CN103632382A - Compressive sensing-based real-time multi-scale target tracking method - Google Patents

Abstract

The invention discloses a real-time multi-scale target tracking method based on compressed sensing. The sample is modeled by extracting the normalized rectangular feature of the sampled image, which is robust to multi-scale target tracking. Due to the high dimensionality of normalized rectangular features, the invention compresses high-dimensional features based on compressed sensing, extracts scale-invariant compressed feature vectors and uses integral maps to greatly reduce computational complexity to meet the needs of real-time tracking. Use the naive Bayesian classifier to classify the compressed feature vector of the sample to determine the most likely position of the target, and use the classifier response to estimate the weight of the particles, resample the particles to prevent the degradation of the particle tracking ability, and consider the speed of the target using two factors First-order models estimate and predict particle states. The method of the invention can track the target in the video image in real time, has high accuracy, low computational complexity, and the tracking frame changes with the scale of the target in real time, meeting the needs of actual tracking applications.

Description

A real-time multi-scale target tracking method based on compressed sensing

technical field

The invention relates to a real-time multi-scale target tracking method based on compressed sensing, which belongs to the technical field of computer vision.

Background technique

Video image moving target tracking is the most important topic in computer vision, and it has a very wide range of applications in the fields of target supervision, motion detection and recognition, and medical images. The task of tracking is the process of estimating the object state information in subsequent video frames under the condition that the object state in the initial video frame is known. Video image moving target tracking is usually described as a dynamic state estimation problem. According to different applications, the state information of the target is generally the kinematic characteristics of the target, such as position coordinates, target scale, etc. Although researchers at home and abroad have proposed various solutions to the problem of video image target tracking after years of research, there are many factors that affect tracking, such as target attitude adjustment, changes in the appearance of the target due to changes in illumination, target occlusion, and target scale. Real-time and reliable video target tracking still faces many challenges.

Generally speaking, video object tracking methods can be divided into two categories: statistical tracking methods and deterministic tracking methods. Deterministic tracking methods obtain the object state location in the current video frame by maximizing the similarity measure between the object model and candidate samples. The most commonly used method to measure the similarity between the target model and the current image sample is the difference sum of squares (SSD). Although this measurement method is simple, it is often less robust. For this reason, various improvement methods have been proposed, such as based on Mean shift algorithm and optimization algorithm to find the best candidate samples. Although the robustness has been greatly improved after the improvement, various iterative search methods consume a lot of computer resources, and the real-time performance cannot be guaranteed. Furthermore, in deterministic tracking methods, there are various methods to model the target, such as point-based models, contour-based models, and kernel-density-space-based models, etc. Point-based object models such as using scale-invariant feature points (SIFT) to model objects, and then using these scale-invariant key feature points to match the original object and the target area in the current frame, while using the RANSAC method to eliminate mismatched feature points , and finally use affine transformation or perspective transformation to realize the mapping between the original target area and the current frame candidate area; the contour-based target model usually needs to model the target contour under offline conditions, and use a large number of samples to train the contour features of the target. In the online tracking and detection stage, the moving target is approached in an adaptive way and the target position state is obtained. It can be seen that the biggest problem of the deterministic tracking method is that the calculation load is relatively large, which is not conducive to real-time applications. In addition, the deterministic tracking method relies heavily on the selection of features in the process of building the target model. In order to have high robustness to factors such as occlusion, scale changes, and illumination changes, the selected features must be strictly required and designed, which will inevitably lead to The amount of calculation is greatly increased.

Statistical tracking methods have received more and more attention in recent years. Although deterministic tracking methods continue to have new results in recent years, these solutions have not fundamentally solved the real-time problem. The statistical tracking method uses the state space and the measurement space to describe the dynamic change process of the entire tracking system target, and the estimation of the state is completed by finding the peak value of the posterior probability of the state under the corresponding measurement conditions. Under the linear Gaussian model, the mean value and variance of the corresponding probability distribution are continuously updated by using the Kalman filter, and the optimal value of the corresponding estimate can be obtained. For tracking problems under nonlinear or non-Gaussian conditions, it is impossible to obtain the analytical optimal value of the state posterior probability distribution. For this reason, many algorithms such as particle filter and extended Kalman filter have been proposed to obtain an approximate estimate of the distribution. Particle filter is the most typical scheme in the statistical tracking method. By continuously transferring and predicting the sampled particles and obtaining the corresponding feature measurement of the sample to update the sample weight, the sample approximation is used to estimate the posterior probability density of the state space. Particle filtering uses a recursive method to continuously update and propagate particles, so the computational load of the algorithm is low during the prediction and update phases. After years of development, a variety of tracking methods based on particle filtering have emerged. These methods usually use contour or color features when observing samples, such as particle filtering based on color histograms, particle filtering methods based on scale-invariant features, and based on Particle filter methods for cascaded features, etc. Although the color histogram is used as an observation model, it is robust to noise, but when the illumination changes significantly or the background interferes strongly with the target, the reliability of the system will often be greatly reduced. In addition, the calculation of the color histogram is related to the target size. As the target size increases, the calculation amount will also increase. The same problem exists in the method using contour and scale-invariant SIFT feature points. This limits the number of sampling particles in the particle filter. When the number of particles is small, the accuracy of the approximate estimation of the posterior probability density will inevitably decrease.

It can be seen that whether it is a deterministic tracking method or a statistical tracking method, the selection of the target appearance model is a very important part of the tracking task, which will directly affect the real-time performance, robustness and adaptability to various factors of the tracking algorithm. . In the field of video tracking in recent years, researchers at home and abroad have done a lot of research on appearance models. Generally speaking, it can be divided into two categories: generative models and discriminative models. The generative model first learns the appearance features of the target, and then uses the learned target appearance features to search for relevant image regions, and obtains the position of the target in subsequent video frames according to the minimum error criterion. In order to adapt to the interference of various factors in the tracking process, it is difficult and rigorous to construct a robust and efficient appearance model. This will also lead to a substantial increase in computational complexity. The more typical ones are appearance models based on sparse representation, appearance models based on orthogonal matching pursuit, and incremental learning methods. The biggest problem with these existing generative appearance models is that the learning of appearance features requires a large number of training samples. In order to reduce the computational complexity, they can only learn offline and assume that the target appearance is unchanged throughout the tracking process. Moreover, the generative model cannot make full use of the background information near the target, which is often beneficial to improve the tracking effect. Discriminative models treat the tracking problem as a binary classification problem, and the main idea is to separate the object from the background. At present, typical discriminative tracking models such as tracking using support vector machine classifier, online boosting tracking algorithm, semi-supervised online boosting tracking algorithm, multiple instance learning (MIL) tracking algorithm, compressed tracking (CT) algorithm, etc. Compressed tracking algorithm has been widely concerned because of its high real-time and reliability, but the method has the following problems, which limit its practicability. First of all, the compressed tracking method uses a fixed-size tracking frame to detect and identify samples. The tracking frame does not change with the change of the target scale, and cannot adapt to the impact of multi-scale changes of the target on the tracking effect. In many practical applications, algorithms are required to have multi-scale tracking properties. As a discriminative tracking method, the tracking process continuously classifies the target and the background, and the classifier parameters need to be updated continuously with the tracking process. Due to the fixed tracking scale, although the algorithm can track the target area under certain conditions, the actual When the target scale changes, the algorithm has considered the target and the background and the combination (the target scale is smaller than the initial scale) or a part of the target (the target scale is larger than the initial scale) as a new target for tracking. Once the target scale has a sudden change, the classifier does not have enough time to learn the changed target features, which will greatly increase the possibility of target loss. Second, the current various discriminative tracking methods often use the time correlation of the target position when collecting samples, select within a fixed radius area, and do not consider the speed and acceleration information of the target movement, and adapt to the fast target moving factor Sex is poor. Third, in the current discriminative tracking methods, the learning parameter values of the classifier are fixed. When the target is covered for a long time, the classifier will inevitably mistake the cover for the target and cause the target to be lost.

Contents of the invention

In order to overcome the shortcomings of existing video tracking methods, the present invention proposes a real-time multi-scale single-target tracking method based on compressed sensing. The method has strong real-time performance, can adapt to changes in target scales, and has high robustness in tracking results.

The real-time multi-scale target tracking method of the present invention is realized by the following technical solutions, including the system initialization phase and the video real-time target tracking phase, and the specific steps are as follows:

System initialization phase:

1. Read the target initial position parameter R _state = [x, y, w, h], where (x, y) represents the coordinates of the upper left corner of the target initial position rectangle, w and h represent the target initial position rectangle width and high;

2. Read the first frame image F ₀ of the video sequence, and convert it into a grayscale image, denoted as I ₀ ;

3. Calculate the integral image of the first frame image;

4. Construct an initial random measurement matrix R ₀ ;

5. Based on the center of the initial target position, collect positive and negative samples with the same width and height as the initial target;

6. Extract the scale-invariant compressed feature vectors of all collected positive and negative samples, and update the parameters of the naive Bayesian classifier;

7. Calculate the scale-invariant compressed feature vector v ₀ of the sample in the initial target rectangular frame, and initialize the particle distribution:

Video real-time target tracking stage:

1. Read the t-th frame image and convert it into a grayscale image, denoted as I _t ;

2. Calculate the current image integral map;

3. Use the second-order model to estimate and predict the particle state;

4. Calculate the scale-invariant compressed eigenvectors of all particles;

5. Use the naive Bayesian classifier to classify all particles, and obtain the classifier response values of all particles;

6. The particle with the largest classifier response is used as the target position, and the corresponding particle scale is used as the estimated value of the current scale of the target, and then the particle weight is calculated using the classifier response value;

7. According to the particle weight, resample all particles. After resampling, the number of samples at the particle with high weight increases, and the particle with too low weight is discarded to avoid degradation caused by low weight particle samples;

8. Based on the currently determined target center, collect positive and negative samples with the same width and height as the determined target frame;

9. Extract the scale-invariant compressed feature vectors of the collected positive and negative samples, and update the classifier parameters;

10. If the video is not over, return to step 1 and continue to read the next frame of video image

表示宽、高分别为w、h的样本图像。本发明利用样本图像的一系列归一化矩形特征作为表示样本特征的原始高维特征，这些高维特征可以通过样本图像z与一系列的归一化矩形滤波器

进行卷积得到：suppose

Indicates a sample image whose width and height are w and h respectively. The present invention uses a series of normalized rectangular features of the sample image as the original high-dimensional features representing the sample features, and these high-dimensional features can pass through the sample image z and a series of normalized rectangular filters

Perform convolution to get:

where i and j denote the width and height of the normalized rectangular filter, respectively. The original high-dimensional feature vector of the sample image representing the feature of the sample is obtained by expressing the filtered sample images obtained by filtering with all rectangular filters as column vectors and connecting the obtained column vectors with each other. For a sample image whose width and height are w and h respectively, the dimension of its original normalized rectangular high-dimensional feature vector is about (wh) ² .

Obviously, calculating the original high-dimensional feature vectors of all sample images requires a lot of computer resources, which cannot meet real-time applications. Next, the compressed sensing principle is used to obtain the low-dimensional feature vector of the compressed sample image.

表示上述样本原始高维特征向量，

表示随机测量矩阵，用于将高维特征映射至低维特征。那么可以得到压缩后的低维特征向量v=Rx，其中

是低维压缩特征向量。本发明利用如下稀疏随机测量矩阵来实现对原始高维特征的压缩。suppose

Represents the original high-dimensional feature vector of the above sample,

Represents a random measurement matrix for mapping high-dimensional features to low-dimensional features. Then the compressed low-dimensional feature vector v=Rx can be obtained, where

is a low-dimensional compressed feature vector. The present invention uses the following sparse random measurement matrix to realize the compression of original high-dimensional features.

的每一行非零元个数最多不超过4个。Among them, s=n/4, the weight w _i is the root of the reciprocal of the number of non-zero elements in the i-th row. It can be seen that the matrix

The maximum number of non-zero elements in each row of is 4.

The specific steps of constructing the initial random measurement matrix are as follows:

(1) w and h represent the width and height of the rectangular frame of the initial position of the target, respectively, and the number of low-dimensional features is m. Set the upper and lower limits of the number of non-zero elements in each row of the initial random measurement matrix to Num _max and Num _min respectively, and set Num _max and Num _min to 4 and 2 respectively according to the above sparse matrix.

(2) Loop to build each row of the initial random measurement matrix:

A. Use a uniform random number generator to generate a random integer in the interval [Num _min , Num _max ] as the number nz _i of non-zero elements in the i-th row of the initial random measurement matrix;

表示随机测量矩阵当前行权重，sign_t表示第t个非零元素符号并通过随机数发生器以等概率随机产生。B. Randomly generate nz _i rectangular areas such that px(i,t)～U(1,w-3), py(i,t)～U(1,h-3), pw(i,t)～ U(1, w-px-2), ph(i, t) ~ U(1, h-py-2), where px(i, t), py(i, t), pw(i, t) , ph(i, t) respectively represent the abscissa of the upper left corner, the ordinate of the upper left corner, width and height of the generated t-th rectangular area, U(a, b) represents the uniform distribution of integers in the interval [a, b], It is used to generate a random integer in the interval, and generate a random measurement matrix value p _value (i, t)=w _i sign _t corresponding to the corresponding rectangular area, where

Indicates the weight of the current row of the random measurement matrix, and sign _t indicates the sign of the tth non-zero element and is randomly generated by the random number generator with equal probability.

Finally obtain and store the values of all non-zero elements of the initial random measurement matrix, the non-zero elements correspond to the abscissa of the upper left corner of the rectangular area, the ordinate of the upper left corner, width and height information, and the set of the number of non-zero elements in each row {nz _i | i=1,2,...,m}.

The specific calculation steps of the sample scale-invariant compressed feature vector extraction are as follows:

Assume that w and h represent the width and height of the rectangular frame of the initial position of the target, respectively, and consider that the initial target size is 1. The integral image of the current video frame image is iH, the width and height of the sampled image are w _s and h _s respectively, the coordinates of the upper left corner of the sampled image are (x, y), and the scale of the sampled image is s, where

(1) According to the scale s of the current sample image, the non-zero elements of the stored initial random measurement matrix R ₀ are adaptively corrected to obtain the random measurement matrix R _s at the scale s, as follows:

Keep the values of all non-zero elements in the initial measurement matrix unchanged, and the parameters of the rectangle corresponding to the non-zero elements px(i, t), py(i, t), pw(i, t), ph(i, t) correspond to Change to the original s times and round up according to the rounding, that is:

Where i=1,...,m, t=1,...,nzi, m is the dimension of the compressed feature vector;

(2) The formula for calculating the compressed eigenvalue of the i-th dimension is as follows:

Among them, P _sum (i, t) represents the sum of pixel values in the rectangle corresponding to the tth non-zero element in row i of the modified random measurement matrix, which can be calculated as follows by using the integral graph:

P _sum (i, t)=iH(maxI, maxJ)-iH(maxI, minJ)-iH(minI, maxJ)+iH(minI, minJ)

表示积分图(u，v)点的值；in

Represents the value of the integral graph (u, v) point;

(3) Finally, the scale-invariant compressed feature vector v=(v _i |i=1, 2, . . . , m} of the sampled image is obtained.

The beneficial effect of the present invention is that the normalized rectangular feature is used as the original high-dimensional feature to describe the target model, so that the feature has strong adaptability to the change of the target scale, can accurately track the position of the target and the change of the target scale, and the target tracking accuracy Improvement; the principle of compressed sensing can be used to effectively model the target only by randomly extracting a small amount of original high-dimensional features, and the integral graph method can be used to quickly calculate the normalized rectangular features, so that the method of the present invention has low computational complexity and can be real-time track target.

Description of drawings

In order to illustrate the embodiments and technical solutions of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the technical solutions.

Figure 1 is a flowchart of a real-time multi-scale target tracking method based on compressed sensing;

Fig. 2 is a schematic diagram of an initial random measurement matrix;

Fig. 3 is a schematic diagram of the sample scale invariant compression feature vector extraction process of the present invention;

Fig. 4 is a schematic diagram of positive sample and negative sample collection area;

Fig. 5 is a schematic diagram of particle state estimation and prediction;

In the figure, 1. Video image area; 2. Target area; 3. Positive sample center point collection area; 4. Negative sample center point collection area.

Detailed ways

The specific implementation of the present invention will be described in detail below with reference to the accompanying drawings. Firstly, the basic flow of the real-time multi-scale single-target tracking method based on compressed sensing will be described. Referring to Figure 1, the specific steps are as follows. The whole process is divided into the system initialization phase and the video real-time target tracking phase:

System initialization phase:

1. Read the target initial position parameter R _state = [x, y, w, h], where (x, y) represents the coordinates of the upper left corner of the target initial position rectangle, w and h represent the target initial position rectangle width and high;

2. Read the first frame image F ₀ ={F _R , F _G , F _B } of the video sequence, and convert it into a grayscale image, denoted as T ₀ ;

The formula for converting a color image to a grayscale image is:

I ₀ (x, y) = 0.299F _R (x, y) + 0.587F _G (x, y) + 0.114F _B (x, y)

I ₀ (x, y) represents the gray value of the grayscale image I ₀ at the point (x, y), and the range of the gray value is [0, 255], where 0 represents black, 255 represents white, F _R , F _G , F _B are the R, G, and B components of the original image, respectively.

3. Calculate the integral map of the first frame image, the calculation formula of the integral map is as follows:

$\mathrm{<span\; class="notranslate">\; i\; H</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; x</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; the\; y</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>$

4. Construct an initial random measurement matrix R ₀ ;

5. Based on the center of the initial target position, collect positive and negative samples with the same width and height as the initial target

Referring to Fig. 4, positive sample center collection radius is set in interval (0.inrad), negative sample center collection radius is set in interval (inrad+4, outrad), in the embodiment of the present invention, inrad and outrad are respectively set to 4 and 30, and Always stay the same. Due to the change of the target scale, the sampling radius of positive and negative samples can also be set proportional to the scale: inrad=4s, outrad=30s, where s represents the current scale that needs to collect positive and negative samples. It should be noted that the determination of the inrad and outrad parameters is related to the specific video image resolution and target size. The parameter values set in this embodiment can achieve ideal tracking effects for targets of different sizes under common video resolutions. In the above sampling interval, 45 positive samples and 50 negative samples were collected through uniform random sampling;

6. Extract the scale-invariant compressed feature vector {v(n)|n=1, 2, ..., n _pos } of all collected positive samples and the scale-invariant compressed feature vector { w(t)|t=1, 2, ..., n _neg }, and use the following formula to update the classifier parameters:

$\begin{array}{c}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&LeftArrow;</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}\\ {\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&LeftArrow;</span>\sqrt{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}\\ {\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; \&LeftArrow;</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}\\ {\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; \&LeftArrow;</span>\sqrt{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\end{array}$

表示压缩特征向量第i维特征值对应的朴素贝叶斯分类器参数， ${\mathrm{\&mu;}}^1=\frac{1}{n_{\mathrm{pos}}}\underset{k=1}{\overset{n_{\mathrm{pos}}}{\mathrm{\&Sigma;}}}{v}_i\left(k\right),{\mathrm{\&sigma;}}^1=\sqrt{\frac{1}{n_{\mathrm{pos}}}\underset{k=1}{\overset{n_{\mathrm{pos}}}{\mathrm{\&Sigma;}}}{(v_i\left(k\right)-{\mathrm{\&mu;}}^1)}^2,}{\mathrm{\&mu;}}^0=\frac{1}{n_{\mathrm{neg}}}\underset{k=1}{\overset{n_{\mathrm{neg}}}{\mathrm{\&Sigma;}}}{w}_i\left(k\right),{\mathrm{\&sigma;}}^0=\sqrt{\frac{1}{n_{\mathrm{neg}}}\underset{k=1}{\overset{n_{\mathrm{nrg}}}{\mathrm{\&Sigma;}}}{(w_i\left(k\right)-{\mathrm{\&mu;}}^0)}^2};$ v_i(k)、w_i(k)分别表示第k个正、负样本压缩特征向量的第i维特征值，初始条件下所有分类器参数

均为0，参数

均为1；Wherein λ is the classifier learning rate parameter, the smaller the λ, the faster the update speed of the classifier parameters, in this embodiment, λ is 0.9, n _pos =45, n _neg =50, i=1, 2,..., m, m represents the dimensionality of the scale-invariant compressed feature vector,

Indicates the parameters of the Naive Bayesian classifier corresponding to the i-th dimension eigenvalue of the compressed eigenvector, ${\mathrm{\&mu;}}^1=\frac{1}{{\mathrm{no}}_{\mathrm{pos}}}\underset{k=1}{\overset{{\mathrm{no}}_{\mathrm{pos}}}{\mathrm{\&Sigma;}}}{v}_i\left(k\right),{\mathrm{\&sigma;}}^1=\sqrt{\frac{1}{{\mathrm{no}}_{\mathrm{pos}}}\underset{k=1}{\overset{{\mathrm{no}}_{\mathrm{pos}}}{\mathrm{\&Sigma;}}}{(v_i\left(k\right)-{\mathrm{\&mu;}}^1)}^2,}{\mathrm{\&mu;}}^0=\frac{1}{{\mathrm{no}}_{\mathrm{neg}}}\underset{k=1}{\overset{{\mathrm{no}}_{\mathrm{neg}}}{\mathrm{\&Sigma;}}}{w}_i\left(k\right),{\mathrm{\&sigma;}}^0=\sqrt{\frac{1}{{\mathrm{no}}_{\mathrm{neg}}}\underset{k=1}{\overset{{\mathrm{no}}_{\mathrm{nrg}}}{\mathrm{\&Sigma;}}}{(w_i\left(k\right)-{\mathrm{\&mu;}}^0)}^2};$ v _i (k) and w _i (k) represent the i-th dimension eigenvalues of the k-th positive and negative sample compressed feature vectors respectively, and all classifier parameters under the initial conditions

are all 0, parameter

Both are 1;

7. Calculate the scale-invariant compressed feature vector v ₀ of the sample in the initial target rectangular frame, and initialize the particle distribution:

In the embodiment of the present invention, the number of particles pn is set to 200, and each particle contains the following parameters: initial state parameters (x ₀ , y ₀ , s ₀ ), state parameters (x _t , y _t , z _t ) at the current time t, t State parameter at time -1 (x _t-1 , y _t-1 , s _t-1 ), state parameter at time t-2 (X _t-2 , y _t-2 , s _t-2 ), where x _t , y _t represents the position coordinates of the center point of the particle sample area at time t, s _t represents the scale value of the particle sample area at time t; it also includes parameters v and w, v represents the scale-invariant compressed feature vector of the particle area at the current time, and w represents the particle weight;

In the process of particle distribution initialization, all particle initial state parameters, current moment state parameters, t-1 moment state parameters, and t-2 moment state parameters are all set as (x+floor(w/2), y+floor(h/2 ), 1), the weights of each particle are initialized to 0, and the scale-invariant compressed feature vector parameters of each particle are initialized to v ₀ .

Video real-time target tracking stage:

1. Read the t-th frame image and convert it into a grayscale image, denoted as I _t ;

2. Calculate the current image integral map;

3. Use the second-order model to estimate and predict the particle state;

For each particle, the state information of particles at time t-1 and t-2 and the second-order model are used to estimate the particle state, the formula is as follows:

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; wx</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; wy</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; ws</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}\end{array}\right.$

Among them, i=1, 2,..., pn, wx _t , wy _t , ws _t represent the zero-mean Gaussian noise on the three state components respectively, and their standard deviations are std _x , std _y , std _s respectively, this In the embodiment of the invention, the standard deviation is set to std _x = 5, std _y = 2.5, std _s = 0.06; in order to prevent the particle sample from exceeding the image range, it is necessary to perform out-of-boundary processing on the estimated state, and use the state after estimation and out-of-boundary processing (x _t (i), y _t (i), s _t (i)) replaces the state parameters of the particle at time t, and replaces the state parameters of the original particle at time t and t-1 with the corresponding particle at time t-1 and t-2 respectively parameter. Referring to Figure 5, using the second-order model to estimate and predict the particle state actually takes into account the velocity information of the target's motion. Figure 5 intuitively shows that the new particle position calculated by the above-mentioned second-order model tends to the direction of the target's motion, rather than the new particle position. The particle sampling point is limited to the vicinity of the sample point at the previous moment. Similarly, the particle-scale state estimation also considers the particle-scale state at the previous two moments, and estimates the new particle-scale state according to its change trend;

4. Calculate the scale-invariant compressed eigenvectors of all particles;

5. Use the naive Bayesian classifier to classify all particles, and obtain the classifier response value {H(v _t )|t=1, 2, ..., pn} of all particles, the formula is as follows:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; log</span>}<span\; class="notranslate">\; (</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Pi;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&Pi;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; log</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span>$

表示第t个粒子的压缩特征向量v_t的第i维特征值，y=1表示样本属于目标，y=0表示样本属于背景，并假设p(y=1)=p(y=0)。由于高维随机向量的低维随机投影都是服从高斯分布，因此低维压缩特征向量的每一维特征均服从如下正太分布 $p\left(v_t^i\right|y=1)~N({\mathrm{\&mu;}}_i^1,{\mathrm{\&sigma;}}_i^1),p\left(v_t^i\right|y=0)~N({\mathrm{\&mu;}}_i^0,{\mathrm{\&sigma;}}_i^0),$ 分类器参数

通过不断采集得到的正负样本更新得到；Where m represents the low-dimensional compressed feature vector dimension,

Indicates the i-th dimension eigenvalue of the compressed eigenvector v _t of the t-th particle, y=1 means the sample belongs to the target, y=0 means the sample belongs to the background, and assume p(y=1)=p(y=0). Since the low-dimensional random projections of high-dimensional random vectors are subject to Gaussian distribution, each dimension feature of the low-dimensional compressed feature vector is subject to the following normal distribution $p\left(v_t^i\right|\mathrm{the\; y}=1)~N({\mathrm{\&mu;}}_i^1,{\mathrm{\&sigma;}}_i^1),p\left(v_t^i\right|\mathrm{the\; y}=0)~N({\mathrm{\&mu;}}_i^0,{\mathrm{\&sigma;}}_i^0),$ Classifier parameters

It is obtained by updating the positive and negative samples obtained through continuous collection;

6. Use the particle with the largest response of the classifier as the target position, and the corresponding particle scale as the estimated value of the current scale of the target, and then use the classifier response value to calculate the particle weight. The weight calculation formula is as follows:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; pn</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$

表示第i个粒子在t时刻状态为

条件下观测到t时刻目标z_t的概率， $p\left(z_t\right|x_t^{*}\left(i\right))\&Proportional;\exp \left(H\left(v_i\right)\right)=\underset{j=1}{\overset{m}{\mathrm{\&Pi;}}}\frac{p\left(v_j^j\right|y=1)}{p\left(v_i^j\right|y=0)};$ in

Indicates that the state of the i-th particle at time t is

The probability of observing the target z _t at time t under the condition, $p\left(z_t\right|x_t^{*}\left(i\right))\&Proportional;\exp \left(h\left(v_i\right)\right)=\underset{j=1}{\overset{m}{\mathrm{\&Pi;}}}\frac{p\left(v_j^j\right|\mathrm{the\; y}=1)}{p\left(v_i^j\right|\mathrm{the\; y}=0)};$

7. Resample all particles according to the formula p{newp=p(i)}=w _t , so that the probability that particle newp in the new particle set after resampling is the i-th particle p(i) before resampling is equal to its weight w _i . Assuming that the set of particles before resampling is {p(i)|i=1, 2, ..., pn}, the specific method is as follows:

(1) Sort all pn particles according to their weights from large to small, and get a new particle set {p′(i)|i=1, 2, ..., pn}, and the corresponding particle weight after sorting is

b.计算以第i个排序后粒子为基础，重采样所需复制衍生出的粒子个数c.复制得到n_i个重采样后的新粒子。(2) Perform the following steps in a loop until all pn resampled particles are obtained: a. Read the i-th sorted particle weight

b. Calculate the number of particles derived from the replication required for resampling based on the i-th sorted particle c. Copy to obtain n _i new particles after resampling.

After resampling, the number of samples at the particles with high weight is increased, and the particles with too low weight are discarded to avoid the degradation caused by low weight particle samples.

8. Based on the currently determined target center, collect positive and negative samples with the same width and height as the determined target frame;

9. Extract the scale-invariant compressed feature vectors of the collected positive and negative samples, and update the classifier parameters;

10. If the video is not over, return to step 1 and continue to read the next frame of video image.

With reference to Fig. 2, the concrete steps of described construction initial random measurement matrix are as follows:

(1) w and h respectively represent the width and height of the rectangular frame of the initial position of the target, and the number of low-dimensional features is m (m=150 in this embodiment). Set the upper and lower limits of the number of non-zero elements in each row of the initial random measurement matrix to Num _max and Num _min respectively (Num _max = 4, Num _min = 2 in this embodiment).

(2) Circularly construct each row of the initial random measurement matrix:

A. Use a uniform random number generator to generate random integers in the interval [Num _min , Num _max ] as the number nz _i of non-zero elements in the current i-th row of the matrix;

表示随机测量矩阵当前行权重，sign_t表示第t个非零元素符号，以等概率随机产生。B. Randomly generate nz _i rectangular areas such that px(i,t)～U(1,w-3), py(i,t)～U(1,h-3), pw(i,t)～ U(1, w-px-2), ph(i, t) ~ U(1, h-py-2), where px(i, t), py(i, t), pw(i, t) , ph(i, t) respectively represent the abscissa of the upper left corner, the ordinate of the upper left corner, width and height of the generated t-th rectangular area, U(a, b) represents the uniform distribution of integers in the interval [a, b], And generate a random measurement matrix value p _value (i, t)=w _i sign _t corresponding to the corresponding rectangular area, where

Indicates the weight of the current row of the random measurement matrix, and sign _t indicates the sign of the tth non-zero element, which is randomly generated with equal probability.

Finally obtain and store the values of all non-zero elements of the initial random measurement matrix, the non-zero elements correspond to the abscissa of the upper left corner of the rectangular area, the ordinate of the upper left corner, width and height information, and the set of the number of non-zero elements in each row {nz _i | i=1,2,...,m}.

The specific calculation steps for extracting the scale-invariant compressed feature vector of the sampled image are described below in conjunction with the accompanying drawing 3 of the description:

为了方便算法描述，本实施例中认为目标尺度在宽度和高度方向上变化一致，不再分别设置宽度和高度方向上的尺度参数，在不付出创造性劳动的前提下，可以扩展得到在宽度、高度方向尺度独立变化的实施例。Assume that w and h represent the width and height of the rectangular frame of the initial position of the target, respectively, and consider that the initial target size is 1. The integral image of the current video frame image is iH, the width and height of the sampled image are w _s and h _s respectively, the coordinates of the upper left corner of the sampled image are (x, y), and the scale of the sampled image is s, where

In order to facilitate the description of the algorithm, in this embodiment, it is considered that the target scale changes uniformly in the width and height directions, and the scale parameters in the width and height directions are no longer set separately. On the premise of no creative work, it can be expanded to obtain Embodiments where orientation scales vary independently.

(1) According to the scale s of the current sample image, the non-zero elements of the stored initial random measurement matrix R ₀ are adaptively corrected to obtain the random measurement matrix R _s at the scale s:

Keep the values of all non-zero elements in the initial measurement matrix unchanged, and the parameters of the rectangle corresponding to the non-zero elements px(i, t), py(i, t), pw(i, t), ph(i, t) correspond to Change to the original s times and round up according to the rounding, that is:

Where i=1,...,m, t=1,...,nz _i , m is the dimension of the compressed feature vector;

(2) The formula for calculating the compressed eigenvalue of the i-th dimension is as follows:

Among them, P _sum (i, t) represents the sum of pixel values in the rectangle corresponding to the tth non-zero element in row i of the modified random measurement matrix, which can be calculated as follows by using the integral graph:

P _sum (i, t)=iH(maxI, maxJ)-iH(maxI, minJ)-iH(minI, maxJ)+iH(minI, minJ)

in

Referring to Figure 3, the above formula What is represented is the normalized rectangular feature corresponding to the tth non-zero element in the i-th row in the modified random measurement matrix R _s , and p _value (i, t) represents the non-zero element in the corresponding random measurement matrix The value of the element. In Fig. 3, r _ij and x _j represent each element in the random measurement matrix R _s and each eigenvalue in the original high-dimensional normalized rectangular eigenvector. Since the random measurement matrix R _s is a very sparse matrix, in the above calculation In , only the non-zero elements need to be operated, and there is no need to calculate all the eigenvalues in the high-dimensional eigenvector.

Finally, the sampled image scale-invariant compressed feature vector v={v _i |i=1, 2, ..., m} is obtained.

Claims (5)

1. A real-time multi-scale target tracking method based on compressed sensing, characterized in that, comprising the following steps: System initialization phase: (1). Read the target initial position parameter R _state = [x, y, w, h], where (x, y) represents the coordinates of the upper left corner of the target initial position rectangle, w and h represent the target initial position rectangle respectively width and height; (2). Read the first frame image F ₀ of the video sequence, and convert it into a grayscale image, denoted as I ₀ ; (3). Calculate the integral image of the first frame image; (4). Construct an initial random measurement matrix R ₀ ; (5). Based on the center of the initial target position, collect positive and negative samples with the same width and height as the initial target; (6). Extract the scale-invariant compressed feature vectors of all collected positive and negative samples, and update the parameters of the naive Bayesian classifier; (7). Calculate the sample scale-invariant compressed feature vector v ₀ in the initial target rectangular frame, and initialize the particle distribution; Video real-time target tracking stage: (1). Read the t-th frame image, and convert it into a grayscale image, which is denoted as I _t ; (2). Calculate the current image integral map; (3). Use the second-order model to estimate and predict the particle state; (4). Calculating the scale-invariant compressed eigenvectors of all particles; (5). Utilize the naive Bayesian classifier to classify all particles, and obtain the classifier response values of all particles; (6). The particle with the largest classifier response is used as the target position, and the corresponding particle scale is used as the estimated value of the current scale of the target, and then the particle weight is calculated using the classifier response value; (7). Resample all particles according to the weight of the particles. After resampling, the number of samples at the particles with high weight increases, and the particles with too low weight are discarded to avoid degradation caused by low weight particle samples; (8). Based on the currently determined target center, collect positive and negative samples with the same width and height as the determined target frame; (9). Extract the scale-invariant compressed feature vectors of the collected positive and negative samples, and update the classifier parameters; (10). If the video is not over, return to step 1 and continue to read the next frame of video image. 2. a kind of real-time multi-scale target tracking method based on compressed sensing according to claim 1, is characterized in that, described sample image scale-invariant compressed feature vector extraction comprises the following steps: Let w and h denote the width and height of the rectangular frame of the initial position of the target respectively, and consider the initial target size scale to be 1; the width and height of the sample image are w _s and h _s respectively, and the position coordinates of the upper left corner of the sample image are (x, y) , the sample image scale is s,

The sample image is collected from a frame image in the video sequence and the integral image of the frame image is iH; (1). According to the scale s of the sample image, the non-zero elements of the initial random measurement matrix R ₀ are adjusted to obtain the random measurement matrix R _s under the scale s: Keep the values of all non-zero elements in R ₀ unchanged, and the rectangular parameters corresponding to the non-zero elements px(i, t), py(i, t), pw(i, t), and ph(i, t) become The original s times and according to rounding, the formula is as follows:

Where i=1,...,m, t=1,...,nz _i , m is the compressed feature vector dimension, and nz _i is the number of non-zero elements in row i of the initial random measurement matrix; (2). Calculate the i-th dimension eigenvalue of the scale-invariant compressed eigenvector, the formula is as follows: Among them, P _sum (i, t) represents the sum of the pixel values in the rectangle corresponding to the t non-zero element in the i-th row of the random measurement matrix R _s , which can be calculated by using the integral graph as follows: P _sum (i, t)=iH(maxI, maxJ)-iH(maxI, minJ)-iH(minI, maxJ)+iH(minI, minJ) in

In the formula,

Represents the corresponding eigenvalue in the original high-dimensional normalized rectangular eigenvector corresponding to the t non-zero element in the i-th row in the random measurement matrix R _s ; (3). The sampled image scale-invariant compressed feature vector v={v _i |i=1, 2,..., m}. 3. A kind of real-time multi-scale target tracking method based on compressed sensing according to claim 2, wherein the original high-dimensional normalized rectangular feature vector of the sample image has the following characteristics: make

Represents a sample image with width and height w and h respectively, the original high-dimensional normalized rectangular feature vector can be described as a sample image z and a series of normalized rectangular filters

The convolution, normalized rectangular filter formula is as follows: Among them, i and j respectively represent the width and height of the normalized rectangular filter; the filtered sample image obtained by filtering with all rectangular filters is represented as a column vector and these column vectors are connected to each other to form the sample image The original high-dimensional feature vector. 4. a kind of real-time multi-scale target tracking method based on compressed sensing according to claim 1, is characterized in that, described utilize second-order model to estimate particle state and prediction formula is as follows: $\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; wx</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; wy</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; ws</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}\end{array}\right.$ Among them, i=1, 2,..., pn, pn is the total number of particles, wx _t , wy _t , ws _t respectively represent the 0-mean Gaussian noise on the three state components of the particles at time t, and their standard deviations are std _x , std _y , std _s ; in order to prevent particle samples from exceeding the range of the image, it is necessary to perform out-of-boundary processing on the estimated state, and use the state after estimation and out-of-boundary processing (x _t (i), y _t (i), s _t (i) ) to replace the state parameters of the particle at time t, and replace the original particle state parameters at time t and t-1 with the corresponding particle state parameters at time t-1 and t-2, and use the second-order model to estimate and predict the particle state considering the speed information of the target movement . 5. a kind of real-time multi-scale object tracking method based on compressed sensing according to claim 1, is characterized in that, described particle weight calculation formula is as follows: ${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; pn</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$ Where i=1, 2,..., pn, pn is the total number of particles,

Indicates that the state of the i-th particle at time t is

The probability of observing the target at time t under the condition, $p\left(z_t\right|x_t^{*}\left(i\right))\&Proportional;\exp \left(h\left(v_i\right)\right)=\underset{j=1}{\overset{m}{\mathrm{\&Pi;}}}\frac{p\left(v_i^j\right|\mathrm{the\; y}=1)}{p\left(v_i^j\right|\mathrm{the\; y}=0)},$ H(v _i ) represents the Bayesian classifier response value of the i-th particle.

Images