Target tracking method and system fusing co-occurrence statistics and fhog gradient features

Abstract

The application provides a target tracking method and system fusing co-occurrence statistics and fhog gradient features, and the method comprises the following steps: determining an initial position according to a target region; calculating a target position p t and a scale factor s t‑1 of a previous frame of a current frame I t‑1 , so as to determine a target region and extract relevant features of the current frame; co-occurrence filter is used to obtain co-occurrence statistical information between pixel pairs in a tracked target image, fDSST algorithm is used to obtain fhog features of the tracking target, and the co-occurrence statistical information and the fhog features of the tracking target are fused, so as to be used as target image features; when the model is updated, the target region is determined according to a target center p t and a target scale s t of the current frame I t , and the target region is extracted; the target region extraction features are sent into a position filter and a scale filter, so as to obtain a target tracking result through iterative updating. The application solves the technical problems that the prior art cannot completely and effectively use feature expression of a target, has low robustness, and has poor tracking effect in a specific scene.

Description

Technical Field

[0001] This invention relates to the field of computer vision, specifically to a target tracking method and system that integrates co-occurrence statistics and fhog gradient features. Background Technology

[0002] Target tracking is a crucial research topic in computer vision, with wide applications in both military and civilian scenarios. After decades of research, numerous high-performance algorithms have emerged in target tracking in recent years. For example, trackers based on Discriminative Correlation Filters (DCF) have achieved excellent performance in terms of accuracy and robustness. The Minimum Output Sum Of Squared Error (MOOSE) filter proposed by Bolme et al. first introduced the concept of correlation filters into the field of target tracking, and its remarkable performance in both tracking accuracy and speed demonstrates the enormous potential of correlation filtering in target tracking. Henriques et al. proposed a kernelized correlation filter (KCF), which utilizes discrete Fourier transform for fast frequency domain computation and extends the correlation filter from a single-channel to a multi-channel model using the Histogram of Oriented Gradients (HOG) feature. For example, the existing invention patent application document CN108470355A, entitled "Target Tracking Method Integrating Convolutional Network Features and Discriminative Correlation Filters," establishes an end-to-end lightweight network architecture. It trains convolutional features by learning rich streaming information from consecutive frames, improving feature representation and tracking accuracy. The correlation filter tracking component is constructed as a special layer in the network to track a single image patch. During tracking, it simultaneously tracks the target patch and multiple background patches. By perceiving the structural relationship between the target and surrounding background patches, it builds a model for the highly identifiable parts of the target and its surrounding environment. The target tracking effect is measured by the peak-to-sidelobe ratio and the peak relationship of the confidence map. In situations where tracking is difficult, such as large-area occlusion, extreme target deformation, or drastic changes in illumination, it automatically utilizes discriminative background parts for localization. As can be seen from the description of this existing patent application, this prior art utilizes discriminative correlation filters and a discriminative correlation filter, and achieves target tracking through Fourier transform. However, due to the complexity of scene characteristics and the difficulty in fully and effectively utilizing features to represent the target, the aforementioned technical solutions still face certain challenges in achieving robust and real-time target tracking.

[0003] When the target scale changes, tracking drift is prone to occur. Danelljan et al. addressed this scale variation problem by learning a Discriminative Scale Space Tracking (DSST) filter. This algorithm constructs a 33-layer pyramid feature vector to respond to changes in target scale, but using 33 scales leads to computational complexity. Danelljan et al. proposed the Fast Discriminative Scale Space Tracking (FDSST) algorithm, which uses PCA dimensionality reduction to reduce HOG features to 18 dimensions. Compared to the DSST algorithm, fDSST further reduces the computational cost of tracking without sacrificing robustness and accuracy, but its tracking performance is poor when the target is in a fast-moving, rapidly deforming, or severely occluded environment for extended periods.

[0004] To address this issue, existing target tracking methods have proposed a series of improvements. Ma et al. proposed an effective scale estimation method by adding Average Peak-to-Correlation Energy (APCE) to the multi-resolution conversion filter framework, thus solving the problem of fixed template size in standard discriminative correlation filters. Yazd et al. proposed a rotation-aware discriminative scale space tracking method, which can effectively estimate the rotational position of the target and integrate it into the scale estimation model. Mueller et al. proposed a context-aware correlation filter tracking algorithm that incorporates the global context into the filter training stage, improving target tracking performance when the background is cluttered and the target appearance changes drastically. For example, the existing invention patent document CN110414439B, entitled "Anti-Occlusion Pedestrian Tracking Method Based on Multi-Peak Detection," first performs pedestrian detection to obtain the initial position and initializes the tracker parameters and pedestrian template. In each subsequent frame, the location of the feature fusion response peak is used as the pedestrian prediction position center, and the target response peak Fmax, average peak correlation energy (APCE), and their thresholds are calculated. The joint confidence score formed by these calculations is used to detect the multi-peak filter response, thereby realizing pedestrian occlusion judgment. In occluded frames, the updating of filter parameters and pedestrian target template is paused, achieving the anti-occlusion pedestrian tracking task. As can be seen from the prior art specification, the prior art disclosed in this document uses FHOG features and Color Naming features for adaptive fusion as feature descriptors. The algorithms disclosed in the aforementioned prior art still have many practical problems to be solved in challenging and complex scenarios. Moreover, when the tracking process encounters significant challenges such as rapid target movement or severe occlusion, a single shallow gradient feature cannot fully and effectively express the target information.

[0005] In summary, existing technologies suffer from limitations in fully and effectively utilizing features to represent targets, low robustness, and poor tracking performance in specific scenarios. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to solve the problems of existing technologies being unable to fully and effectively utilize features to represent targets, having low robustness, and poor tracking performance in specific scenarios.

[0007] This invention solves the above-mentioned technical problems by employing the following technical solution: a target tracking method that integrates co-occurrence statistics and fhog gradient features includes:

[0008] S1. Determine the initial target position based on the given tracking target area;

[0009] S2, Calculate the current frame I t The target position p in the previous frame t-1 and scale factor s t-1 Based on this, the target region can be determined and relevant features of the current frame can be extracted.

[0010] S3. Obtain co-occurrence statistics between pixel pairs in the tracked target image using a co-occurrence filter, obtain the target fhog features using the fDSST algorithm, and combine the co-occurrence statistics with the target fhog features as the target image features. Step S3 includes:

[0011] S31. Relevant features of the current frame With position filter Perform relevant calculations to obtain the target position in the current frame. Among them, the relevant features of the current frame are taken. The point with the largest response value in the current frame is used as the target center p. t ;

[0012] S32, Calculate the current frame I t The target position p in the previous frame t-1 and scale factor s t-1 Based on this, the target area is defined, and after scaling, a preset number of target image areas are obtained;

[0013] S33. Obtain co-occurrence statistics and fhog gradient features through pre-set logic processing. Step S33 includes:

[0014] S331. The co-occurrence filtering module uses the co-occurrence filter (COF) to process frequently occurring pixel data in the tracked target image;

[0015] S332. Based on the frequently occurring pixel data, assign high weights to high-frequency pixel values ​​and low weights to low-frequency pixel values ​​in the preset conformal matrix to obtain co-occurrence matrix weight data.

[0016] S333. Extract target detail features within the texture region of the tracked target image based on the co-occurrence matrix weight data to obtain co-occurrence statistics.

[0017] S34. Combine co-occurrence statistics and fhog gradient features to obtain tracking fusion data, and extract multi-scale features from the tracking fusion data. Multiscale features With difference scale filter Perform relevant calculations to obtain the current frame I. t Scale information Among them, scale information is taken. The largest response value in the current frame is used as the target scale s. t ;

[0018] S4. When the model is updated, in the current frame I t According to the target center p in the current frame t and the target scale s of the current frame t The target region is identified and its features are extracted. These features are then fed into a position filter and a scale filter, and the target tracking results are obtained through iterative updates.

[0019] This invention addresses the challenge of effectively utilizing features to fully represent target information during target tracking in complex scenes. It proposes a target tracking algorithm that integrates co-occurrence statistics and fhog features. Building upon the fDSST tracking algorithm, a pixel filtering module based on image co-occurrence statistics is introduced. This module leverages the high weight of frequently occurring pixel values ​​in the co-occurrence matrix, fusing them with the fhog features from the original tracking algorithm to enhance co-occurrence statistics based on texture similarity. By combining the fhog features from the original fDSST tracking algorithm with the pixel module features based on co-occurrence statistics, this invention fuses semantic features of the target from different perspectives, achieving a complete representation of the tracked target.

[0020] This algorithm combines the co-occurrence filter with the fhog gradient features from the original position and scale filters, enabling it to represent global feature information of the target image based on different image details, thereby improving target tracking accuracy. This invention continuously updates the co-occurrence filter model by combining the position information obtained from the position filter and the target scale obtained from the scale filter in the discriminative scale-space tracking algorithm, addressing the problem of decreased tracking rate in scenarios where the target position and scale change.

[0021] In a more specific technical solution, step S31 includes:

[0022] S311. Use a two-dimensional position filter to estimate the position of the tracked target, assuming f is composed of d-dimensional feature vectors. The image patch centered on the tracked target is used to construct the optimal filter between the target-related response of the training samples and the desired output using the following logic:

[0023]

[0024] In the formula, g, and Let be an M×N matrix, l∈{1,…,d} represent the feature dimension, ★ represent the convolution operation, and λ represent the regularization coefficient;

[0025] S312. Using the new training samples, update the filter parameters iteratively according to the following logic:

[0026]

[0027] In the formula, η represents the learning rate. and For filter The numerator and denominator of the training samples;

[0028] S313. Use Passavar's theorem to transform the linear least squares data in step S312 into the complex frequency domain for solution.

[0029] S314. Extract the image patch features z centered on the target from the image of the tracked target in a new frame, and perform a two-dimensional Fourier transform on the image patch features to obtain the transformed image feature data.

[0030] S315, Utilize preset logic to process the filter The numerator and denominator of the training samples and the transformed image feature data The maximum response value is obtained accordingly, and the position corresponding to the maximum response value is taken as the position of the target in the current frame.

[0031] This invention utilizes Passavar's theorem to transform the linear least squares problem into a solution in the complex frequency domain, thereby improving computational speed. Simultaneously, this invention uses new training samples to iteratively update the filter parameters, making the filter robust.

[0032] In a more specific technical solution, in step S313, the following logic is used to transform the linear least squares data in step S312 into the complex frequency domain for solving:

[0033]

[0034] In the formula, k∈{1,…,d} represents the range of values ​​for the feature dimension. Let g represent the conjugate transpose of G, and the discrete Fourier transform of the M×N matrix. and Therefore, the corresponding variables G and H are obtained. l F l .

[0035] In a more specific technical solution, in step S315, the following logic processing filter is used. The numerator and denominator of the training samples and The maximum response value is thus obtained:

[0036]

[0037] In a more specific technical solution, step S332 includes:

[0038] S3321. Obtain the co-occurrence information between pixel pairs in the image of the tracked target, and detect the co-occurrence filter (COF) of the edge based on the co-occurrence matrix using the following logic:

[0039]

[0040] In the formula, I q and J P Let p and q represent the input and output images, respectively, where p and q are the indices of pixels at different positions in the image, and w(p,q) represents the input and output pixel weights.

[0041] S3322. The first input / output pixel weights w(p,q) are obtained by processing them in a preset Gaussian filter using the following logic:

[0042]

[0043] In the formula, d(p,q) is the Euclidean distance between pixels p and q on the image plane, and σ s Represents the spatial domain kernel parameters, Indicates a Gaussian filter;

[0044] S3323. The spatial distance data and pixel similarity data of the tracked target image are processed using the following logic, and the second input-output pixel weights are obtained by processing them in a preset bilateral filter:

[0045]

[0046] In the formula, I p and I q σ represents the grayscale values ​​of pixels p and q, respectively. r It is a pixel-range domain kernel parameter;

[0047] S3324. Using a co-occurrence filter (COF) combined with a normalized co-occurrence matrix and a bilateral filter, the co-occurrence matrix weight data is obtained through processing.

[0048]

[0049]

[0050] In the formula, M(a,b) is a 256×256 normalized co-occurrence matrix based on the co-occurrence matrix C(a,b), and h(a) and h(b) are the frequencies of a and b in the image.

[0051] This invention maps the input tracking image to the embedded image, making the Euclidean distance between pixel values ​​in the embedded image space similar to the co-occurrence statistics in the original space, thereby obtaining better tracking results.

[0052] In a more specific technical solution, step S333 includes:

[0053] S3331. Construct a pre-defined co-occurrence matrix, and use the following logic to calculate co-occurrence information and co-occurrence matrix relationship data:

[0054]

[0055]

[0056] In the formula, σ is the Gaussian filter parameter, and [·] indicates that if the expression in the parentheses is true, the result is 1, otherwise the result is 0;

[0057] S3332. Using a hard clustering method, the pixel values ​​of the tracked target image are clustered using the k-means algorithm to generate a guide image T, where k is the number of clusters.

[0058] S3333, Let M T To obtain the co-occurrence matrix of the guiding image T, the following logical processing is used to obtain the guiding image co-occurrence matrix of T, where the size of the guiding image co-occurrence matrix becomes k×k:

[0059]

[0060] S3334. Calculate the co-occurrence statistics in the guiding image T using the following logic, in order to... Representing pixel values ​​clustered into k clusters, the co-occurrence matrix relationship data is expanded to obtain the following logical representation of the hard clustering co-occurrence matrix relationship data:

[0061]

[0062] In the formula, τ a and τb Representing two clusters, T p =a indicates that pixel p belongs to cluster τ a T q =b indicates that pixel q belongs to cluster τ b ;

[0063] S3335. Using a soft allocation strategy, the probability of pixel values ​​belonging to different clusters is specified using the following logic:

[0064]

[0065] S3336. Use a kernel function to make Pr(p∈τ)=K(I p ,τ), based on pixel value I p Calculate the probability of assigning pixel p to cluster τ by taking the distance between the cluster center τ and the original space.

[0066] S3337. Using the following logic, I p The distance between τ and τ is approximately equal to τ. p Distance between τ and τ:

[0067] Pr(p∈τ)≈K(τ p ,τ);

[0068] S3338. Utilize the following logic to process τ p The distance between τ and the hard clustering co-occurrence matrix data are used to obtain global co-occurrence statistics of the tracked target image:

[0069]

[0070]

[0071] Where Z is the normalization constant, σ r It is a specified parameter, τ p Is it away from I p The nearest cluster center.

[0072] The co-occurrence filter used in this invention does not consider the color space information in the image, but only the probability that pixel pairs in the image always appear simultaneously. Based on the characteristic that frequently occurring pixel values ​​in the image have higher weights in the co-occurrence matrix, the global statistical information of the image can be obtained to a greater extent when performing feature sampling or other processing on the image.

[0073] To address the shortcomings of traditional color feature matching techniques, which struggle to fully and effectively represent the local color distribution and color space location of images, and the difficulty of handling occlusion issues with HOG gradient direction histograms, this invention employs a CoF filter. This ensures that frequently occurring and co-occurring pixel pairs in the target image always have higher weights, regardless of changes in the target image scene. This combines the co-occurrence statistics of the target image with gradient features, thereby improving the accuracy of target tracking.

[0074] In a more specific technical solution, the size of the co-occurrence matrix of the guiding image in step S3333 becomes k×k.

[0075] In more specific technical solutions, the position filter and the difference scale filter are independent of each other.

[0076] This invention uses a standard position filter to estimate the target's position, and simultaneously uses a set of target samples at different scales to learn a scale filter, thereby obtaining an accurate estimate of the target's scale.

[0077] In a more specific technical solution, step S34 includes:

[0078] S341. Use at least two differential scaling filters and a one-dimensional Gaussian function as the desired output.

[0079] S342. Iteratively update the scaling filter model h using the following logic:

[0080]

[0081] To continuously acquire an M×N×S feature pyramid composed of S scales in a new frame image;

[0082] S343. Calculate the maximum response value using the following logic, and use the scale corresponding to the maximum response value as the target scale s of the current frame. t :

[0083]

[0084] In more specific technical solutions, target tracking systems that integrate co-occurrence statistics and fhog gradient features include:

[0085] The initial position module is used to determine the initial target position based on a given tracking target area;

[0086] The current frame feature module is used to calculate the current frame I. t The target position p in the previous frame t-1 and scale factor s t-1 Based on this, the target region can be determined and relevant features of the current frame can be extracted. The current frame feature module is connected to the initial position module;

[0087] The target image feature module is used to obtain co-occurrence statistics between pixel pairs in the tracked target image using a co-occurrence filter, obtain the target fhog features using the fDSST algorithm, and combine the co-occurrence statistics with the target fhog features to form the target image features. The target image feature module is connected to the current frame feature module. The target image feature module includes:

[0088] The current frame target location module is used to store relevant features of the current frame. With position filter Perform relevant calculations to obtain the target position in the current frame. Among them, the relevant features of the current frame are taken. The point with the largest response value in the current frame is used as the target center p. t ;

[0089] The target image region module is used to calculate the current frame I. t The target position p in the previous frame t-1 and scale factor s t-1 The target area is defined based on this, and after scaling, a preset number of target image areas are obtained. The target image area module is connected to the target position module of the current frame.

[0090] The co-occurrence statistics and FHOG gradient feature processing module is used to obtain co-occurrence statistics and FHOG gradient features through pre-defined logic processing. This module is connected to the target image region module. The co-occurrence statistics and FHOG gradient feature processing module includes:

[0091] The co-occurrence filtering module is used to process frequently occurring pixel data in the image of the tracked target using a co-occurrence filter (COF).

[0092] The weight allocation module is used to assign high weights to frequently occurring pixel values ​​and low weights to infrequently occurring pixel values ​​in a preset conformal matrix based on frequently occurring pixel data, thereby obtaining co-occurrence matrix weight data. The weight allocation module is connected to the frequent pixel processing module.

[0093] The co-occurrence statistics module is used to extract target detail features within the texture region of the tracked target image based on the co-occurrence matrix weight data, thereby obtaining co-occurrence statistics. The co-occurrence statistics module is connected to the weight allocation module.

[0094] The co-occurrence information and gradient feature fusion module is used to fuse co-occurrence statistical information and FHOG gradient features to obtain tracking fusion data and extract multi-scale features from the tracking fusion data. Multiscale features With difference scale filter Perform relevant calculations to obtain the current frame I. t Scale information Among them, scale information is taken. The largest response value in the current frame is used as the target scale s. t The co-occurrence information and gradient feature fusion module is connected to the co-occurrence information and gradient feature processing module;

[0095] The target tracking module is used to track targets in the current frame I during model updates. t According to the target center p in the current frame t and the target scale s of the current frame t The target region is determined and its features are extracted. These features are then fed into the position filter and scale filter, and the target tracking result is obtained through iterative updates. The target tracking module is connected to the target image feature module.

[0096] Compared with existing technologies, this invention has the following advantages: Addressing the problem of difficulty in fully and effectively utilizing features to comprehensively represent target information during target tracking in complex scenes, this invention proposes a target tracking algorithm that integrates co-occurrence statistics and fhog features. Based on the fDSST tracking algorithm, a pixel filtering module based on image co-occurrence statistics is introduced. This module leverages the high weight of frequently occurring pixel values ​​in the co-occurrence matrix and fuses them with the fhog features from the original tracking algorithm, adding co-occurrence statistical information based on texture similarity. This invention combines the fhog features from the original fDSST tracking algorithm with the pixel module features based on co-occurrence statistics to fuse target semantic features from different perspectives, achieving a complete representation of the tracked target.

[0097] This algorithm combines the co-occurrence filter with the fhog gradient features from the original position and scale filters, enabling it to represent global feature information of the target image based on different image details, thereby improving target tracking accuracy. This invention continuously updates the co-occurrence filter model by combining the position information obtained from the position filter and the target scale obtained from the scale filter in the discriminative scale-space tracking algorithm, addressing the problem of decreased tracking rate in scenarios where the target position and scale change.

[0098] This invention utilizes Passavar's theorem to transform the linear least squares problem into a solution in the complex frequency domain, thereby improving computational speed. Simultaneously, this invention uses new training samples to iteratively update the filter parameters, making the filter robust.

[0099] This invention maps the input tracking image to the embedded image, making the Euclidean distance between pixel values ​​in the embedded image space similar to the co-occurrence statistics in the original space, thereby obtaining better tracking results.

[0100] The co-occurrence filter used in this invention does not consider the color space information in the image, but only the probability that pixel pairs in the image always appear simultaneously. Based on the characteristic that frequently occurring pixel values ​​in the image have higher weights in the co-occurrence matrix, the global statistical information of the image can be obtained to a greater extent when performing feature sampling or other processing on the image.

[0101] To address the shortcomings of traditional color feature matching techniques, which struggle to fully and effectively represent local color distribution and each color space location, and the difficulty of handling occlusion issues with HOG gradient direction histograms, this invention employs a CoF filter. This ensures that frequently occurring and co-occurring pixel pairs in the target image always have higher weights, regardless of changes in the target image scene. This combines co-occurrence statistics with gradient features, thereby improving the accuracy of target tracking.

[0102] This invention uses a standard position filter to estimate the target's location, and simultaneously uses a set of target samples at different scales to learn a scale filter, thereby obtaining an accurate estimate of the target's scale. This invention solves the technical problems of existing technologies, such as inability to fully and effectively utilize features to represent the target, low robustness, and poor tracking performance in specific scenarios. Attached Figure Description

[0103] Figure 1 This is a schematic diagram of the feature extraction process of a multi-scale image through a scale filter in Embodiment 1 of the present invention;

[0104] Figure 2 The image consists of the feature map of the original image after fhog transformation and the co-occurrence matrix image after co-occurrence statistics in Embodiment 1 of the present invention.

[0105] Figure 3 This is a schematic diagram of the overall algorithm flow of the target tracking method that integrates co-occurrence statistics and fhog gradient features according to Embodiment 1 of the present invention;

[0106] Figure 4a The graphs show the distance accuracy curves of each algorithm in Embodiment 2 of the present invention under OPE evaluation.

[0107] Figure 4b This is a success rate curve of each algorithm in Embodiment 2 of the present invention under OPE evaluation;

[0108] Figure 5a This is a schematic diagram of the qualitative comparison of the first vehicle video sequence in Embodiment 2 of the present invention;

[0109] Figure 5b This is a schematic diagram of the qualitative comparison of the second vehicle video sequence in Embodiment 2 of the present invention;

[0110] Figure 5cThis is a schematic diagram of the qualitative comparison of the first person's video sequence in Embodiment 2 of the present invention;

[0111] Figure 5d This is a schematic diagram of the qualitative comparison of the second person's video sequence in Embodiment 2 of the present invention;

[0112] Figure 5e This is a schematic diagram of the qualitative comparison of the first item video sequence in Embodiment 2 of the present invention;

[0113] Figure 5f This is a schematic diagram illustrating the qualitative comparison of the second item video sequence in Embodiment 2 of the present invention;

[0114] Figure 5g This is a schematic diagram of the qualitative comparison of the first event video sequence in Embodiment 2 of the present invention;

[0115] Figure 5h This is a schematic diagram of the qualitative comparison of the second event video sequence in Embodiment 2 of the present invention. Detailed Implementation

[0116] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0117] Example 1

[0118] This invention proposes a target tracking algorithm that integrates co-occurrence statistics and fhog gradient features. By combining the fhog features of the original fDSST tracking algorithm with pixel module features based on co-occurrence statistics, it fuses the semantic features of the target from different perspectives to achieve a complete representation of the tracked target. Quantitative and qualitative experiments on the OTB50 and OTB100 datasets verify the robust feature learning ability and accuracy of the proposed method.

[0119] The target tracking method that integrates co-occurrence statistics and fhog gradient features provided by this invention includes:

[0120] Fast discriminative scale-space filter:

[0121] Accurately estimating changes in target scale is challenging when the tracked target moves along the camera axis or its appearance changes. The discriminative scale-space correlation filter estimates target localization and scale information by learning discriminative correlation filters and scale filters based on multi-dimensional features. This algorithm uses a standard position filter to estimate the target's location, while simultaneously learning a scale filter using a set of target samples at different scales to obtain an accurate estimate of the target's scale. The position filter and scale filter are independent of each other.

[0122] The position of a target is estimated using a two-dimensional position filter, assuming f is composed of d-dimensional eigenvectors. The image patches centered on the target are used to construct the optimal filter between the target-related response of the training samples and the desired output, as shown in the following equation:

[0123]

[0124] Among them, g, h l and f l Let be an M×N matrix, l∈{1,…,d} representing the feature dimension, ★ representing the convolution operation, and λ representing the regularization coefficient. The linear least squares problem in equation (1) is transformed into a problem in the complex frequency domain using Passavar's theorem, thereby improving the computational speed. The formula is as follows:

[0125]

[0126] In the formula, k∈{1,…,d} represents the range of values ​​for the feature dimension. Let g and h represent the conjugate transpose of G. l and f l The corresponding variables G and H are obtained by performing a Discrete Fourier Transform (DFT). l F l To make the filter robust, the filter parameters are iteratively updated using new training samples:

[0127]

[0128] Where η represents the learning rate, and B t For filter The numerator and denominator of the training samples.

[0129] Finally, the image patch feature z centered on the target in the new frame is extracted, and the extracted features are subjected to a two-dimensional Fourier transform to obtain... The maximum response value is obtained using equation (2), and the position corresponding to the maximum response value is the position of the target in the current frame.

[0130]

[0131] A one-dimensional scaling filter is used to evaluate changes in target scale. Assuming the target size in the current frame is P×R and the scaling filter size is S, for... At the center of the target, obtain a scale level of n and a size of a. n P×a n Image patches of R are used as feature descriptors of different dimensions for the scaling filter. Here, 'a' is the scaling factor, used to obtain image patches of different sizes.

[0132] like Figure 1 As shown, this illustrates the process of feature extraction from a multi-scale image using a scale filter. The scale filter employs the same design method as the location filter, but uses a one-dimensional Gaussian function as the desired output. I represents image patches at different scales centered on the target location, h is the scale filter model, and f represents the image patch features at different scales obtained through computation. Equation (1) is used to iteratively update the scale filter model h, continuously acquiring an M×N×S feature pyramid composed of S scales in a new frame image. Equation (2) is used to obtain the maximum response value, and the scale corresponding to the maximum response value is the scale information of the current frame.

[0133] Optimization algorithm:

[0134] The co-occurrence filtering module uses co-occurrence information in the image, rather than intensity information, to assign corresponding weights. Pixel values ​​that appear frequently in the image have higher weights in the co-occurrence matrix, while those that do not appear frequently have lower weights. Therefore, applying this co-occurrence idea to target tracking, for targets of interest in the image, their pixel values ​​always appear in pairs, and regardless of scene changes, co-occurring pixel pairs in the target image always have higher weights. Thus, the CoF filter exhibits excellent performance in extracting detailed features within texture regions of the tracked image.

[0135] Given the powerful feature extraction capabilities of co-occurrence filters, this invention proposes a multi-scale fusion target tracking method relying on co-occurrence matrices and FHOG gradient features. This method adds a co-occurrence filter tracking module to the discriminative scale-space tracking algorithm. The method resizes each frame of the image to a specified model size and inputs it into the co-occurrence filter tracking module to perform calculations with the original gradient features of the position filter, ultimately obtaining a feature matrix of the target image patch based on position information. Similarly, target images at different scales are fed into the co-occurrence filter tracking module and the scale filter for feature concatenation, obtaining a feature matrix of the target image patch based on scale information. This feature matrix is ​​then input into the target tracking module. Unlike conventional target image feature extraction algorithms, this algorithm combines the co-occurrence filter with the FHOG gradient features from the original position and scale filters, allowing for the representation of global feature information of the target image based on different image details, thereby improving the accuracy of target tracking. This method continuously updates the co-occurrence filter model by combining the position information obtained from the position filter in the discriminative scale-space tracking algorithm and the target scale obtained from the scale filter, addressing the problem of decreased tracking rate in scenarios where the target position and scale change.

[0136] In this embodiment, a co-occurrence filtering (COF) method is used to detect edges based on the co-occurrence matrix and the co-occurrence information between image pixel pairs.

[0137]

[0138] In the formula, I q and J P Let p and q represent the input and output images, respectively. p and q are the indices of pixels at different positions in the image, and w(p,q) represents the weight of the input pixel q in relation to the output pixel p.

[0139] In a Gaussian filter, w(p,q) is defined as:

[0140]

[0141] Where d(p,q) is the Euclidean distance between pixels p and q on the image plane, and σ s Represents the spatial domain kernel parameters, This represents a Gaussian filter.

[0142] In a bilateral filter, w(p,q) is defined as:

[0143]

[0144] In the formula, I p and I q σ represents the grayscale values ​​of pixels p and q, respectively.r These are the pixel-range domain kernel parameters. The bilateral filter considers the spatial distance relationship between pixels, and compared to the Gaussian filter, it also considers the similarity between pixels.

[0145] When the co-occurrence filter combines the normalized co-occurrence matrix with a bilateral filter, the form of w(p,q) becomes:

[0146]

[0147] in,

[0148]

[0149] In the formula, M(a,b) is a 256×256 normalized co-occurrence matrix based on the co-occurrence matrix C(a,b), and h(a) and h(b) are the frequencies of a and b in the image (i.e., histograms of pixel values).

[0150]

[0151]

[0152] Where σ is the Gaussian filter parameter, and [·] means that if the expression in the parentheses is true, the result is 1, otherwise the result is 0.

[0153] Co-occurrence filters do not consider color space information in an image; they only consider the probability that pixel pairs in the image always occur simultaneously. Based on the characteristic that frequently occurring pixel values ​​have higher weights in the co-occurrence matrix, feature sampling or other processing of the image can yield a greater degree of global statistical information.

[0154] Calculating the co-occurrence information of the color space using equation (4) requires constructing a 2563×2563 co-occurrence matrix, which makes the calculation overly complex. Using the k-means algorithm to cluster the pixel values ​​of image I to generate the co-occurrence matrix of the guiding image TT, the size becomes k×k, where k is the number of clusters. Let M... T Let be the co-occurrence matrix of image T. Then, the co-occurrence matrix operation of guided image T is as follows:

[0155]

[0156] The co-occurrence statistic in the guiding image T is calculated using a formula, assuming... Let represent the pixel values ​​clustered into k clusters. Expanding formula (4) yields:

[0157]

[0158] Where, τ a and τb Representing two clusters, T p =a indicates that pixel p belongs to cluster τ a T q =b indicates that pixel q belongs to cluster τ b Since each pixel is assigned to its nearest cluster center, this method is also known as hard clustering.

[0159] However, pixel values ​​near the original space may be mapped to two different clusters during the clustering step, causing artifacts. To address this issue, a soft assignment strategy is used to specify the probability of pixel values ​​belonging to different clusters, as follows:

[0160]

[0161] The computational cost of collecting co-occurrence statistics using soft allocation is high, so a kernel function (i.e., a Gaussian function) is used, where Pr(p∈τ)=K(I p ,τ), based on pixel value I p Given the distance between I and the cluster center τ in the original space, calculate the probability of assigning pixel p to cluster τ. Then, set I... p The distance between τ and τ is approximately equal to τ. p The distance between τ and τ is Pr(p∈τ)≈K(τ). p ,τ), the improved result is:

[0162]

[0163]

[0164] Where Z is the normalization constant, σ r It is a specified parameter, τ p Is it away from I p The nearest cluster center. The formula does not calculate all pixel values, only the distance between cluster centers.

[0165] In target tracking, feature extraction is crucial. Traditional visual feature extraction algorithms are based on color feature matching, but they struggle to fully and effectively represent the local color distribution and spatial location of each color in an image. HoG gradient orientation histograms can effectively reflect the edge information of targets in an image using gradient information and can characterize the local appearance and shape of an image by measuring the magnitude of local gradients. However, due to the nature of gradients, they are highly sensitive to noise and struggle to handle occlusion issues. CoF filters, regardless of changes in the target image scene, always give higher weights to frequently occurring and co-occurring pixel pairs in the target image. Based on this characteristic, this invention combines co-occurrence statistics in the target image with gradient features to improve the accuracy of target tracking.

[0166] like Figure 2 As shown, in this embodiment, the co-occurrence matrix M is a 256×256 matrix. According to equation (3), co-occurrence statistics are collected from the original image. White and orange often appear simultaneously in the image, and light green and dark green also often appear. Therefore, they have higher weights, and the corresponding pixel regions in the matrix have higher scores.

[0167] This invention is based on the original fDSST algorithm and delves into the feature fusion method of co-occurrence statistics and fhog gradient to improve the robustness of the algorithm in scenarios where the target undergoes drastic deformation and rapid movement.

[0168] like Figure 3 As shown, the target tracking method that integrates co-occurrence statistics and fhog gradient features provided by this invention includes the following steps:

[0169] S1. Determine the target location; In this embodiment, the initial target location is determined based on a given tracking target area;

[0170] S2, Fhog transform; in this embodiment, the current frame I t Based on the calculated target position p of the previous frame t-1 and scale factor s t-1 Determine the target region and extract relevant features of the current frame.

[0171] S3, Co-occurrence Statistics; In this embodiment, the relevant features extracted by the algorithm of the present invention are the combination of the co-occurrence filter and the fhog feature in the original algorithm as the final target image features;

[0172] S4. The position filter determines the target position; in this embodiment, the relevant features of the current frame are used. With position filter Perform correlation calculations to obtain the target position of the current frame. in The point with the largest response value is the target center p in the current frame. t ;

[0173] Similarly, current frame I t Based on the calculated target position p t-1 and scale factor s t-1 Define the target region and scale it to obtain S target image regions;

[0174] S5. Scale filter determines the target scale; in this embodiment, the multi-scale features of fhog in the fusion co-occurrence filter and the original algorithm are extracted. Multiscale features With scale filter Perform correlation calculations to obtain the scale information of the current frame. in The largest response value in the current frame is the target scale s. t ;

[0175] S6. Extract the test set based on the target position and scale of the previous frame;

[0176] S7 updates the target position; S8 updates the target scale and executes step S9, determining whether the current frame is the last frame. If yes, target tracking ends; otherwise, steps S4 to S8 are executed repeatedly. In this embodiment, when the model is updated, in the current frame I... t The position p estimated based on the above steps t and scales t Features are extracted from the identified target region and then fed into the position filter and scale filter for continuous iterative updates to obtain the final tracking result.

[0177] Example 2

[0178] Experimental Analysis

[0179] The experiment was conducted on a Windows system using the MATLAB R2016a software platform, an Intel(R) Core(TM) i7-12700H processor with a 2.30GHz processor and 16GB of RAM on a laptop. The experimental parameters were set as follows: the learning rate η of the tracking model was set to 0.025, and the scaling factor a was set to 1.02.

[0180] Experimental data and evaluation indicators

[0181] To verify the effectiveness of the method of this invention, we conducted experiments on two large benchmark datasets, OTB50 and OTB100. The experimental test videos included complex scene environments such as rapid target movement, changes in lighting, target rotation, occlusion, and scale transformation. Table 1 shows a partial description of the test video sequences:

[0182] Table 1 describes some of the test video sequences.

[0183] Table 1 Partial test video sequence description

[0184]

[0185]

[0186] In this invention, we use commonly used evaluation metrics, success rate and accuracy, to assess the algorithm's performance. The ratio between the intersection and union of the target's actual position and the tracking bounding box is called the overlap rate. When the overlap rate of each frame in the test video sequence is greater than a certain threshold (the overlap threshold is set to 0.5 in this invention), the current frame is considered to have been successfully tracked. The ratio of the total number of successfully tracked frames to the total number of frames in the video sequence is called the success rate. The pixel distance between the actual center position of the target in the video sequence and the center position of the tracking result bounding box is called the center position error. The percentage of frames where the actual center position error of the tracked target is less than a manually calibrated center position error threshold (set to 20 in this invention) out of the total number of video frames is called the accuracy map.

[0187] Quantitative analysis

[0188] To evaluate the accuracy of the algorithm of this invention, it was compared with nine other target tracking algorithms on the OTB50 dataset. An overlap threshold of 0.5 and a center error threshold of 20 pixels were used to rank the tracking algorithms.

[0189] like Figure 4a and Figure 4b As shown, the accuracy and success rate of the algorithm of this invention are 82.6% and 62.7%, respectively. Compared with the fDSST algorithm, it improves the distance accuracy and success rate by 0.7 and 0.5 percentage points, respectively. Furthermore, the improved algorithm has higher accuracy and success rate than other tracking algorithms, which proves the robustness of the tracking performance of the algorithm of this invention.

[0190] Qualitative analysis

[0191] like Figures 5a to 5h As shown, this invention selects four video sequences from the OTB100 dataset to perform a qualitative analysis and comparison of the algorithm of this invention with three tracking algorithms: fDSST, VTS, and CXT.

[0192] In the Car4 video sequence, the tracker easily loses track of the target due to its rapid movement. This invention's algorithm utilizes the co-occurrence information of the image to track the target, achieving robust tracking. In the David video sequence, the target's tracking fails due to changes in lighting. This invention's algorithm leverages the high weight of frequently occurring pixel values ​​in the co-occurrence matrix to achieve tracking stability. In the Fish video sequence, the target is prone to tracking drift due to camera shake and changes in lighting. This invention's algorithm employs a dual-filter adaptive tracking mechanism to reduce background interference and accurately locate the target. In the Soccer video sequence, the target undergoes severe deformation and occlusion during movement, and the target image contains significant background noise. Other tracking algorithms have already lost track of the target by frame 111. This invention's algorithm can robustly track the target with high accuracy.

[0193] Twelve different tracking algorithms were compared on the OTB50 dataset in terms of Mean Overlap Accuracy (Mean OP) and Mean Distance Accuracy (Mean DP), and the results are shown in Table 2. The top two tracking accuracy results are represented in columns 12-3 and 13-2, and columns 12-2 and 13-3, respectively. Experimental results show that the proposed algorithm and the fDSST algorithm are the top two algorithms in terms of tracking accuracy. The proposed algorithm outperforms fDSST by 7.1% in Mean Overlap Accuracy and by 1.1% in Mean Distance Accuracy.

[0194] Table 2. Average overlap accuracy and average distance accuracy of 12 tracking algorithms across 50 videos.

[0195] Table 2 Average overlap accuracy and average distance accuracy of 12tracking algorithms on 50 videos

[0196]

[0197] To fully explore the accuracy of the algorithm under different conditions, experiments were conducted on 41 datasets from the OTB50 dataset, and the results are shown in Table 3. The evaluation metrics were center position error, distance accuracy, and overlap accuracy. The experimental results show that the method of this invention improves tracking accuracy compared to the fDSST algorithm.

[0198] Table 3 Performance comparison between the method of the present invention and the fDSST method

[0199] Table 3 Performance comparison between this method and fdsst method

[0200]

[0201]

[0202] This invention applies the concept of co-occurrence to target tracking, proposing a target tracking method that integrates co-occurrence statistics and fhog features to fully extract texture details from the input target image. Regardless of changes in the tracking scene, pixels surrounding the target always appear together, thus enabling better extraction of image details and target localization. Algorithm validation and comparison on two large benchmark datasets, OTB50 and OTB100, show that the proposed algorithm outperforms 11 other popular tracking algorithms in average overlap accuracy and average distance accuracy. Compared to the most accurate fDSST algorithm, the proposed algorithm achieves 7.1% higher average overlap accuracy and 1.1% lower average distance accuracy, while maintaining consistent and accurate target tracking with good robustness. Therefore, the target tracking method integrating co-occurrence statistics and fhog features proposed in this invention can not only effectively learn the detailed feature information of the target image but also further improve the accuracy of target tracking.

[0203] In summary, this invention addresses the problem of insufficient and ineffective feature representation in complex scene target tracking. It proposes a target tracking algorithm that integrates co-occurrence statistics and fhog features. Based on the fDSST tracking algorithm, a pixel filtering module based on image co-occurrence statistics is introduced. This module leverages the high weight of frequently occurring pixel values ​​in the co-occurrence matrix, fusing it with the fhog features from the original tracking algorithm to add texture similarity-based co-occurrence statistics. This invention combines the fhog features from the original fDSST tracking algorithm with the pixel module features based on co-occurrence statistics to fuse semantic features of the target from different perspectives, achieving a complete representation of the tracked target.

[0204] This algorithm combines the co-occurrence filter with the fhog gradient features from the original position and scale filters, enabling it to represent global feature information of the target image based on different image details, thereby improving target tracking accuracy. This invention continuously updates the co-occurrence filter model by combining the position information obtained from the position filter and the target scale obtained from the scale filter in the discriminative scale-space tracking algorithm, addressing the problem of decreased tracking rate in scenarios where the target position and scale change.

[0205] This invention utilizes Passavar's theorem to transform the linear least squares problem into a solution in the complex frequency domain, thereby improving computational speed. Simultaneously, this invention uses new training samples to iteratively update the filter parameters, making the filter robust.

[0206] This invention maps the input tracking image to the embedded image, making the Euclidean distance between pixel values ​​in the embedded image space similar to the co-occurrence statistics in the original space, thereby obtaining better tracking results.

[0207] The co-occurrence filter used in this invention does not consider the color space information in the image, but only the probability that pixel pairs in the image always appear simultaneously. Based on the characteristic that frequently occurring pixel values ​​in the image have higher weights in the co-occurrence matrix, the global statistical information of the image can be obtained to a greater extent when performing feature sampling or other processing on the image.

[0208] To address the shortcomings of traditional color feature matching techniques, which struggle to fully and effectively represent local color distribution and each color space location, and the difficulty of handling occlusion issues with HOG gradient direction histograms, this invention employs a CoF filter. This ensures that frequently occurring and co-occurring pixel pairs in the target image always have higher weights, regardless of changes in the target image scene. This combines co-occurrence statistics with gradient features, thereby improving the accuracy of target tracking.

[0209] This invention uses a standard position filter to estimate the target's location, and simultaneously uses a set of target samples at different scales to learn a scale filter, thereby obtaining an accurate estimate of the target's scale. This invention solves the technical problems of existing technologies, such as inability to fully and effectively utilize features to represent the target, low robustness, and poor tracking performance in specific scenarios.

[0210] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target tracking method fusing co-occurrence statistics and fhog gradient features, characterized in that, The method comprises: S1, determining an initial target position according to a given tracking target region; S2, compute target position p and scale factor s for previous frame I t of current frame I t-1 and scale factor s t-1 from which to determine target region and extract features related to current frame S3, obtaining co-occurrence statistical information between pairs of pixels in a tracked target image using a co-occurrence filter, obtaining tracking target fhog features using an fDSST algorithm, and combining the co-occurrence statistical information and the tracking target fhog features to obtain target image features, wherein the step S3 comprises: S31, taking the maximum response value point of the current frame correlation features and the position filter to do correlation operation, so as to obtain the target position of the current frame wherein, taking the maximum response value point of the current frame correlation features as the target center p of the current frame t ; S32, calculating the target position p of the previous frame of the current frame I t and the scale factor s t-1 and the scale factor s t-1 , so as to define a target region, and to obtain a preset number of target image regions after scaling S33, processing the co-occurrence statistical information and fhog gradient features using a preset logic, and the step S33 comprises: S331, processing frequently occurring pixel data in the tracked target image using a co-occurrence filter COF in a co-occurrence filtering module; S332, assigning high weights to high-frequency pixel values and low weights to low-frequency pixel values in a preset conformal matrix according to the frequently occurring pixel data, and obtaining co-occurrence matrix weight data based thereon; S333, extracting the target detail features in a texture region of the tracked target image according to the co-occurrence matrix weight data, and obtaining the co-occurrence statistical information based thereon; S34, fuse the co-occurrence statistical information and the fhog gradient feature, to obtain tracking fusion data, extract multi-scale features in the tracking fusion data Fuse the multi-scale features with a difference scale filter to obtain scale information of the current frame I t wherein the maximum response value point in the scale information is taken as a current frame target scale s t ;​ S4, when the model is updated, in the current frame I t according to the current frame target center p t and the current frame target scale s t determine the target region extraction features, send the target region extraction features into the position filter and the scale filter, and update the target tracking results iteratively.

2. The target tracking method of claim 1, wherein, The step S31 comprises: S311. Estimate the position of the tracked target using a two-dimensional position filter, assuming f is composed of d-dimensional feature vectors. The image patch centered on the tracked target is used to construct an optimal filter between the target-related response of the training samples and the desired output using the following logic: where g, h l and f l is an M x N matrix, l e {1, …, d} represents the feature dimension, ★ represents the convolution operation, and l represents the regularization term coefficient; S312, using new training samples to iteratively update the parameters of the filter according to the following logic: In the formula, η represents a learning rate, and B t is a filter the numerator and denominator of the training sample; S313, converting the linear least squares data in the foregoing step S312 to the complex frequency domain for solving using the Parseval theorem; S314, extracting a new frame of target-centered image block features z from the tracked target image, two-dimensional Fourier transforming the image block features to obtain transformed image feature data S315, processing the filter by using preset logic molecular denominators of the training samples and the transformed image feature data The maximum response value is obtained, and a position corresponding to the maximum response value is taken as a position of a target of the current frame. 3.The target tracking method of claim 1, wherein, In the step S313, the linear least squares data in the step S312 is converted to the complex frequency domain for solving using the following logic: In the formula, k∈{1,…,d} represents the characteristic dimension value range, represents the conjugate transpose of G, the Fourier transform MxN matrix is discretely processed to obtain the corresponding variable G, H l , F l .

4. The method of claim 1, wherein the fusion of co-occurrence statistics and fhog gradient features for object tracking is characterized by, In the step S315, the filter is processed by using the following logic the numerator and the denominator of the training sample and from which the maximum response value is obtained:

5. The method of claim 1, wherein the fusion of co-occurrence statistics and fhog gradient features for object tracking is characterized by, The step S332 comprises: S3321, obtaining co-occurrence information between pairs of pixels in the tracked target image, and detecting edges using the co-occurrence filter COF according to the following logic based on the co-occurrence matrix: where I q and J P denote the input and output images, respectively, p and q are indices of different pixels in the images, and w(p,q) denotes the input-output pixel weight; S3322, processing first input-output pixel weights w(p,q) in a preset Gaussian filter according to the following logic: where d(p, q) is the Euclidean distance between pixels p and q on the image plane, σ s denotes a spatial domain kernel parameter, denotes a Gaussian filter; S3323, processing spatial distance data and pixel similarity data of the tracked target image according to the following logic, and processing second input-output pixel weights in a preset bilateral filter based thereon: where I p and I q represent the gray values of pixels p and q, respectively, and σ r is a pixel range domain kernel parameter; S3324, using the co-occurrence filter COF to combine the normalized co-occurrence matrix and the bilateral filter to process the co-occurrence matrix weight data: In the formula, M(a,b) is a 256x256 normalized co-occurrence matrix based on the co-occurrence matrix C(a,b), and h(a) and h(b) are the frequencies of a and b in the image.

6. The method of claim 1, wherein the fusion of co-occurrence statistics and fhog gradient features for object tracking is characterized by, The step S333 comprises: S3331, constructing the co-occurrence matrix of a preset specification, and calculating co-occurrence information and co-occurrence matrix relationship data using the following logic: In the formula, σ is a Gaussian filtering parameter, and [·] represents 1 if the expression in the brackets is true, and 0 otherwise; S3332, using a hard clustering method, clustering processing pixel values of the tracked target image using a k-means algorithm to generate a guide image T, wherein k is the number of clusters; S3333、set M T is the co-occurrence matrix of the guide image T, the guide image co-occurrence matrix of the guide image T is obtained by the following logic processing, wherein the size of the guide image co-occurrence matrix becomes k x k: S3334, compute co-occurrence statistics in the guidance image T using the logic representing pixel values clustered into k clusters, the co-occurrence matrix relationship data is extended to hard clustering co-occurrence matrix relationship data represented by the following logic: where τ a and τ b denote two clusters, T p = a denotes that pixel p belongs to cluster τ a , T q = b denotes that pixel q belongs to cluster τ b ; S3335, using the following logic to specify the probability that the pixel values belong to different clusters through a soft assignment strategy: S3336, using a kernel function, Pr(p e T) = K(I p , T), compute the probability of assigning pixel p to cluster T as a function of the distance between the pixel value I p and the cluster center T in the original space; S3337, I is set to 0 using the logic p The distance between τ and τ is approximated by τ p The distance between τ and τ is approximated by τ Pr(p e T) ~ K(T p , T); S3338, processing the τ p and the distance between τ and the hard clustering co-occurrence matrix relationship data, so as to obtain the co-occurrence statistical information of the tracked target image globally: where Z is a normalization constant, σ r is a specified parameter, τ p is the distance from I p the most recent cluster center.

7. The target tracking method of claim 6, wherein, The size of the guide image co-occurrence matrix in the step S3333 becomes kxk.

8. The method of claim 1, wherein the fusion of co-occurrence statistics and fhog gradient features for object tracking is characterized by, The position filter and the difference scale filter are independent of each other.

9. The method of claim 1, wherein the fusion of co-occurrence statistics and fhog gradient features for object tracking is characterized by, The step S34 comprises: S341, using a one-dimensional Gaussian function as the expected output by using not less than two difference scale filters; S342, iteratively updating the scale filter model h using the following logic: to continuously obtain an MxNxs feature pyramid composed of S scales in a new frame of image; S343, find the maximum response value by using the following logic, and use the scale corresponding to the maximum response value as the target scale s of the current frame t :

10. A target tracking system fusing co-occurrence statistics with fhog gradient features, characterized in that, The system comprises: An initial position module configured to determine an initial target position according to a given tracking target region; a current frame feature module for calculating a target position p t and a scale factor s t-1 of a previous frame of the current frame I t-1 from which a target region is determined and a current frame related feature is extracted the current frame feature module is connected with the initial position module; A target image feature module configured to obtain co-occurrence statistical information between pixel pairs in a tracked target image by using a co-occurrence filter, to obtain tracking target fhog features by using an fDSST algorithm, and to combine the co-occurrence statistical information and the tracking target fhog features as target image features, wherein the target image feature module is connected to the current frame feature module, and wherein the target image feature module comprises: a current frame target position module for determining a target position of the current frame based on the current frame related features and the position filter to obtain the target position of the current frame wherein the maximum response value point of the current frame related features is taken as the target center p t of the current frame a target image region module, configured to calculate the target position p t of the previous frame of the current frame I t-1 and the scale factor s t-1 , so as to define a target region, and scale to obtain a preset number of target image regions, the target image region module being connected with the current frame target position module; A co-occurrence statistical information and fhog gradient feature processing module configured to process the co-occurrence statistical information and the fhog gradient features by using a preset logic, wherein the co-occurrence statistical information and fhog gradient feature processing module is connected to the target image region module, and wherein the co-occurrence statistical information and fhog gradient feature processing module comprises: A co-occurrence filter module configured to process frequently occurring pixel data in the tracked target image by using a co-occurrence filter COF; A weight distribution module configured to distribute high weights to high-frequency pixel values and low weights to low-frequency pixel values in a preset conformal matrix according to the frequently occurring pixel data, to obtain co-occurrence matrix weight data, wherein the weight distribution module is connected to the frequently occurring pixel processing module; A co-occurrence statistical information module configured to extract the target detail features in a texture region of the tracked target image according to the co-occurrence matrix weight data, to obtain the co-occurrence statistical information, wherein the co-occurrence statistical information module is connected to the weight distribution module. The co-occurrence information and gradient feature fusion module is used for fusing the co-occurrence statistical information and the fhog gradient feature, so as to obtain tracking fusion data, and extracting multi-scale features in the tracking fusion data The multi-scale features are filtered by a difference scale filter to obtain scale information of the current frame I t The scale information is obtained by correlation operation , and the maximum response value point in the scale information is taken as a target scale s of the current frame t The co-occurrence information and gradient feature fusion module is connected with the co-occurrence information and gradient feature processing module. a target tracking module, configured to determine a target region extraction feature according to a target center p t and a target scale s t of the current frame I t at the time of model updating, and send the target region extraction feature into the position filter and the scale filter, so as to iteratively update a target tracking result, wherein the target tracking module is connected with the target image feature module.

Images