set threshold t2. If yes, it is considered that a similar target has been detected, and proceed to step S5. If no, expand the set neighborhood and reselect the next frame image to execute steps S2 to S4.

[0079] The method for expanding the defined neighborhood in step S4 is as follows:

[0080] Assume that the length of the neighborhood before expansion is k*h and the width is k*w, where h is the set minimum detection unit length, w is the set minimum detection unit width, and k represents the number of times the judgment result in step S4 is that the maximum response value in the improved NCC response map does not exceed the set threshold t2.

[0081] Increment k by 1, then multiply it by the length and width of the smallest detection unit to obtain the expanded set neighborhood.

[0082] When the defined neighborhood is expanded to the size of the entire graph, the defined neighborhood is downsampled by a factor of two, and steps S2 to S4 are executed using the downsampled defined neighborhood.

[0083] S5: Obtain the KCF template corresponding to the nearest uncrowded image to the current frame image, and use the KCF template to perform KCF detection on the first M response values ​​selected from the improved NCC response map that exceed the set threshold t2 in turn, to obtain M KCF response values, and then proceed to step S6, where M is at most 3;

[0084] The method for obtaining the KCF template corresponding to the nearest uncrowded image to the current frame image in step S5 is as follows:

[0085] F t =(1-η-η) max )F t-1 +η max F max +ηF t ′

[0086] Where t is the frame number of the nearest uncrowded image to the current frame, and F t For the optimized KCF template of the t-th frame image that is the nearest uncrowded image to the current frame image, F t-1 F is the optimized KCF template of the frame preceding the nearest t-th frame image that is not occluded to the current frame image. t ' is the original KCF template of the nearest uncrowded frame t to the current frame image, and η is the set baseline template learning rate. max F is the template learning rate corresponding to the image with the highest WPSR response value in the historical frame images; maxAn optimized KCF template for the image with the highest WPSR response value in historical frame images.

[0087] In other words, in order to integrate historical information, the present invention uses an improved template update mechanism that can record the optimal template, thereby further mitigating the impact of mislearning the features of the occluded object when the WPSR value does not decrease significantly when entering the partially occluded state.

[0088] S6: Determine whether there is at least one KCF detection response value greater than the set threshold t3. If yes, it is considered that the position of the target in the current frame image has been detected, and the position of the maximum value among the KCF detection response values ​​greater than the set threshold t3 is taken as the position of the target in the current frame image. If no, the next frame image is selected again and steps S2 to S4 are executed.

[0089] If the judgment result in step S6 is yes, it is considered that the occlusion has ended or the current occlusion is insufficient to affect the target tracking, and the Kalman filter is restarted to obtain the target's motion information and position information. It should be noted that restarting the filter can make the filter update process after restarting unaffected by the previous tracking. Since the target often does not perform linear motion during the occlusion period, restarting the filter can enable the Kalman filter to quickly learn new motion information.

[0090] Specifically, the state update of the Kalman filter can be performed according to the following formula:

[0091]

[0092] Where K is the Kalman gain, z t =[x,y] T This represents the new location detected by KCF. Let H represent the state vector at time t, H be the observation matrix, P be the covariance matrix, Q be the system noise covariance matrix, R be the observation noise matrix, and I be the identity matrix.

[0093] Following the above process, the KCF template, NCC template queue, and Kalman filter can be continuously updated when occlusion does not occur.

[0094] The aforementioned target reconstruction and retrieval mechanism combines the target's motion information with appearance features at two granularities, effectively solving the problem of interference from similar targets and finding a more accurate restart location. The solution to interference from similar targets is mainly due to the relatively accurate image matching capability of NCC. At the same time, the normalization process of NCC can effectively alleviate the interference of illumination changes. With the powerful foreground and background resolution capability of the relevant filter template, the target's reconstruction location can be further precisely located. Compared with the slow speed of ordinary NCC algorithms, this invention uses integral images and FFT technology to improve the running speed of NCC.

[0095] It should be noted that steps S3 to S6 are actually the target re-identification process based on the fast NCC algorithm and Kalman filter to resist interference from similar targets. The NCC algorithm can perform a wide-range search near the location predicted by the Kalman filter, thus alleviating the limitation of the Kalman filter in only performing linear motion modeling. However, using only the NCC response result as the re-identification location is not robust. Therefore, this invention combines NCC template detection and KCF template detection for multi-stage target re-identification. The specific process is as follows:

[0096] Centered on the predicted position of the Kalman filter, a region of size k*w and k*h (pre-set) is selected for NCC template matching, where k represents the nth time the target was not detected. When k = K, k*w and k*h are set to the full image size for full image search. To improve running speed, when the selected region is larger than 200*200 resolution, downsampling is performed by a factor of two. Meanwhile, since the robustness of NCC is poor, especially when the position with a high NCC response value is too far from the predicted position, the reliability of the NCC response value is low. Therefore, this invention also incorporates motion information to weight the NCC response map. The weighting method is to multiply the NCC response map with the motion information weighted map to obtain a new NCC response map.

[0097] Furthermore, in an i7-9700KF 16g Ubuntu 20.04 environment, with a target size of 150*200 pixels, the algorithm can perform normal tracking even with occlusion, achieving an average tracking speed of 60fps, demonstrating high tracking speed while resisting occlusion.

[0098] Therefore, the present invention has the following advantages:

[0099] During the occlusion phase, this invention does not rely solely on the Kalman filter for position prediction. Instead, it combines NCC template detection and KCF template detection for multi-stage target re-identification when predicting the position using the Kalman filter. Furthermore, it proposes a mechanism where the NCC algorithm first provides candidate recurrence regions, and then the KCF template performs a secondary determination. Compared to the traditional KCF method, this invention is more resistant to occlusion, meaning it can predict the target position and perform continuous tracking even during the occlusion phase. In contrast, the traditional KCF method cannot handle occlusion issues; once occlusion occurs, the tracking process will fail regardless of whether the target subsequently escapes the occlusion.

[0100] This invention proposes a motion information weighting mechanism for NCC detection, which has better robustness compared to simply adding the NCC algorithm.

[0101] Compared to traditional KCF and related improved methods, this invention can resist interference from similar targets during the target reconstruction stage. Because this invention combines motion information to design two layers of re-retrieval mechanisms with different granularities, it can resist interference from similar targets. For example, when tracking pedestrians, if other similar pedestrians appear during reconstruction after the tracked pedestrian is occluded, this invention can also track the correct target more accurately.

[0102] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A correlation filtering tracking method that resists occlusion and interference from similar targets, characterized in that, Includes the following steps: S1: Obtain the WPSR response value corresponding to the current frame image and determine whether the WPSR response value is greater than the set threshold t1. If yes, the target is not occluded in the current frame image. Use the KCF algorithm to detect the position of the target in the current frame image. If no, the target is occluded in the current frame image. Proceed to step S2. S2: Use a Kalman filter to obtain the predicted position of the target in the current frame image, and obtain the NCC response map corresponding to the sub-image formed by the set neighborhood centered on the predicted position; wherein, the NCC response map corresponding to the sub-image is obtained by performing NCC operation on the sub-image using the three NCC templates with the highest WPSR response values ​​in the NCC template queue, and then filtering the three response maps formed by the channel maximum value. Each template in the NCC template queue is the target region of the five images with the highest WPSR response values ​​in the previous frame; The NCC template queue is decayed before being used to obtain the NCC response graph corresponding to the subgraph. The decay method of the NCC template queue is as follows: Where i is the template number in the NCC template queue, with a value ranging from 1 to 5. This represents the value of the i-th NCC template in the NCC template queue at time T, where T represents the frame number of the current frame image, and T-1 represents the frame number of the previous frame image. The set attenuation factor; S3: Obtain a motion information weighted map based on the motion information corresponding to the nearest uncrowded image to the current frame image, and multiply the motion information weighted map with the NCC response map corresponding to the current frame image to obtain an improved NCC response map; S4: Determine whether the maximum response value in the improved NCC response map exceeds the set threshold t2. If yes, proceed to step S5; if no, expand the set neighborhood and reselect the next frame image to execute steps S2~S4. The method for expanding the set neighborhood is as follows: Assume the initial neighborhood length is k. h, width is k w, where h is the set minimum detection unit length, w is the set minimum detection unit width, and k represents the number of times in step S4 the judgment result is that the maximum response value in the improved NCC response graph does not exceed the set threshold t2; Each time k increases by 1, it is multiplied by the length and width of the smallest detection unit to obtain the expanded set neighborhood. S5: Obtain the KCF template corresponding to the nearest uncrowded image to the current frame image, and use the KCF template to perform KCF detection sequentially at the first M positions selected from the improved NCC response map that exceed the set threshold t2, to obtain M KCF detection response values, and then proceed to step S6, where M is at most 3; S6: Determine whether there is at least one KCF detection response value greater than the set threshold t3. If yes, it is considered that the position of the target in the current frame image has been detected, and the position of the maximum value among the KCF detection response values ​​greater than the set threshold t3 is taken as the position of the target in the current frame image. If no, the next frame image is selected again and steps S2~S4 are executed.

2. The correlation filtering tracking method for resisting occlusion and interference from similar targets as described in claim 1, characterized in that, The WPSR response value corresponding to the current frame image is calculated as follows: in, This is the maximum value in the KCF response graph corresponding to the current frame image. These represent the mean and standard deviation of the sidelobe region, respectively.

3. The correlation filtering tracking method for resisting occlusion and interference from similar targets as described in claim 1, characterized in that, In step S3, a motion information weighted map is obtained based on the motion information corresponding to the nearest uncrowded image to the current frame image. Specifically: in, , These are the standard deviations of the target's position along the x-axis and y-axis, respectively. If the target remains occluded, the standard deviations will gradually increase. , These represent the distances of the target from the predicted position obtained by the Kalman filter in this frame of the image, along the x-axis and y-axis directions, respectively. The width of the defined neighborhood, The length of the defined neighborhood.

4. The correlation filtering tracking method for resisting occlusion and interference from similar targets as described in claim 1, characterized in that, The method for obtaining the KCF template corresponding to the nearest uncrowded image to the current frame image in step S5 is as follows: Where t is the frame number of the nearest uncrowded image to the current frame image. The optimized KCF template is the nearest uncrowded frame t-th image to the current frame image. The optimized KCF template is the frame preceding the nearest t-th frame image that is not occluded to the current frame image. The original KCF template for the t-th frame image that is the closest uncrowded image to the current frame image. The learning rate is set as the baseline template. The template learning rate is the image with the highest WPSR response value in the historical frame images. An optimized KCF template for the image with the highest WPSR response value in historical frame images.

5. The correlation filtering tracking method for resisting occlusion and interference from similar targets as described in claim 1, characterized in that, When the defined neighborhood is expanded to the size of the entire graph, the defined neighborhood is downsampled by a factor of two, and steps S2 to S4 are executed using the downsampled defined neighborhood.

6. The correlation filtering tracking method for resisting occlusion and interference from similar targets as described in claim 1, characterized in that, If the judgment result in step S6 is yes, the Kalman filter is restarted first, and then the target's motion information and position information are updated.

7. The correlation filtering tracking method for resisting occlusion and interference from similar targets as described in claim 1, characterized in that, If the judgment result in step S1 is negative, the Kalman filter update step is paused.

Images