CN104036528A - Real-time distribution field target tracking method based on global search - Google Patents

Abstract

The invention discloses a real-time distribution field target tracking method based on global search, which comprises the following steps: 1) manually marking the position of the target on the target image I selected in the first frame; 2) when the distribution field model of the target image I passes through After the Gaussian smoothing process, the target model is established, and the next step is to determine the area to be searched in the next frame of image and perform Gaussian smoothing on it to obtain the distribution field of the candidate area, and then find the image block with the largest correlation coefficient with the target model; The two target models obtained in two frames are fused according to a certain learning rate ρ to update the target model of the current frame; 4) Cycle 2) and 3) until the end of the entire video sequence. The present invention adopts the global template matching search strategy based on the correlation coefficient, which not only overcomes the limitation that the original distribution field is easy to fall into the local optimal value by using the gradient descent method, but also avoids the disadvantage that the similarity of the L1 distance measure is easily affected by noise, and improves the tracking performance.

Description

A real-time distributed field target tracking method based on global search

【Technical field】

The invention belongs to the field of computer vision and image analysis, in particular to a global search-based real-time distribution field target tracking method.

【Background technique】

Tracking technology is an important problem in computer vision and has been widely used in video surveillance, human-machine interface, robot perception, behavior understanding and action recognition and other fields. Due to the influence of complex factors such as target rotation, deformation, occlusion and illumination changes in the tracking process, visual tracking technology has always been a problem worthy of further research ^{[1, 2]} .

Generally, tracking algorithms mainly include three aspects: target representation, search strategy and model update. Among them, the target representation is the first problem to be solved by the tracking algorithm. The early target is often represented by a template, which contains the brightness, gradient information or other features of the target ^[3] , but it is sensitive to the change of the spatial structure of the target. Another template-based object representation method is the histogram representation method ^{[4, 5]} , which is simple in calculation, fast in speed, insensitive to deformation and posture changes of the target, and can avoid drift to a certain extent. However, it is a statistics-based object representation method that loses some spatial information. Moreover, the expressiveness of these methods decreases when the object-background similarity is high. In 2001, Viola ^[6] and others first introduced the Adaboost algorithm based on Haar-like features into face detection. Since the idea of integral image is applied to the calculation of Haar-like features, the feature acquisition speed is greatly improved. Inspired by this, Babenko ^[7] et al. used the method of online multi-instance learning to train a classifier, and used Haar-like features to train a discriminative model for targets and backgrounds to achieve robust tracking. The calculation of Haar-like features is simple, but it is more sensitive to edges and line segments, and can only describe the characteristics of a specific direction, which is relatively rough. Yao Zhijun ^[8] proposed a new spatial histogram similarity measure, which regards the spatial distribution of each interval in the spatial histogram as a Gaussian distribution, and uses JSD (Jensen-Shannon Divergence) and histogram intersection method to measure The spatial distribution and the similarity in the color histogram are applied to the particle filter tracking algorithm to improve the tracking results.

Recently, Laura Sevilla-Lara ^[9] adopted a novel distribution field (Distribution Fields, DF for short) object representation method and introduced it into the field of object tracking. This method first preserves the basic information of the original image by natural layering of the image, and introduces "fuzziness" into the target representation after Gaussian smoothing of each layer and between layers of the image, which overcomes the deformation and illumination to a certain extent. effects of such changes. However, in the search strategy, the original distribution field is searched according to the gradient descent method, and the search is stopped when the minimum value appears, which partially reduces the amount of calculation. But for the target function is non-convex, the gradient descent method only detects the maximum response position of the target according to the previous frame detection in the specific tracking process, and calculates the L1 norm search target in a limited area from this position in the current frame, so for When the target moves relatively fast, it is easy to fall into a local optimal solution when searching in a limited area, which limits the tracking effect.

references:

[1] Yilmaz A, Javed O, ShahHAH M. Object Tracking: a Survey [J], ACM Computing Surveys (CSUR), 2006, 38(4): 13.

[2] Yang Han-xuan, Zheng Feng, Wang Liang, et al.Recent advances and trends in visualtracking: A review[J].Neurocomputing.2011,74(18):3823-3831.

[3] Baker S, Matthews I. Lucas-Kanade20years on: A unifying framework [J]. International Journal of Computer Vision, 2004, 56(3): 221-255.

[4]Collins R T, Mean-shift blob tracking through scale space[C]//Proc of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Madison:IEEE Press,2003,2:II-234-40.

[5] Ning Ji-feng, Zhang Lei, Zhang David, et al. Robust Mean Shift Tracking with Corrected Background-Weighted Histogram [J], IET Computer Vision. 2012, 6(1): 62-69.

[6]Viola P, Jones M, Rapid object detection using a boosted cascade of simple features[C]//Proc of IEEE Conference on Computer Vision and Pattern Recognition,Hawaii:IEEE Press2001,1:511-518.

[7] Babenko B, Yang M, Belongie S, Robust Object Tracking with Online Multiple Instance Learning [J], IEEE Transaction on Pattern Analysis and Machine Intelligence, 2011, 33(8): 1619-1632.

[8] Yao Zhijun. A new spatial histogram similarity measurement method and its application in target tracking [J]. Journal of Electronics and Information Technology, 2013,35(7):1644-1649.

[9]Laura S L, Erik L M, Distribution fields for tracking[C]//Proc of IEEE Conference on Computer Vision and Pattern Recognition.Providence:IEEE Press,2012:1910-1917.

【Content of invention】

The purpose of the present invention is to provide a real-time distributed field target tracking method based on global search to address the difficulties and challenges in video target tracking such as target being partially occluded, rotated, zoomed, illumination changes, motion blur, and complex backgrounds. Changes to the target appearance model.

To achieve the above object, the present invention is achieved through the following technical solutions:

A real-time distributed field target tracking method based on global search, comprising the following steps:

1) Manually mark the target position on the selected target image I in the first frame, that is, delineate the target area with a rectangular frame, mark the coordinates of the upper left corner of the rectangular frame and the width and height of the rectangular frame, and then delineate the target area with The distribution field model representation, the construction of the distribution field model is as follows:

Using the Kronecker delta function to represent the target image I with a distribution field model, the distribution field model df(i,j,k) of the target image I is obtained, and its formula is as follows:

$\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; if\; i</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 255</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

In the formula: i and j represent the row and column of the target image I respectively;

K represents the number of layers to be divided into the target image I;

k represents the serial number of each layer, k=1, 2, 3..., K;

df(i, j, k) represents the value of the i-th row and the j-th column on the k layer after the target image I is decomposed, and its value range is 0 or 1;

A collection with a depth of 255/K is called a "layer";

I(i,j) represents the pixel value of the i-th row and j-th column on the target image I;

Next, Gaussian smoothing is performed on the distribution field model of the target image I. Gaussian smoothing is divided into spatial domain smoothing and feature domain smoothing. First, the distribution field model of the target image I is smoothed in the spatial domain. Smoothing in two directions of y, the calculation formula is as follows:

${\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

In the formula: df _s (k) represents the distribution field model of the target image I smoothed in the space domain;

df(k) represents the kth layer of the distribution field model of the target image I;

is a 2D Gaussian kernel with standard deviation σ _s ;

"*" is the convolution symbol;

Then perform feature domain smoothing on the distribution field model of the target image I smoothed in the spatial domain, and the calculation formula is as follows:

${\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; ss</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

In the formula: df _ss (i, j) represents the distribution field model of the target image I after feature domain smoothing;

df _s (i, j) represents the smoothed value of the i-th row and the j-th column of the distribution field model of the target image I after spatial domain smoothing;

is a 1D Gaussian kernel with a standard σ _f difference of ;

The distribution field model of the target image I after Gaussian smoothing is 1 for each column of each pixel;

In the tracking process, in order to adapt to the appearance changes of the target environment, it is necessary to dynamically update the distribution field model of the target image, that is, mix the distribution field model of the old target image and the distribution field model corresponding to the target image obtained by the new tracking in a certain proportion, Furthermore, a simple target image distribution field model update formula is as follows:

df _t+1 (i,j,k)=ρdf _t (i,j,k)+(1-ρ)df _t-1 (i,j,k) (4)

In the formula: ρ represents the learning rate, its value is between 0 and 1, generally set to 0.9, which is used to control the update speed of the target model;

t represents the current frame of the tracking target image, t+1 represents the next frame of the current frame of the tracking target image, and t-1 represents the previous frame of the current frame of the tracking target image;

df _t+1 (i, j, k) represents the value of the kth layer, row i, and column j of the next frame target distribution field model of the current frame in the tracking target image;

df _t (i, j, k) represents the value of the kth layer, row i, and column j of the current frame target distribution field model in the tracking target image;

df _t-1 (i, j, k) represents the value of the kth layer, row i, and column j of the previous frame target distribution field model of the current frame in the tracking target image;

2) After the distribution field model of the target image I has been processed by Gaussian smoothing, the target model is established, and the next step is to determine the area to be searched in the next frame of image and perform Gaussian smoothing on it to obtain the distribution field of the candidate area, and then find The image block with the maximum correlation coefficient, where the correlation coefficient between the target model and the distribution field of the candidate area can be expressed as:

${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

In the formula: df(m,n,k) is the target model;

df _i,j (m+1,n+1,k) is the distribution field of the candidate area;

i, j are the coordinates of row i and column j of the candidate area;

C _{i, j} is the correlation coefficient matrix between the candidate area distribution field and the target model;

m and n represent the coordinates of the mth row and nth column of the candidate area;

Suppose the candidate image is M×M, and the target image is N×N. In the frequency domain, the candidate image and the target image need to be zero-filled and extended to (M+N-1) ² , let S=M+N-1, The complexity of the algorithm will be reduced to O(S ² log ₂ S). Since the convolution in the time domain such as formula (5) can be realized by the product in the frequency domain, the correlation between the target model and the distribution field of the candidate area is calculated in the frequency domain Coefficient, the formula is as follows:

${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; ifft</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; fft</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; fft</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

In the formula:

ifft(fft(df(k))·(fft(df _i,j (k))) ^* ) represents the k-th kth of the target image distribution field model df(k) and the candidate image distribution field df _i,j (k) Each layer performs fast Fourier transform fft(df(k)), fft(df _i,j (k)), and finds the conjugate of the transformed candidate image distribution field (fft(df _i,j (k)) ) ^* , then multiply (fft(df _i,j (k))) ^* with fft(df(k)) to obtain the frequency domain correlation coefficient matrix, and finally obtain the time domain correlation coefficient matrix C _i by inverse Fourier transform _,j ;

fft(df(k)) means performing fast Fourier transform on the kth layer of the target image distribution field model df(k);

(fft(df _i,j (k))) ^* means to perform fast Fourier transform on the kth layer of the candidate image distribution field df _i,j (k) to obtain fft(df _i,j (k)) and find it conjugate(fft(df _i,j (k))) ^* ;

df(k) and df _i,j (k) are the target model and the kth layer of the candidate region distribution field respectively;

C _{i, j} is the correlation coefficient matrix of the distribution field of the candidate area and the distribution field model of the target image;

3) Use the formula (4) to fuse the two target models obtained by the two frames before and after according to a certain learning rate ρ to update the target model of the current frame;

4) Cycle 2) and 3) until the entire video sequence ends.

Compared with the prior art, the present invention has the following beneficial effects:

(1) Due to the use of the global template matching search strategy based on the correlation coefficient, it not only overcomes the limitation that the original distribution field is easy to fall into the local optimal value using the gradient descent method, but also avoids the disadvantage of using the L1 distance measure similarity that is easily affected by noise, and improves tracking performance.

(2) Due to the adoption of a dense sampling strategy, based on the circular matrix theory, Fourier transform is used to convert the correlation coefficient from the time domain with high computational complexity to the frequency domain with low computational complexity. The complexity of the algorithm is large In order to reduce, to meet the real-time requirements.

【Description of drawings】

Figure 1 is a schematic diagram of converting image Twinings into a distribution field, where Figure 1(a) is the original image, Figure 1(b) is a schematic diagram of the distribution field of the original image, and Figure 1(c) is the smoothed distribution field of the original image schematic diagram.

Fig. 2 is a schematic diagram of a time-domain global search algorithm.

Fig. 3 is a schematic diagram of an FFT-based correlation coefficient matching algorithm.

Fig. 4 is a video sequence center error diagram.

Fig. 5 is a comparison diagram of some tracking results of various video sequences.

【Detailed ways】

The present invention will be further described below in conjunction with accompanying drawing.

A kind of real-time distributed field target tracking method based on global search of the present invention comprises the following steps:

1) Manually mark the target position on the selected target image I in the first frame, that is, delineate the target area with a rectangular frame, mark the coordinates of the upper left corner of the rectangular frame and the width and height of the rectangular frame, and then delineate the target area with The distribution field model representation, the construction of the distribution field model is as follows:

Using the Kronecker delta function to represent the target image I with a distribution field model, the distribution field model df(i,j,k) of the target image I is obtained, and its formula is as follows:

$\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; if\; i</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 255</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

In the formula: i and j represent the row and column of the target image I respectively;

K represents the number of layers to be divided into the target image I;

k represents the serial number of each layer, k=1, 2, 3..., K;

df(i, j, k) represents the value of the i-th row and the j-th column on the k layer after the target image I is decomposed, and its value range is 0 or 1;

A collection with a depth of 255/K is called a "layer";

I(i,j) represents the pixel value of the i-th row and j-th column on the target image I;

Next, Gaussian smoothing is performed on the distribution field model of the target image I. Gaussian smoothing is divided into spatial domain smoothing and feature domain smoothing. First, the distribution field model of the target image I is smoothed in the spatial domain. Smoothing in two directions of y, the calculation formula is as follows:

${\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

In the formula: df _s (k) represents the distribution field model of the target image I smoothed in the space domain;

df(k) represents the kth layer of the distribution field model of the target image I;

is a 2D Gaussian kernel with standard deviation σ _s ;

"*" is the convolution symbol;

Then perform feature domain smoothing on the distribution field model of the target image I smoothed in the spatial domain, and the calculation formula is as follows:

${\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; ss</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

In the formula: df _ss (i, j) represents the distribution field model of the target image I after feature domain smoothing;

df _s (i, j) represents the smoothed value of the i-th row and the j-th column of the distribution field model of the target image I after spatial domain smoothing;

is a 1D Gaussian kernel with a standard σ _f difference of ;

The distribution field model of the target image I after Gaussian smoothing is 1 for each column of each pixel;

In the tracking process, in order to adapt to the appearance changes of the target environment, it is necessary to dynamically update the distribution field model of the target image, that is, mix the distribution field model of the old target image and the distribution field model corresponding to the target image obtained by the new tracking in a certain proportion, Furthermore, a simple target image distribution field model update formula is as follows:

df _t+1 (i,j,k)=ρdf _t (i,j,k)+(1-ρ)df _t-1 (i,j,k) (4)

In the formula: ρ represents the learning rate, its value is between 0 and 1, generally set to 0.9, which is used to control the update speed of the target model;

t represents the current frame of the tracking target image, t+1 represents the next frame of the current frame of the tracking target image, and t-1 represents the previous frame of the current frame of the tracking target image;

df _t+1 (i, j, k) represents the value of the kth layer, row i, and column j of the next frame target distribution field model of the current frame in the tracking target image;

df _t (i, j, k) represents the value of the kth layer, row i, and column j of the current frame target distribution field model in the tracking target image;

df _t-1 (i, j, k) represents the value of the kth layer, row i, and column j of the previous frame target distribution field model of the current frame in the tracking target image;

2) After the distribution field model of the target image I has been processed by Gaussian smoothing, the target model is established, and the next step is to determine the area to be searched in the next frame of image and perform Gaussian smoothing on it to obtain the distribution field of the candidate area, and then find The image block with the maximum correlation coefficient, where the correlation coefficient between the target model and the distribution field of the candidate area can be expressed as:

${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

In the formula: df(m,n,k) is the target model;

df _i,j (m+1,n+1,k) is the distribution field of the candidate area;

i, j are the coordinates of row i and column j of the candidate area;

C _{i, j} is the correlation coefficient matrix between the candidate area distribution field and the target model;

m and n represent the coordinates of the mth row and nth column of the candidate area;

Suppose the candidate image is M×M, and the target image is N×N. In the frequency domain, the candidate image and the target image need to be zero-filled and extended to (M+N-1) ² , let S=M+N-1, The complexity of the algorithm will be reduced to O(S ² log ₂ S). Since the convolution in the time domain such as formula (5) can be realized by the product in the frequency domain, the correlation between the target model and the distribution field of the candidate area is calculated in the frequency domain Coefficient, the formula is as follows:

${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; ifft</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; fft</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; fft</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

In the formula:

ifft(fft(df(k))·(fft(df _i,j (k))) ^* ) represents the k-th kth of the target image distribution field model df(k) and the candidate image distribution field df _i,j (k) Each layer performs fast Fourier transform fft(df(k)), fft(df _i,j (k)), and finds the conjugate of the transformed candidate image distribution field (fft(df _i,j (k)) ) ^* , then multiply (fft(df _i,j (k))) ^* with fft(df(k)) to obtain the frequency domain correlation coefficient matrix, and finally obtain the time domain correlation coefficient matrix C _i by inverse Fourier transform _,j ;

fft(df(k)) means performing fast Fourier transform on the kth layer of the target image distribution field model df(k);

(fft(df _i,j (k))) ^* means to perform fast Fourier transform on the kth layer of the candidate image distribution field df _i,j (k) to obtain fft(df _i,j (k)) and find it conjugate(fft(df _i,j (k))) ^* ;

df(k) and df _i,j (k) are the target model and the kth layer of the candidate region distribution field respectively;

C _{i, j} is the correlation coefficient matrix of the distribution field of the candidate area and the distribution field model of the target image;

3) Use the formula (4) to fuse the two target models obtained by the two frames before and after according to a certain learning rate ρ to update the target model of the current frame;

4) Cycle 2) and 3) until the entire video sequence ends.

See Figure 1 and Figure 2, Figure 1 is a schematic diagram of converting image Twinings into a distribution field, where Figure 1(a) is the original image, Figure 1(b) is a schematic diagram of the original image distribution field, and Figure 1(c) is Schematic illustration of the distribution field after smoothing the original image. Fig. 2 is a schematic diagram of a time-domain global search algorithm.

The image matching algorithm based on the global search strategy is a point-by-point search and matching algorithm on the candidate image: as shown in Figure 2: each dot represents the position to be searched, and the black dot represents the best matching position. This ergodic global search strategy makes it have the characteristics of a global optimal solution and avoids local extremums, but it has a disadvantage: when the target is large or the candidate area is relatively large, the calculation is huge and cannot meet the real-time requirements .

Figure 3 is a schematic diagram of FFT-based correlation coefficient matching algorithm

Using FFT to achieve fast image matching can effectively improve the time of correlation calculation, so that large-scale correlation matching algorithms can be processed in real time. Usually, the correlation value in the part where the boundary of the target image exceeds the boundary of the candidate image does not contribute to the final calculation result, so let the search image be X(k), the size is L×L, the template image is H(k), the size is M×M, Both images can be enlarged until N>=L, and N=2 ^r , (r is an integer), at this time, the upper left part of the relevant surface is the part required for image matching, and the lower right part is the confusing part of the circular correlation . The obfuscation part is ignored during the specific matching process and does not affect the matching result.

The specific operation steps are as follows:

(1) For the search graph X(k), the template graph H(k) is first extended to N×N, and then FFT is performed to obtain X(k) and H(K), and the total Yoke X ^* (k).

(2) Take the conjugate after multiplying the complex matrix X ^* (k)H(K) point by point, and obtain:

Y(k)＝(H(k)X ^* (k)) ^* ＝H ^* (k)X(k)

(3) Do N×N point IFFT on Y(k) to get the correlation matrix Y(n), and then remove the mixed part in the lower right corner to get the required correlation matrix y(n);

(4) The maximum value of the search matrix y(n) is the best matching point, invert its (x, y) coordinates, and return the result.

Example:

In order to verify the performance of the algorithm in this paper, this paper selects the video library edited by Babenko (2009) as the test set, which covers various difficulties in the field of visual tracking, such as long-term occlusion (Occluded Face, Occluded Face2), target rotation ( Sylv, Twinings), special shape (Surfer, Coke11), target deformation (Cliffbar, Twinings), illumination change (David, Sylv, Coke11), fast movement and occlusion (Tiger1, Tiger2), analog induction (Dollar, Cliffbar), etc. , in terms of comparison algorithms, the multi-instance tracking algorithm (MIL) representing sparse sampling and the original distributed field tracking algorithm (DF) are selected, and compared in terms of success rate, average error and speed, to test the performance of the proposed algorithm. The algorithm is run on Matlab R2010a under the windows7 system, and the computer configuration is Inter(R)Core(TM)i3-2130.3.40GHz CPU, 4GB RAM.

Parameter setting

The parameters of the algorithm are set as follows: for the number of layers K of the distribution field, considering the characteristics of different video sequences, set it to 4 or 8 layers. The fewer layers, the faster the speed, but the lower the accuracy rate; for some videos, too many layers may not have a better tracking effect. The parameter width and variance of the Gaussian smoothing in the spatial domain of the image are consistent with the original distributed field tracking algorithm. The larger the target, the larger the parameter, and vice versa. In the present invention, the variance of Gaussian smoothing in space domain is 0.3, and the variance of Gaussian smoothing in feature domain is 1. For the search radius of the candidate area, it is slightly different to set it to 5 or 20 due to the different size and motion range of each video target. Finally, the learning rate is set to 0.95 when the target model is updated, except for Coke11 and Tiger1 which are 0.8, and Sylv and Surf which are 0.75 to meet different video requirements. For multiple-instance learning tracking and the original distributed field tracking algorithm, we call out the best tracking results for each algorithm.

quantitative analysis

The experiment uses three different strategies to analyze the tracking results, which are the center offset distance between the test position and the real position (Table 1), the tracking success rate (Table 2) and the tracking speed (Table 3). Among them, the center error reflects the size of the deviation between the test position and the actual position. The smaller the value, the closer the tracking is to the actual position, and the better the effect; the larger the value, the farther the tracking position is from the real position, and the tracking performance is poor. Since each video sequence targets different scenes, its value fluctuates greatly. For a video frame, if (A∩B)/(A∪B)>0.5, it is considered that the tracking is successful, where A represents the tracking result rectangle, and B represents the real value rectangle of the target position. It is also an indicator to measure the tracking performance of the algorithm. It is generally considered that the tracking rectangle and the real rectangle of the target position overlap more than 50% and the tracking is considered successful. Otherwise, the tracking is considered to be failed. The larger the value, the better the tracking effect.

It can be seen from Table 1 and Table 2 that for most of the video sequences, the proposed algorithm has better tracking effect than the two algorithms of multi-instance and distributed field in terms of tracking center error and tracking success rate. Table 3 is the comparison of three algorithms in terms of tracking speed. Obviously, the proposed algorithm has a faster tracking speed, and it still has real-time performance under the Matlab environment. Figure 4 is the relative position error (in pixels) between the tracking result and the true value of the target position.

Table 1 The distance between the tracking results and the center of the real position

Table 2 Tracking success rate

Note: Bold fonts in Tables 1 and 2 indicate the best results, while italic fonts indicate the next best results.

It can be seen from Table 1 and Table 2 that for most video sequences, the proposed algorithm has better tracking effect than the two algorithms of multi-instance and distributed field. Table 3 is the comparison of three algorithms in terms of tracking speed. Obviously, the proposed algorithm has a faster tracking speed, and it still has real-time performance under the Matlab environment. Figure 4 is the relative position error (in pixels) between the tracking result and the true value of the target position. Analyzing the reason, this is mainly based on the special distributed field model target representation method. By extending the image to the three-dimensional space and smoothing in the feature domain and the spatial domain, not only can all the information of the target be retained, but also the attractive domain of the target can be expanded. Properly handle the influence of "uncertainty" factors to deal with the influence of deformation, occlusion, and illumination changes of the target during motion. However, the original distribution field adopts the gradient descent method in the target search strategy, that is, the new frame image is expanded and smoothed to obtain the distribution field model, and then the gradient of the difference between the target distribution field and the L1 norm of the candidate sample is calculated, and then descends along the gradient Continue to search until a local optimal solution is reached, and use it to update the appearance model of the target. However, when the target moves violently or is occluded for a long time, it is easy to fall into the local optimum when searching only in a limited area, and the target model will add wrong background information, which cannot effectively represent the target, resulting in matching when searching for the target again. Accuracy drops, limiting tracking effectiveness. Therefore, the present invention proposes a global search strategy using the correlation coefficient instead of the L1 norm to overcome the limitations of local search and poor real-time performance in the original distribution field.

Table 3 Tracking speed (frame/second)

Note: Bold font indicates the best result, while italic font indicates the next best result.

qualitative analysis

Figure 5 lists the tracking effects of the proposed method and other two tracking algorithms on some frames in 12 video sequences. The original distributed field (DF) and the algorithm of the present invention have achieved better tracking results. Since both have inherited the advantages of the distributed field algorithm in processing blur and maintaining the sensitivity of the spatial structure, the algorithm is not suitable for long-term and large-scale occlusion (see #732Occluded face, #387Occluded face2), complex background changes (see #886Sylvester, #281David) and scenes where the target is similar to the background (#236Dollar) and other scenes.

The distribution field target description operator proposed by the original distribution field tracking algorithm has the characteristics of a wide area of attraction, can retain the spatial information of the target and uncertain representation of the target, so that the target description operator can use a simple gradient descent method to search for the target , and achieved a good tracking effect. However, the gradient descent method has a local optimal solution problem, which causes inaccurate tracking. The accumulation of tracking inaccuracies eventually leads to tracking drift. And what the present invention adopts is dense sampling strategy to carry out global search matching, has avoided this defective, and adopts fast Fourier transform, algorithm complexity is low, and running speed is fast, has better tracking result (referring to Twinings#429, Girl# 311).

For the multi-instance learning method, it can adaptively update the appearance model to correctly reflect the change of the appearance of the target. It also introduces the concept of package, which alleviates the tracking drift problem caused by inaccurate target matching to a certain extent. It has good performance on occlusion, but when the occlusion leaves, there will be drift. This may be because the Noisy-OR model used by multiple examples does not fully use the correct information of the previous frame, resulting in the next frame collecting positive samples. Inefficient, the error increases when training the classifier and finally causes drift (see Occluded face#732, Occluded face2#387).

In general, the proposed algorithm first uses a dense sampling strategy for global search and matching, avoiding the algorithm from falling into a local optimal solution, and then converts the complex calculations in the time domain to the frequency domain for fast calculations with the help of Fourier transform, improving tracking performance. At the same time, it overcomes the shortcomings of multi-instance learning tracking and distribution field tracking, which are slow and easy to fall into local minima. Therefore, the algorithm of the present invention has more advantages in dealing with occlusion, appearance changes, illumination changes, rotation changes, etc., and has real-time performance.

It should be added that the present invention selects the video library edited by Babenko (2009) as a test set, which covers various difficulties in the field of visual tracking, such as long-term occlusion (Occluded Face, Occluded Face2), target rotation (Sylv , Twinings), special shape (Surfer, Coke11), target deformation (Cliffbar, Twinings), illumination change (David, Sylv, Coke11), fast movement and occlusion (Tiger1, Tiger2), analog induction (Dollar, Cliffbar), etc., At present, it has been adopted by most researchers and has become the de facto standard test library in the field of object tracking.

References: Babenko B, M.H. Yang, and S. Belongie. 2009. Visual Tracking with Online Multiple Instance Learning. Computer Vision and Pattern Recognition, IEEE Conference on, IEEE:983-990

Claims (2)

1. a real-time distribution field target tracking method based on global search, is characterized in that, comprises the following steps: 1) Manually mark the target position on the selected target image I in the first frame, that is, delineate the target area with a rectangular frame, mark the coordinates of the upper left corner of the rectangular frame and the width and height of the rectangular frame, and then delineate the target area with The distribution field model representation, the construction of the distribution field model is as follows: Using the Kronecker delta function to represent the target image I with a distribution field model, the distribution field model df(i,j,k) of the target image I is obtained, and its formula is as follows: $\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; if\; i</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 255</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$ In the formula: i and j represent the row and column of the target image I respectively; K represents the number of layers to be divided into the target image I; k represents the serial number of each layer, k=1, 2, 3..., K; df(i, j, k) represents the value of the i-th row and the j-th column on the k layer after the target image I is decomposed, and its value range is 0 or 1; A collection with a depth of 255/K is called a "layer"; I(i,j) represents the pixel value of the i-th row and j-th column on the target image I; Next, Gaussian smoothing is performed on the distribution field model of the target image I. Gaussian smoothing is divided into spatial domain smoothing and feature domain smoothing. First, the distribution field model of the target image I is smoothed in the spatial domain. Smoothing in two directions of y, the calculation formula is as follows: ${\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$ In the formula: df _s (k) represents the distribution field model of the target image I smoothed in the space domain; df(k) represents the kth layer of the distribution field model of the target image I; is a 2D Gaussian kernel with standard deviation σ _s ; "*" is the convolution symbol; Then perform feature domain smoothing on the distribution field model of the target image I smoothed in the spatial domain, and the calculation formula is as follows: ${\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; ss</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$ In the formula: df _ss (i, j) represents the distribution field model of the target image I after feature domain smoothing; df _s (i, j) represents the smoothed value of the i-th row and the j-th column of the distribution field model of the target image I after spatial domain smoothing; is a 1D Gaussian kernel with a standard σ _f difference of ; The distribution field model of the target image I after Gaussian smoothing is 1 for each column of each pixel; In the tracking process, in order to adapt to the appearance changes of the target environment, it is necessary to dynamically update the distribution field model of the target image, that is, mix the distribution field model of the old target image and the distribution field model corresponding to the target image obtained by the new tracking in a certain proportion, Furthermore, a simple target image distribution field model update formula is as follows: df _t+1 (i,j,k)=ρdf _t (i,j,k)+(1-ρ)df _t-1 (i,j,k) (4) In the formula: ρ represents the learning rate, and its value is between 0 and 1, which is used to control the update speed of the target model; t represents the current frame of the tracking target image, t+1 represents the next frame of the current frame of the tracking target image, and t-1 represents the previous frame of the current frame of the tracking target image; df _t+1 (i, j, k) represents the value of the kth layer, row i, and column j of the next frame target distribution field model of the current frame in the tracking target image; df _t (i, j, k) represents the value of the kth layer, row i, and column j of the current frame target distribution field model in the tracking target image; df _t-1 (i, j, k) represents the value of the kth layer, row i, and column j of the previous frame target distribution field model of the current frame in the tracking target image; 2) After the distribution field model of the target image I has been processed by Gaussian smoothing, the target model is established, and the next step is to determine the area to be searched in the next frame of image and perform Gaussian smoothing on it to obtain the distribution field of the candidate area, and then find The image block with the maximum correlation coefficient, where the correlation coefficient between the target model and the distribution field of the candidate area can be expressed as: ${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$ In the formula: df(m,n,k) is the target model; df _i,j (m+1,n+1,k) is the distribution field of the candidate area; i, j are the coordinates of row i and column j of the candidate area; C _{i, j} is the correlation coefficient matrix between the candidate area distribution field and the target model; m and n represent the coordinates of the mth row and nth column of the candidate area; Suppose the candidate image is M×M, and the target image is N×N. In the frequency domain, the candidate image and the target image need to be zero-filled and extended to (M+N-1) ² , let S=M+N-1, The complexity of the algorithm will be reduced to O(S ² log ₂ S). Since the convolution in the time domain such as formula (5) can be realized by the product in the frequency domain, the correlation between the target model and the distribution field of the candidate area is calculated in the frequency domain Coefficient, the formula is as follows: ${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; ifft</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; fft</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; fft</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; df</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$ In the formula: ifft(fft(df(k))·(fft(df _i,j (k))) ^* ) represents the k-th kth of the target image distribution field model df(k) and the candidate image distribution field df _i,j (k) Each layer performs fast Fourier transform fft(df(k)), fft(df _i,j (k)), and finds the conjugate of the transformed candidate image distribution field (fft(df _i,j (k)) ) ^* , then multiply (fft(df _i,j (k))) ^* with fft(df(k)) to obtain the frequency domain correlation coefficient matrix, and finally obtain the time domain correlation coefficient matrix C _i by inverse Fourier transform _,j ; fft(df(k)) means performing fast Fourier transform on the kth layer of the target image distribution field model df(k); (fft(df _i,j (k))) ^* means to perform fast Fourier transform on the kth layer of the candidate image distribution field df _i,j (k) to obtain fft(df _i,j (k)) and find it conjugate(fft(df _i,j (k))) ^* ; df(k) and df _i,j (k) are the target model and the kth layer of the candidate region distribution field respectively; C _{i, j} is the correlation coefficient matrix of the distribution field of the candidate area and the distribution field model of the target image; 3) Use the formula (4) to fuse the two target models obtained by the two frames before and after according to a certain learning rate ρ to update the target model of the current frame; 4) Cycle 2) and 3) until the entire video sequence ends. 2. A global search-based real-time distribution field target tracking method according to claim 1, wherein the learning rate ρ is set to 0.9.