Target tracking method combining saliency perception and adaptive distortion suppression

Abstract

The invention relates to the technical field of computer vision, in particular to a target tracking method combining significance perception and adaptive distortion suppression. Firstly, the expression ability of an image feature enhancement algorithm model to a target is extracted. Then, a self-adaptive spectrum residual weighting strategy is utilized, frequency components of a salient region are highlighted through weighted adjustment of frequency domain information to obtain a saliency sensing result, the saliency sensing result is integrated into a space weight, a tracker can adapt to appearance changes, interference from the surrounding environment is relieved, and when a target deforms, the target can be tracked, so that the tracking accuracy of the target is improved. A tracker can punish a filtering coefficient in a targeted manner and reduce the influence of background noise, then adaptive spatial regularization is introduced, so that the spatial regularization weight can be adjusted according to the dynamic change of a target, the diversity information of the target is fully utilized, and a distortion suppression item is added to constrain a response diagram of a current frame, so that the adaptive spatial regularization is realized. And the discriminating capability of the filter is further improved, so that the problem of filter degradation is effectively relieved, and the target tracking accuracy is improved.

Description

Technical Field

[0001] The present invention relates to a target tracking method combining saliency perception and adaptive distortion suppression, and belongs to the field of computer vision. Background Art

[0002] Object tracking technology has a wide range of applications in computer vision, particularly in autonomous driving, video surveillance, and augmented reality. Ensuring real-time and accurate object tracking is a key challenge. Existing object tracking algorithms typically rely on image appearance models, detecting and predicting the target's position. However, traditional methods have limitations in dynamic scenes, particularly in complex backgrounds, object deformation, and rapid motion, where the target's appearance characteristics can easily change, leading to tracking failure. To address these issues, many researchers have proposed object tracking methods based on frequency domain analysis. Spectral residual methods, as an effective frequency domain processing technique, have been widely used in image processing and object detection. These methods operate on the frequency domain components of an image to enhance meaningful portions (such as the target region) and suppress background noise. However, traditional spectral residual methods typically use a fixed weighting strategy for all frequency components and lack adaptive adjustment to local image features, resulting in suboptimal performance in complex scenes. Furthermore, spatial regularization techniques have been proposed for object tracking to address the dynamic adaptation of appearance changes during rapid motion or deformation. Through spatial regularization, the target area can be locally adjusted, so that the tracking model can maintain good stability under different forms. Despite this, the existing spatial regularization method still faces the problem of how to dynamically adapt to the target motion, deformation and noise influence, resulting in its insufficient robustness in complex environments. With the development of target tracking technology, algorithms that combine multiple information sources and multi-level features have gradually become mainstream. Saliency perception technology is one of the important directions. By analyzing the salient areas in the image, the saliency perception method can effectively separate the foreground target and the background, improve the sensitivity of the tracking algorithm to the target, especially in the case of complex or dynamically changing backgrounds, and can better identify and track the target. However, the existing saliency perception method still faces the problem of difficulty in handling rapidly changing targets and background interference in complex scenes. The present invention aims to combine adaptive spectral residual weighting, adaptive regularization, distortion suppression and saliency perception technology to propose a new target tracking method to improve the robustness and accuracy of target tracking. Summary of the Invention

[0003] The present invention proposes a target tracking method that combines saliency perception with adaptive distortion suppression, which improves the tracking accuracy and robustness of targets in complex backgrounds, fast motion and deformation scenes by dynamically adjusting frequency domain information and spatial models.

[0004] To achieve the above objectives, the present invention provides a target tracking method that combines saliency perception with adaptive distortion suppression. The method comprises the following steps:

[0005] Step 1: First, obtain the input image for target tracking and convert it into a grayscale image. Then, extract the saliency information of the target by preprocessing the image.

[0006] Step 2: Calculate the gradient of the image to obtain the edge information of the image. Use the mean filter to calculate the local variance of the image to reflect the complexity of the texture. The local features are used to construct an adaptive weight matrix to enhance the contribution of important areas in the image.

[0007] Step 3: Based on local features, an adaptive weighting matrix is ​​constructed to dynamically adjust the weights of different frequency components according to the local structure of the image. The weighting matrix is ​​applied to the frequency components in the frequency domain to perform weighted adjustment on the frequency amplitude of the image, thereby highlighting the frequency domain contribution of the salient target area.

[0008] Step 4: Combined with the adaptive spectral residual weighted saliency detection method, the saliency map of the image is calculated in the frequency domain;

[0009] Step 5: Dynamically constrain the spatial models of the target area and background area through an adaptive regularization strategy;

[0010] Step 6: Apply distortion suppression regularization to suppress the distortion effect in the image and ensure that the shape change of the target does not affect the tracking accuracy;

[0011] Step 7: Based on the weighted frequency domain information, the frequency domain image is converted back to the spatial domain through inverse Fourier transform to obtain the current frame image of the target, and the position of the next frame is predicted according to the target's motion law.

[0012] Optionally, during the execution of step 1, a background-aware feature extraction method is first adopted. By constructing a cyclic shift sample matrix containing the target area and its surrounding background area, and utilizing multi-channel feature representation of directional gradient histogram features and color features, the distinction between the target and the background is enhanced and more discriminative target features are extracted.

[0013] Optionally, during step 2, the image gradient is calculated using the Sobel operator to obtain edge information. Simultaneously, a mean filter is used to calculate the local variance of the image, reflecting the complexity of the texture. These local features are used to construct an adaptive weight matrix to enhance the contribution of important regions in the image.

[0014] Optionally, during step 3, an adaptive spectral residual weighted saliency detection method is proposed. First, the image is converted from the spatial domain to the frequency domain using a Fourier transform to obtain the image's frequency components. Next, local image features, such as gradients and local variances, are calculated. These features help capture edge information and texture complexity in the image, reflecting the importance of different regions in saliency detection. Based on these local features, an adaptive weighting matrix is ​​constructed that dynamically adjusts the weights of different frequency components based on the image's local structure. Frequency components in regions containing salient targets are given higher weights, while those in background regions are given relatively lower weights. Finally, this weighting matrix is ​​applied to the frequency components in the frequency domain to weight the image's frequency amplitudes, thereby highlighting the frequency domain contribution of salient target regions, enhancing target representation, and suppressing interference from background noise.

[0015] Optionally, during step 4, an adaptive spectral residual weighted saliency detection method is used to calculate the image's saliency map in the frequency domain and combine it with the target's frequency domain information. Saliency perception techniques can help the algorithm focus more on the salient parts of the target, thereby improving robustness in dynamic scenes.

[0016] Optionally, during the execution of step 5, the spatial models of the target area and the background area are dynamically constrained through an adaptive regularization strategy to improve the stability of target tracking. In the specific implementation, a spatial weight matrix is ​​first generated based on the target detection confidence map, where the pixels in the target area are given a lower regularization coefficient to maintain tracking flexibility, and the pixels in the background area are given a higher regularization coefficient to suppress interference. The spatial model of the target area will be subject to smaller regularization constraints, so as to better adapt to the deformation and movement of the target, while the background area will be subject to stronger regularization constraints to reduce the interference of background noise on the tracking process and enhance the ability to accurately locate and track the target area.

[0017] Optionally, during the execution of step 6, an anomaly suppression regularization term is constructed by calculating the Euclidean distance of the response graphs of adjacent frames. When response distortion is detected, a quadratic constraint is dynamically applied, and its regularization strength is adaptively adjusted according to the current background clutter and occlusion level.

[0018] Optionally, during the execution of step 7, an inverse Fourier transform is performed on the weighted frequency domain features to obtain a spatial domain response, and the target position is determined by detecting the peak position of the spatial domain response graph to achieve target tracking.

[0019] The present invention provides a target tracking method that combines saliency perception with adaptive distortion suppression. First, an adaptive spectral residual weighting strategy is adopted to highlight the frequency components of the salient area through weighted adjustment of frequency domain information, thereby enhancing the representation of the salient target, thereby identifying the salient area in the image, effectively distinguishing the foreground target from the complex background, improving the algorithm's sensitivity to key targets and reducing background interference. On this basis, combined with adaptive spatial regularization technology, the spatial model of the target area is dynamically adjusted to provide higher tracking stability for targets in fast-moving or deforming scenes. Finally, the distortion suppression module corrects image distortion to ensure accurate tracking of the target morphology, thereby improving the robustness and accuracy of the target tracking algorithm in complex backgrounds and dynamic scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 This is a framework diagram of a target tracking method combining saliency perception and adaptive distortion suppression according to the present invention.

[0022] Figure 2 It is a flow chart of the overall solution in a specific embodiment of the present invention.

[0023] Figure 3 2 is a schematic diagram of comparing the accuracy and overlap success rate of the OTB100 dataset in a simulation experiment in a specific embodiment of the present invention.

[0024] Figure 4 Schematic diagram of the evaluation results of the basic properties of OTB100 in the simulation experiment in a specific embodiment of the present invention.

[0025] Figure 5 It is a schematic diagram of the tracking results of some sequences in OTB100 of the simulation experiment in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following describes embodiments of the present invention in detail, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention, and are not to be construed as limiting the present invention.

[0027] See also Figures 1 to 5The present invention provides a target tracking method combining saliency perception and adaptive distortion suppression, comprising the following steps:

[0028] S1: Obtain the input image for target tracking and convert it to a grayscale image. Image preprocessing helps simplify subsequent steps and facilitates the extraction of saliency information from the image. Image preprocessing can remove irrelevant color information and improve the saliency representation of the target area.

[0029] S2: After image preprocessing, the image is further processed using local feature extraction techniques to obtain the key features required for target tracking. First, the edge and texture information in the image is captured by calculating the local features of the image. These features mainly include the image gradient and local variance. The gradient reflects the strength of the edge in the image, while the local variance can reveal the complexity of the texture in the image. In terms of gradient calculation, the Sobel operator is used to process the image to obtain the gradient components Gx and Gy in the horizontal and vertical directions, respectively:

[0030] G x =Sobel(I,x),G y =Sobel(I,y) (1)

[0031] Where I is the input image, Gx and Gy represent the gradients of the image in the horizontal and vertical directions, respectively.

[0032] The gradient amplitude ‖▽I‖ is calculated as follows:

[0033]

[0034] In addition, the image is smoothed by a mean filter and the local variance σ is calculated. 2 , the variance reflects the texture complexity of the image area:

[0035]

[0036] Among them, I i Represents the pixel value in the image, μ I is the mean value of the local area, and N is the total number of pixels in the area.

[0037] S3: After extracting local features, the integrated feature vector is processed using an adaptive spectral residual weighting algorithm to enhance the saliency of the target region against a complex background. This algorithm analyzes the distribution characteristics of the feature vector in the frequency domain and adaptively adjusts the weights to make the features of the target region more prominent, thereby improving tracking accuracy and stability. The specific steps of this algorithm are:

[0038] (1) Adaptive weight matrix construction: Based on the calculated local features, an adaptive weight matrix (weightMatrix) is constructed. This matrix reflects the contribution of different regions to saliency detection. Regions with strong local features (such as edge regions or regions with complex textures) will be given higher weights. Through weighting, these important regions can receive more attention in the calculation of spectral residuals.

[0039] The weight matrix weightMatrix can be defined as the weighted sum of the gradient magnitude and the local variance:

[0040] weightMatrix=α·‖▽I‖+β1·σ 2 (4)

[0041] Among them, α and β1 are adjustment coefficients used to control the relative importance of gradient amplitude and local variance.

[0042] (2) Frequency component weighting: In the frequency domain, the image is Fourier transformed to obtain the amplitude and phase information of the frequency components. Then, an adaptive weight matrix is ​​used to weight each frequency component in the frequency domain. The amplitude of the spectral residual A(f) is multiplied element-by-element by the weight matrix to achieve weighting:

[0043] A′(f)=A(f)·weightMatrix(f) (5)

[0044] Among them, A(f) represents the amplitude in the frequency domain, A′(f) is the weighted amplitude, and weightMatrix(f) is the corresponding value of the weight matrix in the frequency domain.

[0045] (3) Spectral residual calculation: Based on the weighted frequency domain, continue to calculate the spectral residual. First, perform logarithmic processing on the frequency domain representation of the image to obtain the logarithmic value of the amplitude:

[0046] log(A(f))=log(|F(I)|) (6)

[0047] where F(I) is the Fourier transform of image I.

[0048] Then the spectrum residual is calculated. The spectrum residual R(f) is calculated by the difference between the logarithmic amplitude and its smoothing result:

[0049] R(f)=log(A(f))-Smooth(log(A(f))) (7)

[0050] The weighted spectral residual A′(f) is used to replace the original amplitude to obtain the final weighted spectral residual:

[0051] R′(f)=log(A′(f))-Smooth(log(A′(f))) (8)

[0052] S4: In this invention, saliency perception is achieved based on frequency domain information and spectral residual methods. First, the saliency map S of the image is calculated in the frequency domain. Then, the accuracy of target detection is enhanced by fusing the saliency features of the target region. The weighted spectral residual is converted back to the spatial domain by performing an inverse Fourier transform (IFFT) on it, thus obtaining the saliency map S:

[0053] S=|F -1 (exp(R′(f)+iφ(f)))| 2 (9)

[0054] Among them, F -1 represents the inverse Fourier transform operation, exp(R′(f)) represents the weighted spectral residual, and φ(f) is the phase information in the frequency domain.

[0055] The saliency map S reflects the important areas in the image, can highlight the target part, help distinguish the target from the background, avoid background interference, and thus improve the accuracy of target tracking.

[0056] S5: Adaptive regularization enhances the stability of the target area by constraining the frequency domain information, especially in scenes with rapid motion or deformation. Adaptive regularization constrains the frequency components to varying degrees according to the differences between the target and the background. In order to effectively improve the robustness and stability of the target tracking algorithm in complex scenes, the present invention proposes a filter weight optimization method based on adaptive regularization. This method dynamically adjusts the weight distribution of the filter by introducing a spatial regularization mechanism, thereby suppressing the influence of unreliable areas and ensuring that the target area has a higher response peak. The specific steps of this method are:

[0057] (1) Adaptive regularization constraint: The goal of adaptive regularization is to penalize the filter coefficients of unreliable areas and constrain the weight changes in combination with the prior reference weights, so that the filter can adaptively adjust the weights to adapt to changes in the target appearance. The regularization formula is:

[0058]

[0059] Among them, R adaptive represents the adaptive regularization term, u represents the current filter weight vector, q k represents the spatial regularization mask, which is used to dynamically adjust the weights of different regions. η1 and η2 are the weight parameters for adjusting the regularization constraints. K represents the number of feature channels. u ref Represents the prior reference weight, which is usually generated by the filter weight of the previous frame and is used to constrain excessive changes in weights.

[0060] (2) Spatial regularization mechanism: This method constructs a spatial regularization mask q k , dynamically adjust the regularization strength of the filter in different regions. First, calculate the regional weight. Utilize the feature differences between the target region and the background region (such as saliency features or texture features) to calculate the corresponding spatial mask q for each feature channel. k The weight of the target area is smaller, and the weight of the background area is larger, thereby suppressing the interference of background noise. Then regularize the weight distribution. The weight u of each channel of the filter and the spatial mask q k An element-wise product is performed to ensure that the response strength of the target area is higher, while the background area is more strongly constrained.

[0061] (3) Prior reference weight constraint: In order to further improve the stability of target tracking, the present invention introduces the prior reference weight u ref Constrain the change of the current weight u. The reference weight is generated by the filter weight of the previous frame, which can effectively capture the historical appearance characteristics of the target. The term limits the drastic changes in the filter weights and prevents tracking failures due to rapid target deformation or background interference.

[0062] (4) Parameter Adjustment and Optimization: During the implementation of adaptive regularization, the present invention balances the spatial regularization term's suppression of background noise and unreliable regions with the degree to which the reference weight constrains filter weight changes by adjusting the parameters η1 and η2. By alternately optimizing the filter weights and regularization parameters, the filter is able to maintain tracking performance under dynamically changing targets and complex environments.

[0063] S6: During target tracking, the response graph may be abnormally distorted due to factors such as background clutter, target occlusion, and illumination changes. This distortion can cause the response graph between adjacent frames to change dramatically, thereby affecting the stable tracking of the target. Therefore, the present invention proposes a distortion suppression regularization method that constrains the rate of change of the response graph to suppress the occurrence of abnormal responses and ensure the robustness of the model. The specific steps of this method are:

[0064] (1) Distortion suppression regularization: The goal of distortion suppression regularization is to limit the drastic changes in the response graph between adjacent frames, thereby suppressing the response distortion caused by abnormal factors. Its regularization formula is defined as follows:

[0065]

[0066] Among them, R distortion represents the distortion suppression regularization term, β represents the adjustment parameter of distortion suppression, Represents the feature input of the current frame, and is the filter weight of the current frame and the previous frame, P T represents the sample cropping matrix, Ψ p,q represents a similarity measurement function used to model changes in the response map, and D represents the number of feature channels. The formula measures the difference in the response maps between adjacent frames and constrains the corresponding rate of change using Euclidean distance, thereby suppressing the generation of abnormal responses.

[0067] (2) Distortion suppression mechanism: Ideally, the response maps between adjacent frames should maintain high similarity, reflecting the smooth changes of the tracked target. However, when abnormal factors such as background clutter or target occlusion exist, the Euclidean distance of the response map will increase sharply. This method introduces distortion suppression regularization to limit the rate of change between the response maps of adjacent frames. Distortion suppression regularization dynamically adjusts the weight distribution by considering the filter responses of the current frame and the previous frame at the same time. This mechanism ensures that the changes in the response maps of adjacent frames remain within a reasonable range, effectively suppressing response distortion.

[0068] (3) Parameter Adjustment and Optimization: By adjusting the parameter β, the present invention can achieve flexible control of the distortion suppression strength. When the value of β is large, the response map changes are more strongly restricted, thereby reducing the occurrence of abnormal responses; when the value of β is small, the model can more flexibly adapt to rapid changes in the target's appearance.

[0069] Furthermore, the present invention combines distortion suppression regularization with adaptive regularization to uniformly optimize the filter weight parameters. By alternately solving the distortion suppression and adaptive regularization objectives, it ensures a balance between tracking accuracy and robustness in dynamic and complex scenarios.

[0070] S7: During the target tracking process, in order to ensure that the model can accurately locate the position of the tracked target, the present invention adopts a target position update method based on a response graph after processing each frame, and updates the target's position coordinates by calculating the peak position in the response graph. The target position update is based on the target's response graph R(x,y), which represents the possible position distribution of the target in the search area. Usually, the peak position of the response graph corresponds to the most likely position of the target in the current frame. By parsing the value of the response graph R(x,y), the position of the maximum response value (x peak ,y peak ),Right now:

[0071] (x peak ,y peak )=argmax (x,y) R(x,y) (12)

[0072] To ensure that the target position update is highly robust in complex scenarios, the present invention combines distortion suppression and adaptive regularization strategies to correct the target position by dynamically adjusting the response map weights and suppressing abnormal responses. Specifically, the following steps are included:

[0073] (1) Weighted fusion of response maps: The distortion suppression regularization strategy is used to constrain the abnormal areas in the response map. By adjusting the weighting factor ω(x, y), a smooth weighted response map is obtained:

[0074] R weighted (x,y)=ω(x,y)·R(x,y) (13)

[0075] The ω(x,y) is dynamically adjusted according to the change rate of the response graph to suppress abnormal peak interference.

[0076] (2) Historical trajectory constraint: using the historical position of the target in the previous frames (x prev ,y prev ) provides a priori information to constrain the target position of the current frame and avoid drastic jumps caused by interference from the response graph.

[0077] (3) Adaptive step size update: According to the degree of change in the target appearance, the update step size η is adaptively adjusted to control the range of change of the target position:

[0078] (x final ,y final )=(1-η)·(x prev ,y prev )+η·(x opt ,y opt ) (14)

[0079] Where η is the dynamic step size factor, and its value range is [0,1].

[0080] In the target tracking optimization framework proposed in this invention, a comprehensive objective function is designed to give full play to the synergistic effect of each module. This objective function is based on adaptive spectral residual weighting (S3) and saliency perception (S4), and adds adaptive regularization (S5) and distortion suppression regularization (S6) constraints to improve the robustness and accuracy of the tracking model. Adaptive spectral residual weighting enhances the model's ability to capture key features by dynamically adjusting the saliency of the target response; saliency perception strengthens the focus of the target area by introducing saliency weights to guide the update of the filter; and adaptive regularization improves the filter's adaptability to changes in target appearance by penalizing unreliable areas of weights; distortion suppression regularization effectively suppresses the occurrence of abnormal responses by limiting the drastic changes in the response map.

[0081] Based on the above modules, the objective function designed by this invention comprehensively considers data loss, adaptive regularization, distortion suppression regularization, and saliency weighting to construct a unified optimization goal. The objective function expression is as follows:

[0082] L=L data +R adaptive +R distortion (15)

[0083] in

[0084]

[0085] Among them, L data Represents the data loss term, which is used to measure the degree of match between the response map and the target area, and contains the constraints of the significance weight on the target area; R adaptive is the adaptive regularization term; R distortion is the distortion suppression regularization term. The design of this objective function ensures the organic integration of various modules, thereby achieving more accurate and stable target tracking in complex environments.

[0086] Furthermore, to verify the effectiveness of the target tracking algorithm proposed in this paper, we provide systematic simulation experiments and conduct performance evaluation and comparison with several existing mainstream target tracking algorithms. The experiments were conducted on a public target tracking dataset, including typical challenging scenarios such as complex backgrounds, multi-scale changes, occlusion, illumination changes, and rapid motion. By comparing the performance of different algorithms in terms of tracking accuracy (Precision) and success rate (Success Rate), the performance of the proposed method was comprehensively evaluated.

[0087] The experiment selected four excellent target tracking algorithms as comparison, including ARCF, AutoTrack, MUSTER and BACF. In the OPE (One Pass Evaluation) evaluation on the OTB100 dataset, the overall performance of accuracy and overlap success rate was statistically analyzed. Figure 3 As shown, the proposed algorithm performs well across all metrics. On the OTB100 dataset, it achieves 85.4% accuracy and 81.2% overlap success rate, both outperforming the other compared algorithms. For example, on the OTB100 dataset, the proposed algorithm improves accuracy by 4.7% and overlap success rate by 5.7% compared to the baseline algorithm BACF. Table 1 shows the accuracy comparison of the top five algorithms on the OTB100 benchmark.

[0088] Table 1. Comparison of the top 5 trackers in terms of accuracy on the OTB100 benchmark

[0089]

[0090] In order to further analyze the performance of the algorithm in dealing with different challenging scenarios, the basic properties were evaluated. Figure 5 As shown in the figure, the accuracy and success rate of various algorithms were compared and evaluated for video sequences with five challenging attributes: background clutter (BC), deformation (DEF), occlusion (OCC), illumination variation (IV), and scale variation (SV) in the OTB100 dataset. The results show that the proposed algorithm performs well on all challenging attributes. This demonstrates that the proposed method, by combining adaptive spectral residual weighting and regularization strategies, improves the robustness and reliability of the model to complex boundary effects and changes in target appearance.

[0091] In addition, in order to verify the adaptability of the algorithm in different complex scenes, especially in the case of scale changes and background clutter, representative video sequences were selected as analysis objects. The Walking2, Twinnings and Tiger2 sequences contain scale changes, partial occlusions, motion blur and illumination changes. Figure 5 As shown in the figure, the algorithm of the present invention shows excellent tracking performance in these scenarios and can stably cope with various complex challenges.

[0092] In summary, the experimental results fully demonstrate the effectiveness of the target tracking algorithm proposed in this invention in complex scenarios.

[0093] The above disclosure is only a preferred embodiment of the present invention, and certainly cannot be used to limit the scope of the rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A joint saliency perception and adaptive distortion suppression target tracking method, characterized in that: The following steps are involved: Step 1: First, obtain the input image for target tracking and convert it into a grayscale image. Then, extract the saliency information of the target by preprocessing the image. Step 2: Calculate the gradient of the image to obtain the edge information of the image. Use the mean filter to calculate the local variance of the image to reflect the complexity of the texture. The local features are used to construct an adaptive weight matrix to enhance the contribution of important areas in the image. Step 3: Based on local features, an adaptive weighting matrix is ​​constructed to dynamically adjust the weights of different frequency components according to the local structure of the image. The weighting matrix is ​​applied to the frequency components in the frequency domain to perform weighted adjustment on the frequency amplitude of the image, thereby highlighting the frequency domain contribution of the salient target area. Step 4: Combined with the adaptive spectral residual weighted saliency detection method, the saliency map of the image is calculated in the frequency domain; Step 5: Dynamically constrain the spatial models of the target area and background area through an adaptive regularization strategy; Step 6: Apply distortion suppression regularization to suppress the distortion effect in the image and ensure that the shape change of the target does not affect the tracking accuracy; Step 7: Based on the weighted frequency domain information, the frequency domain image is converted back to the spatial domain through inverse Fourier transform to obtain the current frame image of the target, and the position of the next frame is predicted according to the target's motion law.

2. The joint saliency perception and adaptive distortion suppression target tracking method according to claim 1, wherein: During the execution of step 1, a background-aware feature extraction method is first adopted. By constructing a cyclic shift sample matrix containing the target area and its surrounding background area, and using the multi-channel feature representation of directional gradient histogram features and color features, the distinction between the target and the background is enhanced and more discriminative target features are extracted.

3. The joint saliency perception and adaptive distortion suppression target tracking method according to claim 2, wherein: During the execution of step 2, the Sobel operator is used to calculate the gradient of the image to obtain the edge information of the image. At the same time, the mean filter is used to calculate the local variance of the image to reflect the complexity of the texture. These local features are used to construct an adaptive weight matrix to enhance the contribution of important areas in the image.

4. The joint saliency perception and adaptive distortion suppression target tracking method according to claim 3, wherein: During the execution of step 3, an adaptive spectral residual weighted saliency detection method is adopted, which includes: first, using Fourier transform to convert the image from the spatial domain to the frequency domain to obtain the frequency components of the image; then, calculating local features of the image, such as gradient and local variance. These features help capture edge information and texture complexity in the image, reflecting the importance of different regions in saliency detection; then, based on these local features, constructing an adaptive weighting matrix, which can dynamically adjust the weights of different frequency components according to the local structure of the image. For regions containing salient targets, frequency components will be given higher weights, while for background regions, frequency components will be given relatively lower weights; finally, this weighting matrix is ​​applied to the frequency components in the frequency domain to perform weighted adjustment on the frequency amplitude of the image, thereby highlighting the frequency domain contribution of the salient target region, enhancing the representation ability of the target, and suppressing the interference of background noise.

5. The joint saliency perception and adaptive distortion suppression target tracking method according to claim 4, characterized in that: During the execution of step 4, the adaptive spectral residual weighted saliency detection method is combined to calculate the image saliency map in the frequency domain and combine it with the frequency domain information of the target. The saliency perception technology can help the algorithm pay more attention to the salient parts of the target, thereby improving robustness in dynamic scenes.

6. The joint saliency perception and adaptive distortion suppression target tracking method according to claim 5, characterized in that: During the execution of step 5, the spatial models of the target area and background area are dynamically constrained through an adaptive regularization strategy. In the specific implementation, a spatial weight matrix is ​​first generated based on the target detection confidence map, where the pixels in the target area are given a lower regularization coefficient to maintain tracking flexibility, and the pixels in the background area are given a higher regularization coefficient to suppress interference.

7. The joint saliency perception and adaptive distortion suppression target tracking method according to claim 6, characterized in that: During the execution of step 6, the anomaly suppression regularization term is constructed by calculating the Euclidean distance of the response maps of adjacent frames. When response distortion is detected, the quadratic constraint is dynamically applied, and its regularization strength is adaptively adjusted according to the current background clutter and occlusion level.

8. The joint saliency perception and adaptive distortion suppression target tracking method according to claim 7, characterized in that: During the execution of step 7, an inverse Fourier transform is performed on the weighted frequency domain features to obtain a spatial domain response. The target position is determined by detecting the peak position of the spatial domain response graph to achieve target tracking.

Images