Target tracking method based on response fusion and adaptive scale processing

Abstract

This invention discloses a target tracking method based on response fusion and adaptive scale processing. Addressing the shortcomings of correlation filter-based tracking algorithms, such as insufficient effectiveness and richness of feature extraction, poor handling of rapid scale changes, and weak anti-interference capabilities, this invention proposes a scale-adaptive target tracking algorithm based on feature response fusion. First, a probabilistic perception model based on color histogram features is introduced. Based on the distribution differences between the background and target regions, a background suppression model and an interference detection model are established. Then, the fusion weights are adaptively calculated using the probability mean of the target region, and the background suppression model and the interference detection model are adaptively fused according to the real-time scene. Finally, in terms of scale processing, a target prediction model incorporating effective extrema edge detection is used to calculate target scale changes during tracking, improving the algorithm's accuracy when handling scale-changing scenarios.

Description

Technical Field

[0001] This invention belongs to the field of image and video processing technology, and further relates to a target tracking method based on response fusion and adaptive scale processing in the field of computer vision. Background Technology

[0002] Target tracking, a research hotspot in image and video processing, plays a crucial role in practical applications such as moving object tracking, traffic control, intelligent robots, radar guidance, medical image processing, surveillance systems, human-computer interaction, pedestrian detection, and action recognition. The core task of target tracking is to accurately estimate the target's position, motion state, and scale. Correlation filter-based target tracking algorithms initialize a filter using the first frame of information. When detecting subsequent image inputs, feature extraction and modeling calculations are performed, and the filter is multiplied element-wise to obtain the output response. The peak position of the response is then located to obtain the target position in the new frame. In practical applications, targets often exhibit characteristics such as rapid movement, deformation, appearance changes, lighting blur, and scale changes, and the tracking scene often contains complex background interference. Traditional correlation filter-based target tracking algorithms typically have relatively simple features, thus limiting their applicability to certain target types and exhibiting weak resistance to interference from complex backgrounds. When encountering complex situations, these algorithms lack scene judgment and processing capabilities, easily leading to tracking drift and target loss. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a target tracking method based on response fusion and adaptive scale processing. This method solves the problems of insufficient effectiveness and richness of feature extraction in correlation filter-based tracking algorithms, poor ability to cope with rapid scale changes, and weak anti-interference ability, thereby improving the overall performance of target tracking algorithms in scenarios with background clutter, similar interference, and scale changes.

[0004] To achieve the above objectives, the specific steps of the present invention are as follows:

[0005] The target tracking method based on response fusion and adaptive scaling includes the following steps:

[0006] Step 1: Before the algorithm starts, initialize the parameters, including fixed weights and estimated values, and assign values ​​to them;

[0007] Step 2: For the first frame of the input image, calculate the background suppression model and the interference detection model, and fuse them to obtain the background prediction model;

[0008] Step 3: Initialize the correlation filter model and extract basic sample blocks from the target location and scale of the first frame image;

[0009] Step 4: Calculate the correlation filter response res based on the detected image in frame t. cf Color histogram response res co and adaptive fusion parameter v;

[0010] Step 5: In the detected image of frame t, calculate the fusion response res, locate the peak position of the fusion response res, and obtain the target position in the image of frame t+1;

[0011] Step 6: In the detection image of frame t, scale calculation is performed. First, image blocks are extracted from the detection image of frame t and their effective values ​​are calculated. Then, edge detection is performed and adaptive scale processing is applied to the detection results. After the scale calculation is completed, the target scale of frame t+1 is obtained. Combined with the target position in step 5, target tracking can be achieved.

[0012] Step 7: After completing the detection part, proceed to the correlation filter training part;

[0013] Step 8: Process the next frame image, repeating steps 4 to 7.

[0014] In one embodiment, the parameter initialization in step 1, during the process of calculating the color histogram response, involves: adding a color histogram learning rate and a color histogram channel length when processing the color histogram to achieve better feature extraction results; and adding histogram model fusion weights when calculating the color histogram response to fuse the background suppression model and the interference detection model, thereby improving the background suppression capability of the color histogram response and the anti-interference capability of the target similarity.

[0015] In one embodiment, step 2, the model calculation steps are as follows:

[0016] Step 2.1, during the model training phase, the background suppression model probability distribution lookup table formula is expressed as:

[0017]

[0018] Where I is the input image, T is the target region, S is the background region, and C is the target region. o C represents the combination of RGB values ​​for pixel o. Let C be the number of pixels in the target region T of the input image I whose RGB values ​​combine to form C. Let ψ be the number of pixels in the background region S of the input image I whose RGB value combination is C, and let ψ be the background suppression histogram, expressed as ψ = 1 - ξ·(P(C)). o |o∈S)-P(C o |o∈T)), where ξ is the background suppression coefficient, used to control the degree of background suppression, and I(o)∈I(T∪S) represents the condition when pixel o is in the target region T or the background region S;

[0019] Step 2.2, during the model training phase, the probability distribution lookup table formula for the interference detection model is expressed as:

[0020]

[0021] Where G represents the potential interference region, and w represents the set enhancement weight. The number of pixels in the potential interference region G of the input image I whose RGB value combination is C;

[0022] Step 2.3: Fuse the background suppression model and the interference detection model to obtain the background prediction model, expressed as:

[0023] P(o∈T|C o )=τ·P(o∈T|T,S,C o ) * +(1-τ)·P(o∈T|T,G,C o ) *

[0024] In the formula, τ represents the fusion weight of the histogram model.

[0025] In one embodiment, step 3, the initialization steps of the correlation filter model, are as follows:

[0026] Step 3.1: In the process of solving the correlation filter response, the regularized least squares method is used to solve the optimization problem of solving the correlation filter model;

[0027] Step 3.2, through the operator Mapping sample q to the feature space improves the sample's representation ability;

[0028] Step 3.3, the closed-form solution of kernel-regularized least squares is as follows:

[0029] α=(K+λI) -1 y

[0030] In the formula, α is the template filter, and each element is α. i y is the standard output vector, and each element of it is y i I is the identity matrix, λ is a constant that assigns size to the identity matrix, usually set before training, and K is a circular matrix with K elements. ij =κ(q) i q j ), (q i q j ) represents the elements in the q rows of a linear sample;

[0031] Step 3.4: Apply the relevant properties of the circulant matrix K to the following formula to obtain the expression formulas for the template filter α and the template feature x;

[0032]

[0033] F(x t )=(1-η)F(x t-1 )+ηF(x new )

[0034] F(α t )=(1-η)F(α t-1 )+ηF(α new )

[0035] In the formula, η is the learning rate, and x t-1 α t-1 Let x represent the template features and template filter of the previous frame, respectively. new α new x represents the newly calculated template features and template filter in the current frame, respectively. t α t These represent the final template features and template filter of the current frame, respectively.

[0036] In one embodiment, step 4 involves the following calculation steps:

[0037] Step 4.1: In the detected image of frame t, the target location information p from the previous frame image is used... t-1 and scale s t-1 Extract search box modules and region features, and check the color probability distribution obtained from the background suppression model training in the LUT table. T The color probability distribution check table LUT obtained from the training of the anti-interference model G The color probability distribution checklist obtained from the training of the background prediction model (LUT) S Calculate the target probability image M and the background probability image M of the frame to be inspected. s And integrate M to obtain the integral image M. int The formula is as follows:

[0038]

[0039] In the formula, M is calculated using the probability distribution lookup table formula of the interference detection model in the previous frame, (m, n) represents the row and column index of M, and (i, j) represents the index of M. int The row and column indices, the above formula represents the accumulation operation performed on the elements of the target probability image M;

[0040] Step 4.2, via M, M s M int Calculate the color histogram response resco ;

[0041]

[0042] In the formula, These represent the top-left, bottom-right, top-right, and bottom-left sampling regions of the integral image, respectively, with the size of each region being the difference between the size of the integral image and the target size; |T| represents the area of ​​the target region.

[0043] Step 4.3: Calculate the correlation filter response res using the template filter α and the template feature x. cf The formula is as follows:

[0044]

[0045] Step 4.4, for M int Hybrid sampling detection is performed to obtain the potential interference region G and the optimal target region T'. The adaptive fusion parameter v is calculated from T' and M. Finally, a color histogram-based perceptual model is trained, and the LUT is updated using T' and G during the processing of the next frame image. T ,lut G ,lut S ;

[0046]

[0047] In the formula, th is the overfitting weight threshold;

[0048]

[0049]

[0050]

[0051] In the formula, These are the calculated probability distribution lookup tables, where the superscript t represents the final parameter of the current frame, the superscript t-1 represents the parameter of the previous frame, the superscript new represents the most recently calculated parameter, and η c The learning rate is the histogram parameter.

[0052] In one embodiment, the formula for calculating the fusion response res in step 5 is as follows:

[0053] res=υ·res co +(1-υ)·res cf

[0054] Locate the peak position of res to obtain the target position p. t .

[0055] In one embodiment, step 6, the formula for edge detection, is as follows:

[0056]

[0057] In the formula m v The edge strength is expressed as follows:

[0058]

[0059] In the formula th b The edge intensity sensitivity coefficient, max v min v Let g1 and g2 be the valid maximum and minimum values ​​of the image, respectively. Their expressions are:

[0060]

[0061]

[0062] In the formula, M is the gray value of pixel o, and the number in the lower right corner indicates the position index of o in the 3×3 traversal template.

[0063] In one embodiment, p t Image patches are extracted, and effective maximum and minimum values ​​are obtained using gray-level histogram statistics. First, two pixel percentages are defined. Then, the corresponding gray-level values ​​that satisfy these two percentages are extracted from the gray-level distribution histogram, thus yielding the effective maximum value (max) of the image. v and the effective minimum value of the image min v .

[0064] In one embodiment, the adaptive scaling process for the detection results includes the following steps:

[0065] (1) Perform shape edge detection based on effective maximum and minimum values ​​on the input image patch;

[0066] (2) Traverse the detected image and calculate the score s for each traversal box using the traversal box scoring formula. r The calculation formula is:

[0067]

[0068] In the formula, m i For different edge groups e i The edge strength m of all edge points v The sum of, r w and r h These are the width and height of the selected rectangle r, respectively. in κ is the inner center box of r, with a height and width that are half the size of r, and κ is the deviation coefficient.

[0069] (3) Scoring sr Filter and retain candidate boxes with scores higher than the set threshold, and record their positions and sizes;

[0070] (4) The obtained candidate boxes are traversed again, and each time a sub-image patch of the size of the candidate box is extracted. The correlation filter response is calculated with the template features in the tracking according to the following formula:

[0071]

[0072] In the formula, It is the vector form of the Gaussian response of the sub-image patch and the template features;

[0073] (5) If res p If the size of the target image block reaches the set threshold, it means that the scale calculation confidence of the corresponding sub-image block is the highest, and the target scale needs to be updated. The scale size of the sub-image block is then taken as the final target scale size.

[0074] In one embodiment, the training part of step 7, correlation filtering, is performed by p t s t Extract basic sample blocks, train and update template filter α and template feature x, and use them in the next frame.

[0075] Compared with the prior art, the beneficial effects of the present invention are:

[0076] In experiments with complex clutter backgrounds, scale variations, and motion blur, the proposed method significantly improves both AUC and tracking accuracy compared to other comparative algorithms. This is because the proposed method establishes a fusion model combining background suppression and interference region detection on top of the traditional color histogram model, providing resistance to complex background interference. Furthermore, it introduces an adaptive scale processing strategy based on shape edge detection, resulting in better tracking performance in scenes with significant scale variations. The proposed method also fuses responses from different features, overcoming the limitations of single features and maintaining high tracking accuracy even in motion-blurred scenes. In summary, the proposed algorithm outperforms commonly used existing technologies in terms of tracking robustness, tracking success rate, AUC, and tracking accuracy, and demonstrates superior performance in complex backgrounds, scenes with significant scale variations, and motion blur. Attached Figure Description

[0077] Figure 1 This is a framework diagram of the target tracking method based on response fusion and adaptive scaling processing of the present invention.

[0078] Figure 2 This is a schematic diagram of the workflow framework for the algorithm training and detection parts in this invention.

[0079] Figure 3 This is a schematic diagram illustrating the interaction between the background suppression model and the interference detection model in this invention.

[0080] Figure 4 This is a schematic diagram of mixed sampling in this invention.

[0081] Figure 5 This is a schematic diagram of the fusion response in this invention.

[0082] Figure 6 This is a flowchart of the parameter adaptive edge detection algorithm in this invention.

[0083] Figure 7 This is a flowchart of the adaptive scaling process for edge detection in this invention.

[0084] Figure 8 This is a schematic diagram comparing the results of different algorithms in an embodiment of the present invention. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0086] To demonstrate the effectiveness of the method of this invention, target tracking was performed using two authoritative publicly available test sets, OTB-100 and VOT datasets. The experiments were conducted on a computer with an AMD Ryzen 7 5800H CPU (3.2GHz), 16GB of RAM, and a 64-bit Windows 10 operating system, using MATLAB R2021b software. Several sets of sequences from typical complex scenarios in the aforementioned datasets were selected as experimental subjects. These sets of sequences included 10 tracking challenges from the OTB-100 and VOT datasets. For both datasets, refer to... Figure 1 and Figure 2 The basic process of the target tracking method based on response fusion and adaptive scaling in this invention is as follows:

[0087] Step 1: Before the algorithm begins, a series of necessary parameter initializations are performed, mainly including fixing weights, estimating values, etc., and assigning a series of values. New techniques are added to the parameter initialization process. For example, in the process of calculating the color histogram response, basic parameters such as the color histogram learning rate and color histogram channel length are added to achieve better feature extraction results when processing the color histogram. Furthermore, when calculating the color histogram response, histogram model fusion weights are added to fuse the background suppression model and the interference detection model, thereby improving the background suppression capability and the anti-interference capability of the target similarity object in the color histogram response. Specific settings in this embodiment are shown in Table 1.

[0088] Table 1 Initialization parameter settings.

[0089] <![CDATA[σ0]]> Standard output bandwidth factor 0.0625 σ Gaussian kernel response bandwidth 0.2 λ Regularization parameters 10-4 η Correlation filter learning rate 0.025 <![CDATA[η c ]]> Color histogram learning rate 0.05 <![CDATA[n h ]]> Color histogram channel length 16 th Overfitting weight threshold 0.58 τ Histogram model fusion weights 0.5 ξ Background suppression coefficient 9.5 d Scale calculation of search box magnification 1.4 <![CDATA[th b ]]> Edge strength sensitivity coefficient 0.06 μ Relevance Sensitivity Coefficient 2

[0090] Step 2: For the first input frame image, initialize the color histogram probability perception model, that is, calculate the color probability distribution check table LUT obtained from the background suppression model training. T The color probability distribution check table LUT obtained from the training of the anti-interference model G The color probability distribution checklist LUT obtained by training the fusion model of background suppression model and interference region detection model (i.e., background prediction model) S .

[0091] In the embodiments of this invention, the images used are all derived from the publicly available datasets OTB-100 and VOT. When the first frame image is input, the color histogram model is initialized. First, a complete image patch containing the background box is extracted from the target position p1 and scale s1 of the first frame. Then, the probability distribution lookup table lut of the background suppression model, interference detection model, and background prediction model is calculated respectively. T ,lut G ,lut S The results of the interaction between the background suppression model and the interference detection model are as follows: Figure 3 As shown. The specific steps are as follows.

[0092] Step 2.1, during the training phase of tracking, the input image I is an expanded image patch containing the target region. Assuming the target region T has been obtained, the background region S outside T can be determined accordingly. Starting from the definition of the color histogram, using... Let C represent the number of pixels in the target region T of the input image I with RGB value combinations equal to C, and let S represent the number of pixels in the background region S of the input image I with RGB value combinations equal to C, respectively. Finally, a background suppression histogram ψ is added. The probability distribution lookup table formula for the background suppression model can be expressed as follows:

[0093]

[0094] In the formula, C o Let C represent the combination of RGB values ​​of pixel o. The expression for the background suppression histogram ψ is ψ = 1 - ξ·(P(C)). o |o∈S)-P(C o ∣o∈T)), ξ is the background suppression coefficient, which is used to control the degree of background suppression, and I(o)∈I(T∪S) represents the condition when pixel o is in the target region T or the background region S.

[0095] Step 2.2, during the training phase of tracking, assuming that the target region T and a set of potential interference regions G have been obtained, based on the color histogram information and by adding the set enhancement weights w, the probability distribution lookup table of the interference detection model can be obtained, expressed by the following formula.

[0096]

[0097] In the formula, Let C be the number of pixels in the potential interference region G of the input image I whose RGB value combinations are C.

[0098] Step 2.3: To improve background suppression and anti-interference capabilities against similar targets, the background suppression model and interference detection model are fused. Based on the probability distribution lookup table formula for the background suppression model in Step 2.1 and the probability distribution lookup table formula for the interference detection model in Step 2.2, the background suppression model and interference region detection model are fused, and the following formula is used as the background prediction model formula:

[0099] P(o∈T|C o )=τ·P(o∈T|T,S,C o ) * +(1-τ)·P(o∈T|T,G,C o ) *

[0100] In the formula, τ represents the fusion weight of the histogram model.

[0101] Step 3: When inputting the first frame image, the correlation filtering model needs to be initialized. First, basic sample blocks are extracted from the target location p1 and target scale s1 in the input first frame image. Then, the template filter α and template feature x are calculated respectively. The specific steps are as follows:

[0102] Step 3.1: In solving the correlation filter response, the template filter α and template feature x are calculated for the correlation filter model. Regularized least squares (RLS) is used to solve the optimization problem of solving the correlation filter model. The loss function expression of RLS is:

[0103] L(y, f(x)) = (yf(x)) 2

[0104] Step 3.2, to further improve the representational ability of sample q, a kernel operator is defined. Pass sample q through Mapping to the feature space enhances the representational power of the samples.

[0105] Step 3.3, the form of the simple closed-form solution of kernel-regularized least squares is:

[0106] α=(K+λI)-1 y

[0107] In the formula, α is the template filter, and each element is α. i y is the standard output vector, and each element of it is y i I is the identity matrix, λ is a constant that assigns size to the identity matrix, usually set before training, and K is a circular matrix with K elements. ij =κ(q) i q j ), (q i q j ) represents the elements in the q rows of a linear sample.

[0108] Step 3.4: Applying the relevant properties of the cyclic matrix K to the following formula, we can obtain the expression formulas for the template filter α and the template feature x.

[0109]

[0110] F(x t )=(1-η)F(x t-1 )+ηF(x new )

[0111] F(α t )=(1-η)F(α t-1 )+ηF(α new )

[0112] In the formula, η is the learning rate, and x t-1 α t-1 Let x represent the template features and template filter of the previous frame, respectively. new α new x represents the newly calculated template features and template filter in the current frame, respectively. t α t These represent the final template features and template filter of the current frame, respectively.

[0113] Step 4: Calculate the correlation filter response res based on the detected image in frame t. cf Color histogram response res co And the adaptive fusion parameter v.

[0114] In this step, the correlation filter response res is calculated using the previous frame image. cf Color histogram response res co The adaptive fusion parameter v is then calculated and updated as the processing parameter for the next frame. The specific steps are as follows:

[0115] Step 4.1: In the detected image of frame t, the target location information p from the previous frame image is used... t-1 and scale st-1 Extract search box modules and region features, and check the color probability distribution obtained from the background suppression model training in the LUT table. T The color probability distribution check table LUT obtained from the training of the anti-interference model G The color probability distribution check table LUT obtained from the training of the fusion model S Calculate the target probability image M and the background probability image M of the frame to be inspected. s And integrate M to obtain the integral image M. int The formula is as follows:

[0116]

[0117] In the formula, M is the target probability image of the frame to be detected, which can be calculated by the probability distribution lookup table formula of the color histogram interference detection model of the previous frame, (m, n) represents the row and column index of M, and (i, j) represents the value of M. int The row and column indices are given by the above formula, which represents the accumulation operation performed on the elements of the target probability image M.

[0118] Step 4.2, via M, M s M int Calculate the color histogram response res co ;

[0119]

[0120] In the formula, These represent the upper left, lower right, upper right, and lower left sampling regions of the integral image, respectively. The size of each region is the difference between the size of the integral image and the target size; |T| represents the area of ​​the target region.

[0121] Step 4.3: Calculate the correlation filter response res using the template filter α and the template feature x. cf The formula is as follows:

[0122]

[0123] Step 4.3, refer to Figure 4 , for M int Hybrid sampling detection is performed to obtain the potential interference region G and the optimal target region T'. The adaptive fusion parameter v is calculated from T' and M. Finally, a color histogram-based perceptual model is trained, and the LUT is updated using T' and G during the processing of the next frame. T ,lut G ,lut S .

[0124]

[0125] In the formula, th is the overfitting weight threshold;

[0126]

[0127]

[0128]

[0129] In the formula, These are the calculated probability distribution lookup tables, where the superscript t represents the final parameter of the current frame, the superscript t-1 represents the parameter of the previous frame, the superscript new represents the most recently calculated parameter, and η c The learning rate is the histogram parameter.

[0130] Step 5: In the detected image of frame t, calculate the correlation filter response res cf and color histogram response res co The peak position of the fused response res is located to obtain the target position in the (t+1)th frame of the image. Figure 5 As shown, the calculation formula is as follows:

[0131] res=υ·res co +(1-υ)·res cf

[0132] Locate the peak position of res to obtain the target position p. t .

[0133] Step 6: In the detected image of frame t, perform scale calculation, first using the scale size s of the detected image of frame t. t and location information p t Image patches are extracted and their effective values ​​are calculated. Edge detection is then performed, and the detection results are adaptively scaled. After scale calculation, the target scale for frame t+1 is obtained. Combined with the target position from step 5, target tracking can be achieved. The flowchart of the parameter-adaptive edge detection algorithm is as follows: Figure 6 The formula for edge detection is as follows:

[0134]

[0135] In the formula m v The edge strength is expressed as follows:

[0136]

[0137] In the formula th b The edge intensity sensitivity coefficient, max v min v Let g1 and g2 be the valid maximum and minimum values ​​of the image, respectively. Their expressions are:

[0138]

[0139]

[0140] In the formula, M is the gray value of pixel o, and the number in the lower right corner indicates the position index of o in the 3×3 traversal template.

[0141] In this step, image patches are extracted and the effective maximum and minimum values ​​of the image are extracted using gray-level histogram statistics. First, two pixel percentages are defined (usually 10% and 90%). Then, the corresponding gray-level values ​​that satisfy these two percentages are extracted from the gray-level distribution histogram, thus obtaining the effective maximum value (max) of the image. v and the effective minimum value of the image min v .

[0142] This step performs adaptive scaling on the edge detection results, referencing... Figure 7 The steps are as follows:

[0143] (1) Perform shape edge detection based on effective maximum and minimum values ​​on the input image patch;

[0144] (2) Traverse the detected image and calculate the score s for each traversal box using the traversal box scoring formula. r The calculation formula is:

[0145]

[0146] In the formula, m i For different edge groups e i The edge strength m of all edge points v The sum of, r w and r h These are the width and height of the selected rectangle r, respectively. in κ is the inner center box of r, with a height and width that are half the size of r, and κ is the deviation coefficient.

[0147] (3) Scoring s r Filter and retain candidate boxes with scores higher than the set threshold, and record their positions and sizes;

[0148] (4) The obtained candidate boxes are traversed again, and each time a sub-image patch of the size of the candidate box is extracted. The correlation filter response is calculated with the template features in the tracking according to the following formula:

[0149]

[0150] In the formula, α is the vector form of the template filter in the correlation filter, and T represents the transpose operation. It is the vector form of the Gaussian response of the sub-image patch and the template features;

[0151] (5) If res p If the size of the target image block reaches the set threshold, it means that the scale calculation confidence of the corresponding sub-image block is the highest, and the target scale needs to be updated. The scale size of the sub-image block is then taken as the final target scale size.

[0152] Step 7: After completing the detection part, proceed to the correlation filter training part.

[0153] The training part of the correlation filter in this step is done by p t s t Extract basic sample blocks, train and update template filter α and template features x, and use them in the next frame image.

[0154] Step 8: Process the next frame image, repeating steps 4 to 7.

[0155] The target tracking method based on response fusion and adaptive scaling in this embodiment of the invention performs as follows on the BlurOwl sequence (a typical fast motion and motion blur sequence) in the public dataset OTB dataset: Figure 8 As shown, shortly after tracking begins, the target's edges become blurred due to camera shake, causing KCFDP tracking to fail. At frame 120, target motion blur causes SAMF and DSST to drift. At frame 160, the target exhibits more severe shaking and blurring, causing all algorithms to drift. However, the method described in this invention shows the smallest drift and promptly captures the target after it returns to normal. By frame 631, only the method described in this invention (FRFT) tracks correctly; the other algorithms have drifted into the background.

Claims

1. A target tracking method based on response fusion and adaptive scaling, characterized in that, Includes the following steps: Step 1: Before the algorithm starts, initialize the parameters, including fixed weights and estimated values, and assign values. In the process of calculating the color histogram response: when processing the color histogram, add the color histogram learning rate and color histogram channel length to achieve better feature extraction results; when calculating the color histogram response, add histogram model fusion weights to fuse the background suppression model and the interference detection model to improve the background suppression capability and the anti-interference capability of the target similar object of the color histogram response. Step 2: For the first input frame image, calculate the background suppression model and the interference detection model, and fuse them to obtain the background prediction model; the model calculation steps are as follows: Step 2.1, during the model training phase, the background suppression model probability distribution lookup table formula is expressed as: in, I For the input image, T For the target area, S For background area, Represents pixels o The combination of RGB values ​​is C , For the input image I Target area T The RGB value combination inside is C The number of pixels, For the input image I Mid-background area S The RGB value combination inside is C The number of pixels, The background suppression histogram is expressed as follows: In the formula This is the background suppression coefficient, used to control the degree of background suppression. Indicates when pixel o In the target area T or background area S conditions; Step 2.2, during the model training phase, the probability distribution lookup table formula for the interference detection model is expressed as: in, G As a potential area of ​​interference, The set enhancement weights, For the input image I Potential interference areas G The RGB value combination inside is C The number of pixels; Step 2.3: Fuse the background suppression model and the interference detection model to obtain the background prediction model, expressed as: In the formula For histogram model fusion weights; Step 3: Initialize the correlation filter model and extract basic sample blocks from the target location and scale of the first frame image; Step 4: Calculate the correlation filter response based on the detected image in frame t. Color histogram response and adaptive fusion parameters v ; Step 5: Calculate the fusion response in the detected image of frame t. Location fusion response The peak position is used to obtain the target position in the (t+1)th frame of the image; Step 6: In the detection image of frame t, scale calculation is performed. First, image blocks are extracted from the detection image of frame t and their effective values ​​are calculated. Then, edge detection is performed and adaptive scale processing is applied to the detection results. After the scale calculation is completed, the target scale of frame t+1 is obtained. Combined with the target position in step 5, target tracking can be achieved. Step 7: After completing the detection part, proceed to the correlation filter training part; Step 8: Process the next frame image, repeating steps 4-7.

2. The target tracking method based on response fusion and adaptive scaling processing according to claim 1, characterized in that, In step 3, the initialization steps of the correlation filter model are as follows: Step 3.1: In the process of solving the correlation filter response, the regularized least squares method is used to solve the optimization problem of solving the correlation filter model; Step 3.2, through the operator Sample q Mapping to the feature space to improve the representational ability of the samples; Step 3.3, the closed-form solution of kernel-regularized least squares is as follows: In the formula, For template filters, each element is , y This is the standard output vector, and each element is... , I It is the identity matrix. It is a constant that assigns the size of the identity matrix, and is usually set before training. K The matrix is ​​a circular matrix with the following elements: , For linear samples q The elements in the row represent; Step 3.4, circular matrix K The relevant properties are used in the following equation to obtain the template filter. Template features The expression formula; In the formula, For learning rate, , These represent the template features and template filter of the previous frame, respectively. , These represent the newly calculated template features and template filter in the current frame, respectively. , These represent the final template features and template filter of the current frame, respectively.

3. The target tracking method based on response fusion and adaptive scaling processing according to claim 1, characterized in that, In step 4, the calculation steps are as follows: Step 4.1: In the detected image of frame t, the target location information from the previous frame image is used... and scale Extract search box modules and region features, and use a color probability distribution checklist obtained from training a background suppression model. Color probability distribution checklist obtained from anti-interference model training Checklist of color probability distribution obtained from training the background prediction model Calculate the target probability image of the frame to be inspected Background probability image and to Integrating to obtain the integral image The formula is as follows: In the formula, The probability distribution of the interference detection model in the previous frame is calculated using the lookup table formula. express row and column indexes, express The row and column indices, as shown in the formula above, represent the probability image of the target. The elements are incremented. Step 4.2, through , , Calculate the color histogram response ; In the formula, , , , These represent the upper left, lower right, upper right, and lower left sampling regions of the integral image, respectively, with the size of each region being the difference between the size of the integral image and the target size. Indicates the area of ​​the target region; Step 4.3, by template filter Template features Calculate the correlation filter response The formula is as follows: Step 4.4, for Hybrid sampling detection is performed to obtain the potential interference region G and the optimal target region T', which are then combined with... M Calculate adaptive fusion parameters v Finally, a color histogram-based perception model is trained, and T' and G are used to update the model during the next frame image processing. , , ; In the formula, th This is the overfitting weight threshold; In the formula, , , These are the calculated probability distribution lookup tables, with superscripts... t Indicates the final parameters of this frame, superscript t -1 indicates parameters from the previous frame, superscript new This represents the latest calculated parameters. The learning rate is the histogram parameter.

4. The target tracking method based on response fusion and adaptive scaling processing according to claim 1, characterized in that, In step 5, the fusion response is calculated. The formula is shown below: position The peak position is used to obtain the target position. .

5. The target tracking method based on response fusion and adaptive scaling processing according to claim 1, characterized in that, In step 6, the formula for edge detection is as follows: In the formula The edge strength is expressed as follows: In the formula The edge intensity sensitivity coefficient, , These are the maximum and minimum valid values ​​of the image, respectively. , The expression is In the formula, M represents the number of pixels. o The grayscale value, with the number in the lower right corner indicating... o Position index in the 3×3 traversal template.

6. The target tracking method based on response fusion and adaptive scaling processing according to claim 5, characterized in that, Depend on Image patches are extracted, and effective maximum and minimum values ​​are obtained using gray-level histogram statistics. First, two pixel percentages are defined. Then, the corresponding gray-level values ​​that satisfy these two percentages are extracted from the gray-level distribution histogram, thus yielding the effective maximum values ​​of the image. and the minimum effective value of the image .

7. The target tracking method based on response fusion and adaptive scaling processing according to claim 5, characterized in that, The adaptive scaling process for the detection results is as follows: (1) Perform shape edge detection based on effective maximum and minimum values ​​on the input image patch; (2) Traverse the detected image and calculate the score of each traversal box using the traversal box scoring formula. The calculation formula is: In the formula, For different edge groups Edge strength of all edge points the sum of and The selected rectangles are respectively r Width and height, yes r The inner center frame of [the object] has a height and width of [size]. r 1 / 2, This is the deviation coefficient; (3) Scoring Filter and retain candidate boxes with scores higher than the set threshold, and record their positions and sizes; (4) Repeat the process for obtaining the candidate boxes, extracting a sub-image patch of the same size as the candidate box each time, and calculate the correlation filter response with the template features in the tracking according to the following formula: In the formula, It is the vector form of the Gaussian response of the sub-image patch and the template features; (5) If If the size of the target image block reaches the set threshold, it means that the scale calculation confidence of the corresponding sub-image block is the highest, and the target scale needs to be updated. The scale size of the sub-image block is then taken as the final target scale size.

8. The target tracking method based on response fusion and adaptive scaling processing according to claim 1, characterized in that, The training part of the correlation filter in step 7 is... , Extract basic sample blocks and train and update template filters. Template features Used in the next frame.

Images