CN104091350A - Object tracking method achieved through movement fuzzy information - Google Patents

Abstract

The invention discloses an object tracking method achieved through movement fuzzy information. The method includes the steps of firstly, extracting characteristics of a tracked object, and constructing a fuzzy object image through a fuzzy kernel function; secondly, constructing a dictionary for image blocks through the obtained clear object and the obtained fuzzy object; thirdly, representing the tracked object through the constructed dictionary according to a sparse representation method, and extracting movement information through coefficients; fourthly, positioning the tracked object through the particle filter algorithm, wherein the tracked target is determined through the coefficients of the blocks, and the particle sampling distribution in the particle filter algorithm is related to the movement information; fifthly, renewing the dictionary through the newly-tracked target according to sparse coding and incremental learning. According to the method, the problem that the characteristics of an image can not be easily extracted due to movement blurring and degrading can be solved through the constructed dictionary, and by means of the movement information, accuracy of the object tracking algorithm can be improved, and the running speed can be increased.

Description

A kind of object tracking method of utilizing motion blur information

Technical field

The present invention relates to computer vision field, particularly a kind of object tracking method of utilizing motion blur information.

Background technology

In computer vision field, tracking problem is a study hotspot, the effect of track algorithm is affected by many-sided factor, and current work is mainly being processed following influence factor: picture noise, compound movement, object generation non-rigid shape deformations, partly or completely block, background interference, illumination variation, requirement of real-time etc.In many application, generally all suppose not motion suddenly of object, object keeps constant speed to move, and does not have motion blur etc. in video.But motion blur is inevitable in practice, the reason that produces motion blur has: object of which movement is too fast, the time shutter is too short, camera motion etc., and motion blur is very common in object tracking problem.

Motion blur is very large for object tracking impact, and the result that motion blur directly causes is exactly image degradation.For track algorithm, primary two key issues that solve are exactly Target Modeling and target localization, and wherein Target Modeling is exactly mainly to extract the feature of tracking target, and the direct effect characteristics of picture quality extracts.So build the tracker of a robustness, the problem that solves motion blur is very important.And about the processing of motion blur, be at present mainly image processing aspect, mainly study the method for image deblurring, and in the problem aspect tracking, do not do special processing for motion blur.

Summary of the invention

The shortcoming that the object of the invention is to overcome prior art, with not enough, provides a kind of object tracking method of utilizing motion blur information.

Object of the present invention realizes by following technical scheme:

Utilize an object tracking method for motion blur information, the step that comprises following order:

S1. extract the gray feature of tracking target, utilize the fuzzy object image of the fuzzy kernel function different fuzzy yardsticks of structure and blur direction;

S2. utilize the clear target and the fuzzy object that obtain to do fragmental image processing, every width image is divided into several fritters that can be overlapping, and each fritter, as an entry of dictionary, is constructed sparse coding dictionary;

S3. utilize the sparse coding dictionary of structure to adopt rarefaction representation method representation tracking target, to the entry coefficient stack of all directions and yardstick, utilization obtains coefficient and extracts movable information, comprises direction of motion and motion size;

S4. the coefficient of rarefaction representation is processed again, according to the stack of piece position, got diagonal coefficient, in conjunction with particle filter algorithm location tracking target;

S5. utilize the target newly tracing in conjunction with sparse coding and incremental learning, dictionary to be upgraded;

S6. repeat above-mentioned steps S3～S5, until follow the tracks of, finish.

Described step S1, specifically comprises the step of following order:

A, initial frame are assumed to be picture rich in detail, extract the gray feature of tracking target from initial frame;

B, utilize picture rich in detail and PSF function do convolution algorithm obtain simulation blurred picture, convolution algorithm formula is:

$I\left(x\right)=k_v*I_s\left(x\right)={\∫}_{-\∞}^{+\∞}\frac{1}{\sqrt{2\mathrm{\π}}}{e}^{-{\mathrm{\σ}}^2/2}{I}_s(x+\mathrm{v\σ})\mathrm{d\σ}$

Wherein Is (x) represents picture rich in detail, and I (x) represents the motion blur image of computer simulation, and v is two-dimensional vector, represents direction of motion and motion amplitude;

V is as parameter, different v values is set and obtains different blurred pictures, for the direction of v, get 8 different directions, be respectively π/4, pi/2 ... 2 π, and the value of amplitude is relevant with the fog-level of image, get n grade, finally obtain the motion blur image of 8n different directions different motion amplitude.

In step S2, the concrete constitution step of described sparse coding dictionary is as follows:

A, expansion picture rich in detail template, utilize initial frame image to extract tracking target feature, and target signature is done to translation and rotational transform, obtains several picture rich in detail templates;

B, the image of each image template is done to piecemeal process, be divided into the fritter that several can be overlapping, dividing method is as follows, the image that is 32 for a fabric width degree, at width, doing the method cut apart can be for 1～16 being the first fritter, 8～24 is the second fritter, and 16～32 the 3rd fritters, so just can be divided into 3 parts to width, in like manner can process height, the image that is 32 * 32 for a width pixel value, according to above-mentioned processing, can obtain 3 * 3 fritters;

C, each fritter is extracted to gray feature, as an entry of dictionary, can complete dictionary structure; The dictionary finally obtaining can be expressed as:

T＝[t _1,1,1,…,t _1,1,k,t _1,2,1,t _1,j,k?t _2,1,1?t _i,j,k]

Wherein i represents i direction of motion, and j represents j sport rank, and k represents k the fritter that each template is divided into;

D, dictionary is done to normalized.

In step S3, when described tracking target exists motion blur, image degradation, in order to represent the tracking target of degradation, adopts the sparse coding representing based on subspace to represent tracking target and extract movable information, and concrete steps are as follows:

A, sparse representation model are:

$\underset{c}{\min }\frac{1}{2}{\left|\right|Y-\mathrm{Tc}\left|\right|}_F^2+\mathrm{\λ}{\left|\right|c\left|\right|}_1$

Wherein T is sparse coding dictionary, and c is sparse coding coefficient, solves this model and adopts Lasso algorithm; Y is the tracking target feature through piecemeal, and now sparse representation model adopts Frobenius norm, and according to the dictionary of design, coefficient c is (i * j * k) * k matrix;

B, extraction movable information, sparse coding is the popularization of independent component analysis, by the sparse coding dictionary T constructing, each entry of Fuzzy Template structure has represented different direction of motion and motion amplitude, corresponding coefficient c has represented the weight along entry, according to all directions stack, obtains the weight θ of all directions, according to each amplitude stack, obtain each motion amplitude weight l:

${\mathrm{\θ}}_i=\underset{j}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}{c}_{i,j,k,k}$

$l_j=\underset{i}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}{c}_{i,j,k,k}.$

In step S4, described particle filter comprises observation model and forecast model two parts, and concrete steps are as follows:

A, according to suggestion distribution q, produce particle, suggestion distributes and has been integrated into movable information, and suggestion distribution concrete model is:

q(x _t|x _1:t-1,y _1:t)＝p(x _t|x _t-1)+p(x _t|x _t-1,x _t-2)+∑θ _iq _i(x _t| _xt-1,y _t-1)

Wherein, p (x _t| x _t-1) be first order Markov conversion, often getting average is x _t-1gaussian function; p(x _t| x _t-1, x _t-2) be second order markov transform, often getting average is x _t-1+ u _t-1gaussian function, u _t-1velocity contrast for front cross frame; q _i(x _t| x _t-1, y _t-1) representing the distribution along i direction of motion, its distribution is also Gaussian distribution, average is relevant to the movable information of extraction, is x _t-1+ v _t-1, v _t-1for motion vector, be parameter θ _ifunction with amplitude l;

B, the state that each particle is represented extract tracking target, to possible tracking target piecemeal, and use rarefaction representation to represent tracking target, measure state that each particle represents and the similarity of tracking target, measuring similarity adopts a minute block message, and specific practice is:

$p_{k1,k2}=\frac{1}{C}\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{c}_{i,j,k1,k2}$

Wherein C represents normalization constant, p _{k1, k2}represent to represent with k1 piece the weight of k2 piece, the matrix that the p finally obtaining is k * k, k the coefficient maximum that piece should be represented by k piece in theory, so get the diagonal element of matrix p, this diagonal values is larger, the most approaching with tracking target;

C, utilize the reconstructed error of rarefaction representation to set up likelihood function model;

D, utilize likelihood function resampling particle;

E, repetition above-mentioned steps, finish until follow the tracks of.

Described step S5, specifically comprises the step of following order:

A, clear template is done to PCA analyze, the proper vector obtaining is according to the sequence of eigenwert size, and the large several proper vectors of feature are as base U;

A few frame tracking results Y that the base U that B, utilization obtain preserves with rarefaction representation method representation, now tracking results is not done piecemeal processing, and sparse representation model is:

$[\hat{z},\hat{s}]\arg \min \frac{1}{2}{\left|\right|Y-\mathrm{Uz}-s\left|\right|}_F^2+\mathrm{\λ}{\left|\right|s\left|\right|}_1+\mathrm{\γ}{\left|\right|z\left|\right|}_1$

Wherein z is coefficient, and s is noise, and s is laplacian distribution;

Utilize rarefaction representation to proofread and correct tracking results, bearing calibration is as follows:

$Y_{\mathrm{new}}=\left\{\begin{array}{cc}\mathrm{Uz}& s=0\\ \mathrm{\μ}& s\≠0\end{array}\right.$

Wherein μ is the mean vector that PCA analytical calculation obtains;

The tracking results Ynew that C, utilization are proofreaied and correct adopts increment PCA learning method, and training obtains new base U again;

The base U that D, use newly obtain represents current tracking results y, proofreaies and correct current front result:

$[\hat{z},\hat{s}]=\arg \min \frac{1}{2}{\left|\right|y-\mathrm{Uz}-s\left|\right|}_2^2+\mathrm{\λ}{\left|\right|s\left|\right|}_1+\mathrm{\γ}{\left|\right|z\left|\right|}_1$

$y_{\mathrm{new}}=\left\{\begin{array}{cc}\mathrm{Uz}& s=0\\ \mathrm{\μ}& s\≠0\end{array}\right.$

Wherein μ is the mean vector that PCA analytical calculation obtains, and through the tracking results of overcorrect, is used for upgrading clear template;

E, according to the old entry of dictionary, upgrade slow, new entry and upgrade fast principle, generate accumulated probability sequence:

$L=\{0,\frac{1}{2^{n-1}-1},\frac{3}{2^{n-1}-1},...,1\}$

Produce the random number between 0～1, by random number, determine a certain template of replacing in clear template;

After F, clear template renewal, clear template piecemeal is processed, replaced the clear part of sparse coding dictionary T, new dictionary is done to normalized, obtain the dictionary upgrading.

In step S6, described PSF function is gaussian kernel function.

Compared with prior art, tool has the following advantages and beneficial effect in the present invention:

1, utilize rarefaction representation to solve the problem that motion blur causes image degradation, system energy robustness is followed the tracks of the object tracking problem that has motion blur.

2, utilize blocking characteristic extracting method, tracker energy robustness is processed the occlusion issue of tracing process.

3, utilize the movable information extracting, be effectively incorporated into particle filter algorithm, make track algorithm more efficient.

4, tracing process adopts dictionary updating method can effectively solve the variations such as posture in tracing process, illumination, and the dictionary updating blocking with motion blur is unified in same framework, reduces the tracking drifting problem that dictionary updating brings.

Accompanying drawing explanation

Fig. 1 is a kind of process flow diagram that utilizes the object tracking method of motion blur information of the present invention.

Embodiment

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited to this.

As Fig. 1, a kind of object tracking method of utilizing motion blur information, the step that comprises following order:

S1. extract the gray feature of tracking target, utilize the fuzzy object image of the fuzzy kernel function different fuzzy yardsticks of structure and blur direction, specifically comprise the step of following order:

A, initial frame are assumed to be picture rich in detail, extract the gray feature of tracking target from initial frame;

B, utilize picture rich in detail and PSF (point spread function) to do the blurred picture that convolution algorithm obtains simulation, the PSF function here adopts gaussian kernel function, and gaussian kernel function simulates motion blur image, and convolution algorithm formula is:

$I\left(x\right)=k_v*I_s\left(x\right)={\∫}_{-\∞}^{+\∞}\frac{1}{\sqrt{2\mathrm{\π}}}{e}^{-{\mathrm{\σ}}^2/2}{I}_s(x+\mathrm{v\σ})\mathrm{d\σ}$

Wherein Is (x) represents picture rich in detail, and I (x) represents the motion blur image of computer simulation, and v is two-dimensional vector, represents direction of motion and motion amplitude;

V is as parameter, different v values is set and obtains different blurred pictures, for the direction of v, get 8 different directions, be respectively π/4, pi/2 ... 2 π, and the value of amplitude is relevant with the fog-level of image, get n grade, finally obtain the motion blur image of 8n different directions different motion amplitude;

S2. utilize the clear target and the fuzzy object that obtain to do fragmental image processing, every width image is divided into several fritters that can be overlapping, and each fritter, as an entry of dictionary, is constructed sparse coding dictionary, and concrete constitution step is as follows:

A, expansion picture rich in detail template, utilize initial frame image to extract tracking target feature, and target signature is done to translation and rotational transform, obtains several picture rich in detail templates;

B, the image of each image template is done to piecemeal process, be divided into the fritter that several can be overlapping, dividing method is as follows, the image that is 32 for a fabric width degree, at width, doing the method cut apart can be for 1～16 being the first fritter, 8～24 is the second fritter, and 16～32 the 3rd fritters, so just can be divided into 3 parts to width, in like manner can process height, the image that is 32 * 32 for a width pixel value, according to above-mentioned processing, can obtain 3 * 3 fritters;

C, each fritter is extracted to gray feature, as an entry of dictionary, can complete dictionary structure; The dictionary finally obtaining can be expressed as:

T＝[t _1,1,1,…,t _1,1,k,t _1,2,1,t _1,j,k?t _2,1,1?t _i,j,k]

Wherein i represents i direction of motion, and j represents j sport rank, and k represents k the fritter that each template is divided into;

D, dictionary is done to normalized;

S3. utilize the sparse coding dictionary of structure to adopt rarefaction representation method representation tracking target, to the entry coefficient stack of all directions and yardstick, utilization obtains coefficient and extracts movable information, comprises direction of motion and motion size;

When described tracking target exists motion blur, image degradation, in order to represent the tracking target of degradation, adopts the sparse coding representing based on subspace to represent tracking target, and concrete steps are as follows:

A, sparse representation model are:

$\underset{c}{\min }\frac{1}{2}{\left|\right|Y-\mathrm{Tc}\left|\right|}_F^2+\mathrm{\λ}{\left|\right|c\left|\right|}_1$

Wherein T is sparse coding dictionary, and c is sparse coding coefficient, solves this model and adopts Lasso algorithm (minimum definitely reduction and selection algorithm); Y is the tracking target feature through piecemeal, and now sparse representation model adopts Frobenius norm, and according to the dictionary of design, coefficient c is (i * j * k) * k matrix;

B, extraction movable information, sparse coding is the popularization of independent component analysis, by the sparse coding dictionary T constructing, each entry of Fuzzy Template structure has represented different direction of motion and motion amplitude, corresponding coefficient c has represented the weight along entry, according to all directions stack, obtains the weight θ of all directions, according to each amplitude stack, obtain each motion amplitude weight l:

${\mathrm{\θ}}_i=\underset{j}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}{c}_{i,j,k,k}$

$l_j=\underset{i}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}{c}_{i,j,k,k};$

S4. the coefficient of rarefaction representation is processed again, according to the stack of piece position, got diagonal coefficient, in conjunction with particle filter algorithm location tracking target;

Described particle filter comprises observation model and forecast model two parts, and concrete steps are as follows:

A, according to suggestion distribution q, produce particle, suggestion distributes and has been integrated into movable information, and suggestion distribution concrete model is:

q(x _t|x _1:t-1,y _1:t)＝p(x _t|x _t-1)+p(x _t|x _t-1,x _t-2)+∑θ _iq _i(x _t|x _t-1,y _t-1)

Wherein, p (x _t| x _t-1) be first order Markov conversion, often getting average is x _t-1gaussian function; p(x _t| x _t-1, x _t-2) be second order markov transform, often getting average is x _t-1+ u _t-1gaussian function, u _t-1velocity contrast for front cross frame; q _i(x _t| x _t-1, y _t-1) representing the distribution along i direction of motion, its distribution is also Gaussian distribution, average is relevant to the movable information of extraction, is x _t-1+ v _t-1, v _t-1for motion vector, be parameter θ _ifunction with amplitude l;

B, the state that each particle is represented extract tracking target, to possible tracking target piecemeal, and use rarefaction representation to represent tracking target, measure state that each particle represents and the similarity of tracking target, measuring similarity adopts a minute block message, and specific practice is:

$p_{k1,k2}=\frac{1}{C}\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{c}_{i,j,k1,k2}$

Wherein C represents normalization constant, p _{k1, k2}represent to represent with k1 piece the weight of k2 piece, the matrix that the p finally obtaining is k * k, k the coefficient maximum that piece should be represented by k piece in theory, so get the diagonal element of matrix p, this diagonal values is larger, the most approaching with tracking target;

C, utilize the reconstructed error of rarefaction representation to set up likelihood function model;

D, utilize likelihood function resampling particle;

E, repetition above-mentioned steps, finish until follow the tracks of;

S5. utilize the target newly tracing in conjunction with sparse coding and incremental learning, dictionary to be upgraded, specifically comprise the step of following order:

A, clear template is done to PCA analyze, the proper vector obtaining is according to the sequence of eigenwert size, and the large several proper vectors of feature are as base U;

A few frame tracking results Y that the base U that B, utilization obtain preserves with rarefaction representation method representation, now tracking results is not done piecemeal processing, and sparse representation model is:

$[\hat{z},\hat{s}]\arg \min \frac{1}{2}{\left|\right|Y-\mathrm{Uz}-s\left|\right|}_F^2+\mathrm{\λ}{\left|\right|s\left|\right|}_1+\mathrm{\γ}{\left|\right|z\left|\right|}_1$

Wherein z is coefficient, and s is noise, and s is laplacian distribution;

Utilize rarefaction representation to proofread and correct tracking results, bearing calibration is as follows:

$Y_{\mathrm{new}}=\left\{\begin{array}{cc}\mathrm{Uz}& s=0\\ \mathrm{\μ}& s\≠0\end{array}\right.$

Wherein μ is the mean vector that PCA analytical calculation obtains;

The tracking results Ynew that C, utilization are proofreaied and correct adopts increment PCA learning method, and training obtains new base U again;

The base U that D, use newly obtain represents current tracking results y, proofreaies and correct current front result:

$[\hat{z},\hat{s}]=\arg \min \frac{1}{2}{\left|\right|y-\mathrm{Uz}-s\left|\right|}_2^2+\mathrm{\λ}{\left|\right|s\left|\right|}_1+\mathrm{\γ}{\left|\right|z\left|\right|}_1$

$y_{\mathrm{new}}=\left\{\begin{array}{cc}\mathrm{Uz}& s=0\\ \mathrm{\μ}& s\≠0\end{array}\right.$

Wherein μ is the mean vector that PCA analytical calculation obtains, and through the tracking results of overcorrect, is used for upgrading clear template;

E, according to the old entry of dictionary, upgrade slow, new entry and upgrade fast principle, generate accumulated probability sequence:

$L=\{0,\frac{1}{2^{n-1}-1},\frac{3}{2^{n-1}-1},...,1\}$

Produce the random number between 0～1, by random number, determine a certain template of replacing in clear template;

After F, clear template renewal, clear template piecemeal is processed, replaced the clear part of sparse coding dictionary T, new dictionary is done to normalized, obtain the dictionary upgrading;

S6. repeat above-mentioned steps S3～S5, until follow the tracks of, finish.

Above-described embodiment is preferably embodiment of the present invention; but embodiments of the present invention are not restricted to the described embodiments; other any do not deviate from change, the modification done under Spirit Essence of the present invention and principle, substitutes, combination, simplify; all should be equivalent substitute mode, within being included in protection scope of the present invention.

Claims (7)

1. an object tracking method of utilizing motion blur information, is characterized in that, the step that comprises following order:

S1. extract the gray feature of tracking target, utilize the fuzzy object image of the fuzzy kernel function different fuzzy yardsticks of structure and blur direction;

S2. utilize the clear target and the fuzzy object that obtain to do fragmental image processing, every width image is divided into several fritters that can be overlapping, and each fritter, as an entry of dictionary, is constructed sparse coding dictionary;

S3. utilize the sparse coding dictionary of structure to adopt rarefaction representation method representation tracking target, to the entry coefficient stack of all directions and yardstick, utilization obtains coefficient and extracts movable information, comprises direction of motion and motion size;

S4. the coefficient of rarefaction representation is processed again, according to the stack of piece position, got diagonal coefficient, in conjunction with particle filter algorithm location tracking target;

S5. utilize the target newly tracing in conjunction with sparse coding and incremental learning, dictionary to be upgraded;

S6. repeat above-mentioned steps S3～S5, until follow the tracks of, finish.

2. the object tracking method of utilizing motion blur information according to claim 1, is characterized in that, described step S1 specifically comprises the step of following order:

A, initial frame are assumed to be picture rich in detail, extract the gray feature of tracking target from initial frame;

B, utilize picture rich in detail and PSF function do convolution algorithm obtain simulation blurred picture, convolution algorithm formula is:

$I\left(x\right)=k_v*I_s\left(x\right)={\∫}_{-\∞}^{+\∞}\frac{1}{\sqrt{2\mathrm{\π}}}{e}^{-{\mathrm{\σ}}^2/2}{I}_s(x+\mathrm{v\σ})\mathrm{d\σ}$

Wherein Is (x) represents picture rich in detail, and I (x) represents the motion blur image of computer simulation, and v is two-dimensional vector, represents direction of motion and motion amplitude;

V is as parameter, different v values is set and obtains different blurred pictures, for the direction of v, get 8 different directions, be respectively π/4, pi/2 ... 2 π, and the value of amplitude is relevant with the fog-level of image, get n grade, finally obtain the motion blur image of 8n different directions different motion amplitude.

3. the object tracking method of utilizing motion blur information according to claim 1, is characterized in that, in step S2, the concrete constitution step of described sparse coding dictionary is as follows:

A, expansion picture rich in detail template, utilize initial frame image to extract tracking target feature, and target signature is done to translation and rotational transform, obtains several picture rich in detail templates;

B, the image of each image template is done to piecemeal process, be divided into the fritter that several can be overlapping, dividing method is as follows, the image that is 32 for a fabric width degree, at width, doing the method cut apart can be for 1～16 being the first fritter, 8～24 is the second fritter, and 16～32 the 3rd fritters, so just can be divided into 3 parts to width, in like manner can process height, the image that is 32 * 32 for a width pixel value, according to above-mentioned processing, can obtain 3 * 3 fritters;

C, each fritter is extracted to gray feature, as an entry of dictionary, can complete dictionary structure; The dictionary finally obtaining can be expressed as:

T＝[t _1,1,1,…,t _1,1,k,t _1,2,1,t _1,j,k?t _2,1,1?t _i,j,k]

Wherein i represents i direction of motion, and j represents j sport rank, and k represents k the fritter that each template is divided into;

D, dictionary is done to normalized.

4. the object tracking method of utilizing motion blur information according to claim 1, it is characterized in that, in step S3, while there is motion blur in described tracking target, image degradation, in order to represent the tracking target of degradation, adopt the sparse coding representing based on subspace to represent tracking target and extract movable information, concrete steps are as follows:

A, sparse representation model are:

$\underset{c}{\min }\frac{1}{2}{\left|\right|Y-\mathrm{Tc}\left|\right|}_F^2+\mathrm{\λ}{\left|\right|c\left|\right|}_1$

Wherein T is sparse coding dictionary, and c is sparse coding coefficient, solves this model and adopts Lasso algorithm; Y is the tracking target feature through piecemeal, and now sparse representation model adopts Frobenius norm, and according to the dictionary of design, coefficient c is (i * j * k) * k matrix;

B, extraction movable information, sparse coding is the popularization of independent component analysis, by the sparse coding dictionary T constructing, each entry of Fuzzy Template structure has represented different direction of motion and motion amplitude, corresponding coefficient c has represented the weight along entry, according to all directions stack, obtains the weight θ of all directions, according to each amplitude stack, obtain each motion amplitude weight l:

${\mathrm{\θ}}_i=\underset{j}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}{c}_{i,j,k,k}$

$l_j=\underset{i}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}{c}_{i,j,k,k}.$

5. the object tracking method of utilizing motion blur information according to claim 1, it is characterized in that, in step S4, described particle filter comprises observation model and forecast model two parts, and the design that suggestion distributes is the very large factor of influence of track algorithm effect, concrete steps are as follows:

A, according to suggestion distribution q, produce particle, suggestion distributes and has been integrated into movable information, and suggestion distribution concrete model is:

q(x _t|x _1:t-1,y _1:t)＝p(x _t|x _t-1)+p(x _t|x _t-1,x _t-2)+∑θ _iq _i(x _t|x _t-1,y _t-1)

Wherein, p (x _t| x _t-1) be first order Markov conversion, often getting average is x _t-1gaussian function; p(x _t| x _t-1, x _t-2) be second order markov transform, often getting average is x _t-1+ u _t-1gaussian function, u _t-1velocity contrast for front cross frame; q _i(x _t| x _t-1, y _t-1) representing the distribution along i direction of motion, its distribution is also Gaussian distribution, average is relevant to the movable information of extraction, is x _t-1+ v _t-1, v _t-1for motion vector, be parameter θ _ifunction with amplitude l;

B, the state that each particle is represented extract tracking target, to possible tracking target piecemeal, and use rarefaction representation to represent tracking target, measure state that each particle represents and the similarity of tracking target, measuring similarity adopts a minute block message, and specific practice is:

$p_{k1,k2}=\frac{1}{C}\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{c}_{i,j,k1,k2}$

Wherein C represents normalization constant, p _{k1, k2}represent to represent with k1 piece the weight of k2 piece, the matrix that the p finally obtaining is k * k, k the coefficient maximum that piece should be represented by k piece in theory, so get the diagonal element of matrix p, this diagonal values is larger, the most approaching with tracking target;

C, utilize the reconstructed error of rarefaction representation to set up likelihood function model;

D, utilize likelihood function resampling particle;

E, repetition above-mentioned steps, finish until follow the tracks of.

6. the object tracking method of utilizing motion blur information according to claim 1, is characterized in that, described step S5 specifically comprises the step of following order:

A, clear template is done to PCA analyze, the proper vector obtaining is according to the sequence of eigenwert size, and the large several proper vectors of feature are as base U;

A few frame tracking results Y that the base U that B, utilization obtain preserves with rarefaction representation method representation, now tracking results is not done piecemeal processing, and sparse representation model is:

$[\hat{z},\hat{s}]\arg \min \frac{1}{2}{\left|\right|Y-\mathrm{Uz}-s\left|\right|}_F^2+\mathrm{\λ}{\left|\right|s\left|\right|}_1+\mathrm{\γ}{\left|\right|z\left|\right|}_1$

Wherein z is coefficient, and s is noise, and s is laplacian distribution;

Utilize rarefaction representation to proofread and correct tracking results, bearing calibration is as follows:

$Y_{\mathrm{new}}=\left\{\begin{array}{cc}\mathrm{Uz}& s=0\\ \mathrm{\μ}& s\≠0\end{array}\right.$

Wherein μ is the mean vector that PCA analytical calculation obtains;

The tracking results Ynew that C, utilization are proofreaied and correct adopts increment PCA learning method, and training obtains new base U again;

The base U that D, use newly obtain represents current tracking results y, proofreaies and correct current front result:

$[\hat{z},\hat{s}]=\arg \min \frac{1}{2}{\left|\right|y-\mathrm{Uz}-s\left|\right|}_2^2+\mathrm{\λ}{\left|\right|s\left|\right|}_1+\mathrm{\γ}{\left|\right|z\left|\right|}_1$

$y_{\mathrm{new}}=\left\{\begin{array}{cc}\mathrm{Uz}& s=0\\ \mathrm{\μ}& s\≠0\end{array}\right.$

Wherein μ is the mean vector that PCA analytical calculation obtains, and through the tracking results of overcorrect, is used for upgrading clear template;

E, according to the old entry of dictionary, upgrade slow, new entry and upgrade fast principle, generate accumulated probability sequence:

$L=\{0,\frac{1}{2^{n-1}-1},\frac{3}{2^{n-1}-1},...,1\}$

Produce the random number between 0～1, by random number, determine a certain template of replacing in clear template;

After F, clear template renewal, clear template piecemeal is processed, replaced the clear part of sparse coding dictionary T, new dictionary is done to normalized, obtain the dictionary upgrading.

7. the object tracking method of utilizing motion blur information according to claim 1, is characterized in that, in step S6, described PSF function is gaussian kernel function.