Real-time high robustness target tracking method for high-noise gray image based on hybrid denoising coupled kernel correlation filter

Abstract

The application discloses a high-noise grayscale image real-time high-robustness target tracking method based on a hybrid denoising coupled kernel correlation filter, which comprises a hybrid denoising module, an adaptive scale module and a redetection module. The hybrid denoising module realizes the balance between edge preservation and noise suppression by calculating the interframe average grayscale value change, adaptively switching bilateral filtering and median filtering and dynamically adjusting the filtering weight; the adaptive scale module realizes the real-time optimization of the target scale based on the response maximization by constructing the scale pyramid and introducing the dynamic weight for each scale factor; the redetection module triggers a three-stage progressive search strategy when the tracking fails by setting the confidence threshold, and restores the target layer by layer from local to global. The application forms a tracking closed loop through the cooperative work of the three modules, effectively solves the problems of easy feature being submerged, scale drift and target loss caused by occlusion in target tracking in a high-noise environment, and realizes the real-time high-robustness tracking of a dynamic target in a high-noise grayscale image.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to a real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupled kernel correlation filtering. It is suitable for real-time tracking of dynamic targets in high-noise grayscale image sequences and can be widely applied in fields such as industrial automation, robot navigation, traffic monitoring, driver assistance, UAV vision, and security monitoring. Background Technology

[0002] Target tracking is an important research direction in computer vision. Its core task is to detect, estimate, and predict information such as the position, velocity, and trajectory of a target by processing consecutive image frames in a video sequence. Kernel correlation filtering (KCR) algorithms have attracted widespread attention due to their simplicity, fast computation speed, and high tracking accuracy. This algorithm uses a cyclic matrix to densely sample the target template and leverages Fast Fourier Transform (FFT) to convert the correlation operation to the frequency domain for efficient computation. Combined with histogram of oriented gradients (HBP) features, it achieves fast target position estimation and exhibits good tracking performance in low-noise or ideal environments.

[0003] However, in practical applications, target tracking faces many challenges, and existing kernel correlation filtering techniques have the following shortcomings: (I) Existing denoising methods struggle to adapt to dynamically changing noise environments, leading to insufficient accuracy in target feature extraction: Actual video sequences are often affected by various noise sources such as sensor noise, quantization errors, and compression artifacts, resulting in a decrease in image signal-to-noise ratio. In high-noise grayscale images, the edges, textures, and other detailed features of the target are easily submerged or even lost due to noise. In existing technologies, denoising preprocessing and target tracking are usually processed as two independent steps, lacking the ability to dynamically perceive and adaptively adjust the image noise level. Traditional fixed denoising methods cannot adaptively switch filtering strategies based on the degree of grayscale fluctuation between frames, making it difficult to achieve a balance between suppressing noise and preserving edge details, resulting in contaminated filter templates and reduced accuracy in target feature extraction and matching. (II) Existing scale estimation methods struggle to adapt to rapid target deformation, with prominent scale drift issues: During target movement, the target's scale in the image continuously changes due to changes in distance and viewpoint. Existing kernel correlation filtering algorithms typically employ fixed scale factors or simple scale pooling strategies for scale estimation, lacking a dynamic perception and weight adjustment mechanism for the scale change trend of historical frames. When the target undergoes rapid deformation, the rate of scale change between adjacent video frames increases, making it difficult for a fixed scale factor to respond effectively. This leads to an accumulation of scale estimation bias, ultimately causing scale drift and severely affecting tracking accuracy. Furthermore, traditional methods lack a weighting mechanism for evaluating samples at each scale; excessively large or small scale factors can negatively impact the tracking process. (III) Lack of an effective re-detection mechanism, leading to permanent target loss in long-term occlusion scenarios: During target tracking, occlusion, rapid target movement, and cluttered backgrounds are the main factors causing tracking failure. Existing kernel correlation filtering algorithms lack an effective mechanism for evaluating the confidence of tracking results. When the target is occluded or tracking fails, the confidence of the response map decreases significantly, and the algorithm lacks an effective re-detection and recovery strategy. Even if some methods design re-detection schemes, they typically employ a single-scale global search approach, resulting in high computational overhead and low search efficiency. This makes it difficult to balance re-detection success rate and real-time requirements, leading to permanent target loss after long-term occlusion and interruption of the tracking sequence. (iv) Lack of dynamic control in model updates, resulting in a prominent contradiction between template drift and scene adaptation: Existing kernel correlation filtering algorithms typically use a fixed learning rate for model updates, making it difficult to balance the weights of historical information and new observation information. In high-noise environments, a fixed learning rate can easily lead to two extreme situations: when the learning rate is too high, single-frame noise pollution is quickly introduced into the model, causing template drift; when the learning rate is too low, the model cannot adapt to changes in the appearance of the target and scene evolution in a timely manner. This contradiction is particularly prominent in long-term tracking tasks, resulting in insufficient adaptive capability of the tracker in complex scenes. (v) Lack of coordination in independent operation of each stage, resulting in insufficient overall robustness: In existing technologies, each stage, such as denoising preprocessing, target localization, scale estimation, and model updates, is usually designed independently and connected in sequence, lacking an effective data transfer and feedback coordination mechanism between modules.The denoising process does not consider subsequent tracking requirements, scale estimation does not utilize the confidence information of target localization, and model updates are unaware of the reliability of tracking results. The independent operation of each stage leads to the accumulation of errors at each stage. Once a deviation occurs in one stage, subsequent stages cannot correct or compensate for it, making it difficult to guarantee the robustness of the overall system in complex environments.

[0004] In summary, existing kernel correlation filtering target tracking methods struggle to effectively suppress noise interference while handling issues such as target scale changes and occlusion in high-noise grayscale image scenarios. There is an urgent need for a real-time, highly robust target tracking method that can achieve adaptive denoising, scale optimization, re-detection and recovery, and module collaboration. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides a real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupled kernel correlation filtering. By working collaboratively with a hybrid denoising module, an adaptive scaling module, and a re-detection module, a tracking closed loop is formed, enabling reliable, accurate, and real-time tracking of dynamic targets in high-noise grayscale images.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupled kernel correlation filtering is proposed. The method includes a hybrid denoising module, an adaptive scaling module, and a re-detection module. These modules work collaboratively to form a tracking closed loop. The specific steps are as follows: Step 1: Acquire the current frame image and initialize the target position. Input the current frame image as the algorithm's input parameter, and simultaneously acquire the target's position coordinates and initial region of interest (ROI) from the previous frame. Input the previous frame image and the initial ROI into the algorithm to drive it to update the target position.

[0007] Step 2: Adaptive denoising preprocessing is performed on the input frame image using a hybrid denoising module. The hybrid denoising module is used to reduce noise contamination of the filter template and improve the accuracy of extraction and matching of effective target features.

[0008] Step 3: The denoised image is scaled using an adaptive scaling module. This module adapts to deformations caused by changes in distance and viewpoint during target motion, improving the algorithm's robustness to rapid target deformation.

[0009] Step 4: Address tracking failures and occlusions using the re-detection module. The re-detection module is used to re-search for the target in a local area when target tracking fails or the response map confidence level falls below a threshold, preventing target loss due to occlusion.

[0010] Step 5: Ensure target tracking continuity through visualization output and model updates. Draw the target tracking result bounding box on the original image based on the target's location and scale, and label the tracking status to present the target location and tracking status visually for easy result analysis. Simultaneously, update the filter coefficients and base samples using the tracking results of the current frame, allowing the model to adapt to changes in the target's appearance. After the update is complete, return to Step 1 to continue processing the next frame, forming a tracking loop.

[0011] Furthermore, step one specifically involves: The image of the current frame is used as the input parameter of the algorithm. At the same time, the position coordinates of the target in the previous frame and the initial region of interest are obtained. The image of the previous frame and the initial region of interest are used as input parameters to drive the algorithm to update the target position.

[0012] Furthermore, step two specifically involves: (1) Calculate the inter-frame grayscale change to evaluate the image noise level. Calculate the average grayscale value change of the current frame image relative to the previous frame image to characterize the grayscale fluctuation and noise level of the current frame image, and provide a basis for the adaptive selection of subsequent denoising algorithms.

[0013] (2) Adaptively select denoising algorithms based on inter-frame grayscale changes. Compare the average grayscale value change of the current frame image with the set grayscale change threshold, adaptively select different denoising algorithms, and dynamically adjust the filtering weights: If the average grayscale value change of the current frame image exceeds the set grayscale change threshold, it indicates that the grayscale change of the current frame image is drastic, and the detailed features such as target edges and textures in the image may be submerged by noise. In this case, bilateral filtering is used for noise removal. Bilateral filtering can smooth noise while preserving edge structure and avoiding blurring of target contours. The output point of bilateral filtering is calculated based on the pixel values ​​of the input point and its neighborhood, combined with spatial location weights and pixel value weights. The spatial location weight is determined by the spatial distance between the pixel and the center point, and the pixel value weight is determined by the difference in grayscale value between the pixel and the center point. The product of the two is used as the comprehensive weight of the input point in the filtering calculation.

[0014] If the average grayscale value change of the current frame image is less than or equal to the set grayscale change threshold, it indicates that the grayscale change of the current frame image is gradual, and the image mainly contains salt-and-pepper impulse noise. In this case, median filtering is used for noise reduction. Median filtering can effectively remove isolated noise points, has high computational efficiency, and its output is the median of all pixel values ​​sorted within the filter window coverage area.

[0015] Both bilateral filtering and median filtering denoising methods incorporate an adaptive filtering weight adjustment mechanism based on inter-frame grayscale changes. This mechanism dynamically adjusts the filtering weights according to the difference between the current frame's grayscale change value and a set threshold, enabling the adjusted filtering function to automatically adjust with scene changes and noise level fluctuations, thus ensuring the stability of the denoising effect.

[0016] After the above adaptive denoising process, a denoised image with enhanced edge-preserving properties is output, providing high-quality image input for subsequent target feature extraction and inter-frame matching.

[0017] Furthermore, step three specifically involves: (1) Extract histogram of oriented gradient features from the initial region of interest image after denoising. Histogram of oriented gradient features describe the shape and edge information of the target by calculating the gradient orientation histogram of the local region of the image, and have a certain robustness to gray-level changes and local geometric deformations.

[0018] (2) The kernel correlation filter is applied to initialize the tracker and establish the initial state of the target template and the filter. The kernel correlation filter performs dense sampling of the target template through the cyclic matrix and uses the fast Fourier transform to convert the correlation operation to the frequency domain for efficient calculation, thereby establishing the appearance model of the target and ensuring the reliability of the tracking response.

[0019] (3) Multi-scale sampling is performed on the feature-extracted images to construct a scale pyramid. With the current target scale as the center, image patches of different scales are generated according to multiple scale factors to form a scale pyramid, which is used to represent the appearance changes of the target at different scales. During the target tracking process, the scale change between adjacent video frames is generally not too large. Therefore, in order to reduce the impact of excessively large or small scales on the tracking process, a weight is introduced for each scale factor to balance different scales. The weight value is dynamically adjusted according to the scale change rate of historical frames. The larger the change rate, the stronger the penalty on the weight of the larger scale factor, and vice versa.

[0020] (4) The response of each scale sample is weighted and evaluated by the scale correlation filter, the optimal scale is calculated, and the scale with the largest response is selected as the new scale of the target in the current frame, thereby realizing the real-time optimization of the target scale and effectively suppressing scale drift.

[0021] (5) The target features are processed in the frequency domain by kernel correlation filtering. The target template and the search area are cross-correlated to generate a response map. Each pixel value in the response map represents the confidence that the location is the center of the target. The coordinates corresponding to the maximum peak are taken as the estimated position of the target. The candidate scale corresponding to the maximum peak of the response map in the current frame is weighted and fused with the target scale in the previous frame according to the scale smoothing factor to determine the final scale of the target in the current frame. Finally, the center localization and scale estimation of the target in the current frame are completed, and robust target tracking results are output.

[0022] Furthermore, step four specifically involves: (1) Set the re-detection threshold and obtain the maximum response value of the response map to determine the reliability of the current tracking result.

[0023] (2) When the maximum response value is greater than the re-detection threshold, it indicates that the tracking is successful and the target position is reliable. The system will directly jump to the output step to avoid unnecessary re-detection calculations and ensure tracking efficiency and real-time performance.

[0024] (3) When the maximum response value is less than or equal to the re-detection threshold, it indicates that the tracking result confidence is too low, leading to tracking failure. At this time, the re-detection process is initiated, and the re-detection module is activated to prevent the target from being permanently lost. The re-detection process proceeds in three levels: The first level is a local scale search: A dense search is performed within a small area centered on the current target location to extract features from candidate regions and calculate the response map, obtaining the response peak and candidate locations. If the response peak is greater than the re-detection threshold, the re-detection is considered successful, the target location is updated, and the loss counter is reset.

[0025] The second level is the expanded scale search: if the local scale search fails, the search range is expanded and candidate samples are generated according to multiple scale factors. The response map of each candidate region is calculated, and the candidate region with the highest response peak is selected as the re-detection result.

[0026] The third level is the region expansion search: if the first two levels fail, the search range is further expanded to the local area of ​​the entire frame image, and a sliding window search is performed. Features are extracted for each candidate window and the response value is calculated. The candidate box with the highest response value is selected as the new target location.

[0027] (4) In the candidate sample set, search for the target and extract the features of the candidate regions at different candidate search scales, calculate the response map of each candidate region, evaluate the confidence of each candidate region, and select the region most likely to contain the target. Select the candidate box with the highest confidence from the search results as the new target location, and use the learning rate to perform weighted fusion of the current target template and the candidate target template to update the target template. After successful re-detection, output the target location and scale, and pass the recovered tracking result to the next frame to ensure the continuity of the tracking sequence.

[0028] Furthermore, step five specifically involves: Draw the target tracking result box on the original image based on the target location and scale, and mark the tracking status to present the target location and tracking status in a visual form, which facilitates result analysis.

[0029] Simultaneously, the filter coefficients and base samples are updated using the tracking results of the current frame, allowing the model to adapt to changes in the target's appearance. The filter coefficients are weighted and fused with the candidate filter coefficients of the current frame and the filter coefficients of the previous frame using the model learning rate. Similarly, the base samples are weighted and fused with the candidate base samples of the current frame and the base samples of the previous frame using the model learning rate, thus completing the update. The model learning rate controls the step size of the model update, balancing the weights of historical information and new observations.

[0030] After completing the above update, return to step one to continue processing the next frame image, forming a tracking loop to achieve continuous tracking of the target and ensure the continuity of target tracking.

[0031] Compared with the prior art, this patent application has the following advantages: 1. Dynamically adapts to industrial target shapes, significantly improving feature extraction accuracy. This invention assesses image noise levels by calculating the average grayscale value change of the current frame relative to the previous frame, and adaptively selects a denoising algorithm based on the comparison between the grayscale change value and a set threshold. When grayscale changes drastically, details such as target edges and textures in the image may be obscured by noise. In this case, bilateral filtering is used for denoising. Since bilateral filtering considers both spatial location weights and pixel value weights when calculating the output point, it can smooth noise while preserving edge structure and avoiding blurring of target contours. When grayscale changes are gradual, the image mainly contains salt-and-pepper impulse noise. In this case, median filtering is used for denoising. Since its output point is the median of all pixel values ​​sorted within the filter window coverage area, it can effectively remove isolated noise points and has high computational efficiency. In addition, both denoising methods introduce an adaptive filter weight adjustment mechanism based on inter-frame grayscale changes. That is, the filter weights are dynamically adjusted according to the difference between the grayscale change value and the set threshold, so that the filter function can automatically adjust with scene changes and fluctuations in noise levels. After the above adaptive denoising process, the output denoised image has enhanced edge preservation properties, providing high-quality image input for subsequent target feature extraction and inter-frame matching, thereby effectively solving the problem that target detail features are easily submerged by noise in high-noise environments, leading to tracking failure.

[0032] 2. Scale-weighted dynamic balancing strategy, resulting in a leap in deformation adaptability. This invention describes the shape and edge information of a target using histogram of oriented gradients (HARQs), a feature robust to grayscale changes and local geometric deformations. Based on this, a target appearance model is established using kernel correlation filtering, and fast Fourier transform is used to convert correlation operations to the frequency domain for efficient computation, ensuring the reliability of the tracking response. Furthermore, a scale pyramid is constructed, generating image patches of different scales centered on the current target scale and based on multiple scale factors, representing the target's appearance changes at different scales. Dynamic weights are introduced for each scale factor, with the weight values ​​adjusted according to the scale change rate of historical frames. A higher change rate results in a stronger penalty for the weights of larger scale factors, and vice versa, thereby reducing the negative impact of excessively large or small scales on the tracking process. A scale correlation filter is used to weight and evaluate the responses of samples at each scale, selecting the scale with the largest response to achieve real-time optimization of the target scale and effectively suppress scale drift. Finally, in the response map generated by kernel correlation filtering, each pixel value represents the confidence that the location is the center of the target. The coordinates corresponding to the maximum peak value are taken as the target position. The candidate scale of the current frame and the target scale of the previous frame are weighted and fused together with the scale smoothing factor to determine the final scale. This enables the algorithm to adapt to the deformation of the target caused by changes in distance and viewpoint during the movement of the target and maintain stable tracking accuracy in scenarios with rapid target deformation.

[0033] 3. A three-level progressive re-detection mechanism to cope with prolonged occlusion and target loss. This invention determines the reliability of the current tracking result by setting a re-detection threshold and obtaining the maximum response value of the response map. When the maximum response value is greater than the re-detection threshold, it indicates successful tracking, and the result is output directly, avoiding unnecessary re-detection calculations and ensuring tracking efficiency and real-time performance. When the maximum response value is less than or equal to the re-detection threshold, it indicates tracking failure, and the re-detection process is triggered. The re-detection adopts a three-level sequential strategy: the first level performs a dense search within a small range, and if successful, the target is quickly recovered; the second level expands the search range and generates candidate samples according to multiple scale factors, and if successful, the target is recovered within a medium range; the third level expands the search range to a local area of ​​the entire frame image for a sliding window search, and finds the target within a large range. This progressive search strategy, from small to large and layer by layer, achieves a balance between computational efficiency and re-detection success rate. In the candidate sample set, the confidence is evaluated by calculating the response map of each candidate region, the region most likely to contain the target is selected, the candidate box with the highest confidence is selected as the new target location, and the current target template and the candidate target template are weighted and fused using a learning rate to complete the template update. After successful re-detection, the recovered target position and scale are passed to the next frame, effectively preventing permanent loss of the target due to long-term occlusion, rapid target movement, etc., and ensuring the continuity of the tracking sequence.

[0034] 4. The model parameters are dynamically and smoothly updated to ensure long-term drift suppression and scene adaptation. This invention updates the filter coefficients and base samples after each frame of tracking is completed, using the tracking results of the current frame. This allows the model to adapt to changes in the target's appearance over time. The filter coefficients are weighted and fused with the candidate filter coefficients of the current frame and the filter coefficients of the previous frame using the model learning rate. The base samples are updated similarly. The model learning rate controls the step size of the model update, balancing the weights of historical information and new observations. This allows the tracker to dynamically adjust the appearance model to adapt to target deformation and occlusion in high-noise environments, while preventing tracking failures caused by template drift. After the update is complete, the process returns to the first step to continue processing the next frame, forming a tracking loop that ensures the continuity of target tracking and exhibits stable performance in long-term tracking tasks.

[0035] 5. A closed-loop collaborative approach of denoising, tracking, and re-detection enables real-time, highly robust tracking of high-noise grayscale images. This invention forms a complete tracking closed loop through the collaborative work of a hybrid denoising module, an adaptive scaling module, and a re-detection module: the hybrid denoising module performs adaptive denoising preprocessing on the input frame, providing high-quality image input for subsequent modules; the adaptive scaling module completes the accurate localization and scale estimation of the target; the re-detection module recovers the target when tracking fails; and the model update mechanism enables the tracker to adapt to target changes. The collaborative data transfer and feedback between these modules achieves real-time, highly robust tracking of dynamic targets in noisy grayscale images. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the structure of the real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupled kernel correlation filtering used in embodiments of the present invention. Figure 2 This is a flowchart illustrating the structure of the hybrid noise reduction module used in this embodiment of the invention. Figure 3 This is a flowchart illustrating the adaptive scaling module structure used in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the re-detection module structure used in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] A real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupled kernel correlation filtering is proposed. This method is applied to the real-time tracking of dynamic targets in high-noise grayscale image sequences, such as... Figure 1 As shown, the method includes a hybrid denoising module, an adaptive scaling module, and a re-detection module, with each module working together to form a tracking closed loop.

[0039] In this embodiment, the input image sequence is a grayscale video frame with a frame rate of 30fps and an image resolution of 640×480 pixels. The initial position of the target and the initial region of interest are manually marked by the user in the first frame or automatically provided by the target detection algorithm.

[0040] In this embodiment, the hybrid denoising module is used to perform adaptive denoising preprocessing on the input frame image to reduce noise contamination of the filter template. For example... Figure 2 As shown, the specific working process of this module is as follows: Step 1: Acquire the current frame image and initialize the target position. Input the image of the current frame t as the input parameter of the algorithm, and at the same time obtain the position coordinates of the target in frame t-1 and the initial region of interest (ROI). Input the image of frame t-1 and the initial ROI as input parameters into the algorithm to drive the algorithm to complete the update of the target position.

[0041] Step 2: Calculate inter-frame grayscale changes to assess image noise levels. Calculate the average grayscale value change of the current frame t relative to the previous frame t-1. It is used to characterize the grayscale fluctuation and noise level of the current frame image, providing a basis for the adaptive selection of subsequent denoising algorithms.

[0042] Step 3: Adaptively select a denoising algorithm based on inter-frame grayscale changes. This involves adjusting the average grayscale value of the current frame image. The filter weights w are compared with a set grayscale change threshold, and different denoising algorithms are adaptively selected, while the filter weights w are dynamically adjusted.

[0043] like > This indicates that the grayscale of the current frame image changes drastically, and details such as target edges and textures may be obscured by noise. Therefore, bilateral filtering is used for noise reduction. Bilateral filtering smooths noise while preserving edge structure and avoiding blurring of target contours. Its output points... g ( i , j Based on the input point f ( k , l )calculate:

[0044] In the formula, S ( i ,j) means with ( i The neighborhood of size (2N+1)x(2N+1) centered at (j). f ( k , l () represents the pixel value of the input point. w ( i , j , k , l ) represents the weight value calculated by multiplying the spatial distance Gaussian function and the gray value Gaussian function.

[0045] like ≤ θ This indicates that the grayscale change in the current frame is gradual, and the image mainly contains salt-and-pepper impulse noise. Therefore, median filtering is used for noise removal. Median filtering can effectively remove isolated noise points, has high computational efficiency, and its output points... g ( x , y Based on the input point f ( x , y )calculate:

[0046] In the formula, f ( x , y ) represents the pixel values ​​of the original input image. g ( x , y ) represents the processed output pixel value, and W represents the filter kernel size (must be an odd number). m , n This represents the offset within the filter window. After the above adaptive denoising process, the output is a denoised image with enhanced edge-preserving properties, providing high-quality image input for subsequent target feature extraction and inter-frame matching.

[0047] In this embodiment, both bilateral filtering and median filtering denoising methods incorporate an adaptive filtering weight adjustment mechanism based on inter-frame grayscale changes, i.e., according to... and θ The difference dynamically adjusts the filter weight w, enabling the adjusted filter function to automatically adjust with scene changes and noise level fluctuations, ensuring the stability of the denoising effect.

[0048] Step 4: Scale the denoised image using the adaptive scaling module. For example... Figure 3As shown, this module is used to adapt to the deformation of the target caused by changes in distance and viewing angle during movement, improving the robustness of the algorithm to rapid target deformation. The specific working process is as follows: First, histogram of oriented gradients (HOR) features are extracted from the initial region of interest (ROI) image after denoising. HOR features describe the shape and edge information of the target by calculating the gradient orientation histogram of local regions of the image, exhibiting robustness to grayscale changes and local geometric deformations.

[0049] Then, kernel correlation filtering is applied to initialize the tracker, establishing the initial state of the target template and the filter. Kernel correlation filtering performs dense sampling of the target template through a cyclic matrix and uses fast Fourier transform to convert the correlation operation to the frequency domain for efficient calculation, thereby establishing the appearance model of the target and ensuring the reliability of the tracking response.

[0050] Next, multi-scale sampling is performed on the feature-extracted image to construct a scale pyramid. Centered on the current target scale, the pyramid is constructed according to scale factors. S ={ S i | i Five image patches of different scales are generated, forming a scale pyramid, which is used to represent the appearance changes of the target at different scales. It should be noted that during target tracking, the scale change between adjacent video frames is generally not too large. Therefore, to reduce the impact of excessively large or small scales on the tracking process, for each scale factor... S i A weight W is introduced i To balance different scales, the weight values ​​are dynamically adjusted based on the scale change rate of historical frames. The greater the change rate, the stronger the penalty for the weight of larger scale factors, and vice versa.

[0051] Then, the response of each scale sample is weighted and evaluated by a scale-correlated filter to calculate the optimal scale. The scale with the largest response is selected as the new scale of the target in the current frame, thereby achieving real-time optimization of the target scale and effectively suppressing scale drift.

[0052] Finally, the target features are processed in the frequency domain using kernel correlation filtering, and a cross-correlation operation is performed between the target template and the search region to generate a response map R. Each pixel value in the response map represents the confidence level that the location is the center of the target; the coordinates corresponding to the maximum peak value are taken. x , y The estimated position of the target is used as the target's location, and the final scale S of the target in the current frame is determined based on the scale smoothing factor λ. t+1 :

[0053] In the formula, λ is the scale smoothing factor. In this embodiment, the scale smoothing factor λ = 0.75, which effectively suppresses scale abrupt changes while ensuring scale response sensitivity. Finally, the target's center localization and scale estimation in the current frame are completed, and robust target tracking results are output.

[0054] Step 5: Address tracking failures and occlusions using the re-detection module. For example... Figure 4 As shown, this module is used to re-search for the target in a local area when target tracking fails or the confidence level of the response map is below a threshold, in order to prevent target loss due to occlusion. The specific working process is as follows: First, set the re-detection threshold. And obtain the maximum response value P = max(R) of the response map R, which is used to determine the reliability of the current tracking result.

[0055] When the maximum response value P > the re-detection threshold When the target location is found to be credible, the system jumps directly to the output step, thus avoiding unnecessary re-detection calculations and ensuring tracking efficiency and real-time performance.

[0056] When the maximum response value P ≤ the re-detection threshold If the tracking result shows a low confidence level, it indicates that tracking has failed. In this case, the re-detection process begins, activating the re-detection module to prevent permanent target loss. The re-detection process proceeds in three stages: The first level is a local scale search: A dense search is performed within a small area centered on the current target location (the search radius is 1.2 times the current target scale) to extract features of candidate regions and calculate the response map R, obtaining the response peak P. s and candidate positions ( x t , y t If P s > If the re-detection is successful, the target location is updated and the loss counter is reset.

[0057] The second level is the expanded scale search: if the local scale search fails, the search scope is expanded, and a set of local scales is defined. S ={ S 1, S 2, S 3} Calculate the candidate search scale S t+1 =S t ×S i Candidate samples are generated according to multiple scale factors, and the response map of each candidate region is calculated. The candidate region with the highest response peak is selected as the re-detection result.

[0058] The third level is the region expansion search: if the first two levels fail, the search range is further expanded to 1 / 4 of the entire frame image, and a sliding window search is performed with a step size of 0.5 times the target scale. HOG features are extracted for each candidate window and the response value is calculated. The candidate box with the highest response value is selected as the new target location.

[0059] In the candidate sample set, the target is searched and features of candidate regions are extracted at different candidate search scales. The response map R of each candidate region is calculated, and the confidence of each candidate region is evaluated to select the region most likely to contain the target. The candidate box with the highest confidence is selected as the new target location from the search results. x t , y t ), and update its target template α t+1 :

[0060] In the formula, β is the learning rate, and α is the learning rate. t+1 For the current target template, α t This serves as a candidate target template. After successful re-detection, the target location ( x t , y t The system outputs the recovered tracking results to the next frame, along with the scale S, to ensure the continuity of the tracking sequence.

[0061] Step Six: Ensure target tracking continuity through visualization output and model updates. Based on the target location ( x t , y t The target tracking result box is drawn on the original image using scale S, which presents the target position and tracking status in a visual form, facilitating result analysis.

[0062] Simultaneously, the filter coefficients and base samples are updated using the tracking results of the current frame, allowing the model to adapt to changes in the target's appearance. Filter coefficients and base samples x t Update as follows:

[0063] In the formula, η is the model learning rate, which is used to control the step size of model updates and balance the weights of historical information and new observations. After the above update is completed, return to step one to continue processing the next frame image, forming a tracking loop to achieve continuous tracking of the target and ensure the continuity of target tracking.

[0064] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupled kernel correlation filtering, characterized in that, The method includes a hybrid denoising module, an adaptive scaling module, and a re-detection module. These modules work together to form a tracking closed loop. The method includes the following steps: Step 1: Acquire the current frame image and initialize the target position: Input the current frame image as the input parameter of the algorithm, and at the same time obtain the target's position coordinates and initial region of interest in the previous frame. Input the previous frame image and the initial region of interest as input parameters into the algorithm to drive the algorithm to update the target position. Step 2: Adaptive denoising preprocessing is performed on the input frame image through a hybrid denoising module to reduce noise contamination of the filter template and improve the accuracy of extraction and matching of effective target features. Step 3: The scale of the denoised image is adjusted by the adaptive scaling module to adapt to the deformation of the target caused by changes in distance and viewpoint during the movement, thereby improving the robustness of the algorithm to rapid deformation of the target. Step 4: Use the re-detection module to deal with tracking failures and occlusions. When target tracking fails or the confidence of the response map is lower than the threshold, re-search for the target in the local area to prevent target loss due to occlusion. Step 5: Ensure the continuity of target tracking through visualization output and model updates. Draw the target tracking result box on the original image according to the target position and scale and mark the tracking status. At the same time, use the tracking results of the current frame to update the filter coefficients and base samples, so that the model can adapt to the appearance changes of the target. After the update is completed, return to step 1 to continue processing the next frame image, forming a tracking loop.

2. The real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 1, characterized in that, Step two specifically includes: Calculate the average grayscale value change of the current frame image relative to the previous frame image to characterize the grayscale fluctuation and noise level of the current frame image; Compare the average grayscale value change of the current frame image with the set grayscale change threshold: If the average gray value change of the current frame image is greater than the set gray value change threshold, bilateral filtering is used for noise reduction. Bilateral filtering can maintain the edge structure while smoothing noise. Its output point is calculated based on the pixel values ​​of the input point and its neighborhood, combined with spatial position weights and pixel value weights. If the average grayscale value change of the current frame image is less than or equal to the set grayscale change threshold, median filtering is used for noise reduction, and its output point is the median of all pixel values ​​sorted within the filter window coverage area. Meanwhile, the filtering weights are dynamically adjusted based on the difference between the current frame's grayscale change value and the set threshold, so that the filtering function can automatically adjust with scene changes and noise level fluctuations.

3. The real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 1, characterized in that, Step three specifically includes: The initial region of interest image after denoising is subjected to histogram of oriented gradient features. The shape and edge information of the target are described by calculating the gradient orientation histogram of the local region of the image. The tracker is initialized using kernel correlation filtering to establish the initial state of the target template and the filter. Kernel correlation filtering performs dense sampling of the target template through a cyclic matrix and uses fast Fourier transform to convert the correlation operation to the frequency domain for efficient calculation.

4. The real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 3, characterized in that, In step three, image patches of different scales are generated based on the current target scale as the center, according to multiple scale factors, to form a scale pyramid, which is used to characterize the appearance changes of the target at different scales. A weight is introduced for each scale factor to balance different scales. The weight value is dynamically adjusted according to the scale change rate of historical frames. The larger the change rate, the stronger the penalty on the weight of the larger scale factor, and vice versa.

5. The real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 4, characterized in that, In step three, the responses of samples at each scale are weighted and evaluated using a scale-correlation filter, and the scale with the largest response is selected as the new scale of the target in the current frame, thereby achieving real-time optimization of the target scale and effectively suppressing scale drift.

6. The real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 5, characterized in that, In step three, the target features are processed in the frequency domain by kernel correlation filtering, and the target template and the search region are cross-correlated to generate a response map. Each pixel value in the response map represents the confidence that the location is the center of the target. The coordinates corresponding to the maximum peak value are taken as the estimated location of the target. Based on the scale smoothing factor, the candidate scale corresponding to the maximum peak value in the current frame response map is weighted and fused with the target scale in the previous frame to determine the final scale of the target in the current frame.

7. The real-time robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 1, characterized in that, In step four, a re-detection threshold is set, and the maximum response value of the response map is obtained to determine the reliability of the current tracking result. When the maximum response value is greater than the re-detection threshold, it indicates that the tracking is successful and the target location is reliable, and the process jumps directly to the output step; when the maximum response value is less than or equal to the re-detection threshold, it indicates that the tracking has failed and the re-detection process begins.

8. The real-time, robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 7, characterized in that, The re-detection process is carried out in three levels sequentially: The first level is local scale search: with the current target location as the center, a dense search is performed in a small range to extract the features of the candidate region and calculate the response map to obtain the response peak and candidate location; if the response peak is greater than the re-detection threshold, the re-detection is considered successful, the target location is updated and the loss counter is reset. The second level is the expanded scale search: if the local scale search fails, the search range is expanded and candidate samples are generated according to multiple scale factors. The response map of each candidate region is calculated, and the candidate region with the highest response peak is selected as the re-detection result. The third level is the region expansion search: if the first two levels fail, the search range is further expanded to the local area of ​​the entire frame image, and a sliding window search is performed. Features are extracted for each candidate window and the response value is calculated. The candidate box with the highest response value is selected as the new target location.

9. The real-time robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 8, characterized in that, In the candidate sample set, the target is searched and the features of the candidate regions are extracted at different candidate search scales. The response map of each candidate region is calculated, the confidence of each candidate region is evaluated, and the region most likely to contain the target is selected. The candidate box with the highest confidence is selected from the search results as the new target location, and the current target template and the candidate target template are weighted and fused using the learning rate to update the target template; After successful re-detection, the target position and scale are output, and the recovered tracking results are passed to the next frame.

10. The real-time robust target tracking method for high-noise grayscale images based on hybrid denoising coupling kernel correlation filtering according to claim 1, characterized in that, In step five, the filter coefficients are weighted and fused with the candidate filter coefficients of the current frame and the filter coefficients of the previous frame through the model learning rate. The base samples are weighted and fused with the candidate base samples of the current frame and the base samples of the previous frame through the model learning rate. The model learning rate is used to control the step size of the model update and balance the weights of historical information and new observations.

Images