CN103473790A - Online target tracking method based on increment bilateral two-dimensional principal component analysis (Bi-2DPCA) learning and sparse representation - Google Patents

Abstract

The invention relates to an online target tracking method based on increment bilateral two-dimensional principal component analysis (Bi-2DPCA) learning and sparse representation, and provides an increment Bi-2DPCA learning algorithm capable of rapidly and accurately updating a target sub-space model to reflect an appearance change of a target in a tracking process. Aiming at the problem that the target is frequently blocked and polluted by noise in the tracking process to cause that the tracking effect becomes worse, according to the method disclosed by the invention, a Bi-2DPCA-based sub-space model is embedded below a sparse representation frame, so that interferences to target positioning and target sub-space model updating caused by blocking and noise are furthest removed. Meanwhile, a novel method for calculating visual similarity is used. By adopting the method, energy distribution of Bi-2DPCA when an image is presented is considered, and is more accurate in comparison with a classical reconstruction error; tracking is achieved under a Bayesian inference framework; the target state is estimated by using a particle filtering algorithm.

Description

Online method for tracking target based on increment Bi-2DPCA study and rarefaction representation

Technical field

The present invention relates to a kind of online method for tracking target based on increment Bi-2DPCA study and rarefaction representation, it is the online method for tracking target that a kind of subspace in conjunction with bilateral two-dimensional principal component analysis method (Bilateral two-dimensional Principal Component Analysis, Bi-2DPCA) means model and rarefaction representation.

Background technology

Target following is a Basic Problems of computer vision field.It has a wide range of applications: comprise video monitoring, and behavioural analysis, motion event detects, and video frequency searching etc.Although a lot of scholars have made a lot of effort in the research in this field, vision is followed the tracks of and is remained a challenging research field.Because the variation that the outward appearance of target often can face due to illumination variation, blocks in tracing process, deformation, complicated movement background etc. cause.Therefore, a good target appearance model will play decisive role to the robustness of track algorithm.

As a kind of unsupervised learning and data analysis technique of classics, principal component analysis (PCA) (Principal Component Analysis, PCA) has outstanding feature extraction and data representation ability.Target subspace display model based on PCA can be good at describing tracking target, and obtains the tracking effect than robust.But this display model, often need the target image of 2 dimensions to be expanded into to the expression vector of 1 dimension by row or column, so just produced a very feature space of higher-dimension.In this feature space, calculate accurately its covariance matrix and be and be difficult to realize, calculate its proper vector thing very consuming time especially.And PCA be applicable to process the sample based on unimodal probability model, make the accuracy of the display model based on PCA depend on very much the probability distribution of sample.In order to overcome this defect of PCA, document " Bi-2DPCA:A Fast Face Coding Method for Recognition " has proposed the image representation method based on Bi-2DPCA, its main thought is directly processed the view data of 2 dimensions and without the image dress being changed to the vector of 1 dimension, the covariance matrix that the method builds small-sized, make the calculating of eigenwert and proper vector want the many of Simple fast than PCA.The method has obtained the reconstructed error rate less than PCA and good real-time simultaneously, and Bi-2DPCA has firm theoretical foundation to guarantee that it does not rely on the distribution of data.Recently, the tracking based on rarefaction representation has obtained increasing concern, and the method hypothetical target is formed by a series of To Templates and trifling template linear combination, by adding the constraint of sparse property, obtains the rarefaction representation vector of candidate target.The method can be good at processing noise and occlusion issue, but real-time is not high, and the method directly used the base of To Template as sparse dictionary, can not well describe structure and the variation in target signature space.

In sum, anti-the blocking with the noise ability of data representation ability that can Bi-2DPCA is outstanding and rarefaction representation combines, thereby reaches complementation.

Summary of the invention

The technical matters solved

For fear of the deficiencies in the prior art part, the present invention proposes a kind of online method for tracking target based on increment Bi-2DPCA study and rarefaction representation.

Technical scheme

A kind of online method for tracking target based on increment Bi-2DPCA study and rarefaction representation is characterized in that step is as follows:

Step 1: target-marking x in the first frame ₁, x ₁be the affine transformation parameter of target image piece in the first frame, an initialization N particle and weights thereof

Step 2: front T two field picture is used classical particle filtering algorithm tracking target, obtains initial target sample set A={A ₁, A ₂..., A _t, A _imean target image block matrix in the i two field picture, its size criteriaization is to m * n;

Step 3: element in A is carried out to the Bi-2DPCA processing, obtain following target subspace initial model:

element average image in A;

L ∈ R ^m*p: left transformation matrix, its column vector quadrature;

R ∈ R ^n*q: right transformation matrix, its column vector quadrature;

Step 4: input a new two field picture as present frame, and the hypothesis present frame is the t frame.Particle in former frame is pressed to its weights

proportional relation resampled, then use Gaussian motion model, obtain particle in present frame

Step 5: obtain particle in present frame

the outward appearance of correspondence image piece means, both being normalized into size is the image array of m * n

Step 6: calculate particle in present frame

the outward appearance of correspondence image piece means

probability with the visual similarity of target appearance model (this model consists of under the rarefaction representation framework the subspace model insertion based on Bi-2DPCA)

using this value as particle

new weights,

then use maximum posteriori criterion (MAP) criterion, obtain in present frame having the state estimation value of the particle of maximum weights as this frame target be the tracking results to present frame.If present frame is last frame, finishes, otherwise continue to carry out;

Step 7: judge whether to have followed the tracks of the M frame, if perform step 8, otherwise forward step 4 to; Described M is renewal frequency, 2<M<10;

Step 8: by this M tracking results, obtain an Increment Matrix B={A _t+1, A _t+2..., A _t+M, in B, each element is that the outward appearance of target image piece in the two field picture newly traced into means, its size criteria turns to m * n;

Step 9: use increment Bi-2DPCA algorithm to upgrade by C={C, B}, upgrade seasonal C={A, B} for the first time; The target subspace based on Bi-2DPCA built means model; Forward step 4 to.

Beneficial effect

A kind of online method for tracking target based on increment Bi-2DPCA study and rarefaction representation that the present invention proposes, in order to reflect target cosmetic variation in tracing process, proposed a kind of increment Bi-2DPCA learning algorithm and can upgrade fast and accurately the target subspace model.Often be blocked in tracing process due to target and, by noise pollution, cause the tracking effect variation, the present invention is directed to this problem, by the subspace model insertion based on Bi-2DPCA under the rarefaction representation framework.Block thereby eliminated to greatest extent interference target localization and target subspace model modification brought with noise.Simultaneously, the present invention has used a kind of method of new computation vision similarity, and the method has been considered the energy distribution of Bi-2DPCA when presentation video, compares classical reconstructed error more accurate.Tracking realizes, and uses particle filter algorithm estimating target state under the Bayesian inference framework.

The invention has the beneficial effects as follows: the target subspace based on Bi-2DPCA means to have the advantage of Bi-2DPCA at aspects such as image representation, covariance calculating, proper vector and eigenwert calculating, makes track algorithm rapidly and accurately.Target subspace based on Bi-2DPCA is meaned to embed the rarefaction representation framework, can obtain noise image, this noise image has been pointed out noise and has blocked position to occur, write contaminated pixel by getting rid of this, the renewal of target subspace model be can instruct, and then noise and occlusion issue well processed.

The accompanying drawing explanation

Fig. 1 the inventive method process flow diagram

Embodiment

Now in conjunction with the embodiments, the invention will be further described for accompanying drawing:

1) target-marking x in the first frame ₁(x ₁be the affine transformation parameter of target image piece in the first frame), an initialization N particle and weights thereof

2) front T two field picture is used classical particle filtering algorithm tracking target, obtains initial target sample set A={A ₁, A ₂..., A _t, A _imean target image block matrix in the i two field picture, its size is normalized to m * n;

3) calculate the covariance matrix of A

wherein for element average in A.To G _tcarry out Eigenvalues Decomposition (EVD), get its front q larger eigenwert

the characteristic of correspondence vector forms right transformation matrix

by element in A at R _tupper projection obtains: P=AR _t.Calculate P ^tcovariance matrix

f is carried out to Eigenvalues Decomposition (EVD), get its front p larger eigenwert

the characteristic of correspondence vector forms left transformation matrix L _t∈ R ^m*p.The target subspace model is: with

centered by, with L _t, R _tsubspace for span;

4) input a new two field picture as present frame, and the hypothesis present frame is the t frame.By particle in the former frame image by its weights proportional relation resampled, then use Gaussian motion model, obtain particle state parameter in present frame

particle after being about to resample adds the random perturbation of a Gaussian distributed.Gaussian motion model is generally arranged:

Σ wherein _xa diagonal matrix, the variance of element representation affine transformation parameter on diagonal line;

5) obtain particle in present frame the outward appearance of corresponding image block means, both being normalized into size is the image array of m * n

6) to particle in present frame

the outward appearance of correspondence image piece means

display model by target is explained.The present invention uses a pair of matrix of coefficients (E _i, e _i) mean that the outward appearance of candidate target means

meet the following optimal problem based on rarefaction representation:

$(E_i,e_i)=\underset{E,e}{\arg \min }{\left|\right|A_t^i-{\stackrel{\&OverBar;}{A}}_t-L_t{\mathrm{ER}}_t^T-e\left|\right|}_2^2+\mathrm{\&lambda;}{\left|\right|e\left|\right|}_1---\left(1\right)$

Wherein, E ∈ R ^p*q, e ∈ R ^m*n,

the average of current time, L _tthe left transformation matrix of current time subspace model, R _tbe the right transformation matrix of current time, λ is the sparse property of a balances noise image e and the factor of reconstructed error size.The present invention uses iterative algorithm to solve the optimal problem of (1) formula.By (E _i, e _i) can solve particle

the outward appearance of correspondence image piece means

probability with the visual similarity of target appearance model

?

have:

$p\left(A_t^i\right|x_t)\&Proportional;\exp [-\frac{1}{2{\mathrm{\&sigma;}}^2}{\left|\right|A_t^i-{\stackrel{\&OverBar;}{A}}_t-\left|L_t{E}_i{R}_t^T\right||}_F]\&times;\exp [-\frac{1}{2}\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}\underset{l=1}{\overset{q}{\mathrm{\&Sigma;}}}\frac{E_i^2(k,l)}{{\mathrm{\&lambda;}}_L^k+{\mathrm{\&lambda;}}_R^l}]---\left(2\right)$

Wherein, ${\mathrm{\&sigma;}}^2=\underset{i=p+1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_L^i+\underset{i=q+1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_R^i,$ ${\mathrm{\&lambda;}}_R^1\&GreaterEqual;{\mathrm{\&lambda;}}_R^2\&GreaterEqual;...\&GreaterEqual;{\mathrm{\&lambda;}}_R^n$ For G _tn eigenwert, ${\mathrm{\&lambda;}}_L^1\&GreaterEqual;{\mathrm{\&lambda;}}_L^2\&GreaterEqual;...\&GreaterEqual;{\mathrm{\&lambda;}}_L^m$ For F _tm eigenwert, G _t, F _tdefinition as step 3), ‖ ‖ _fthe F norm of matrix, E _i(k, l) means the capable l column element of k of noise image E.Using the MAP criterion, obtain in present frame having the state estimation value of the particle of maximum weights as this frame target, has been both our tracking results to present frame.If present frame is last frame, finishes, otherwise continue to carry out;

7) judge whether to have followed the tracks of M frame (M is renewal frequency, 2<M<10).If perform step 8), otherwise forward step 4) to;

8) obtain an Increment Matrix B={A by this M tracking results _t+1, A _t+2..., A _t+M, in B, each element is that the outward appearance of target image piece in the two field picture newly traced into means, its size criteria turns to m * n;

9) use increment Bi-2DPCA algorithm to upgrade by C={C, B}(upgrades seasonal C={A, B} for the first time) target subspace that builds means model.Forward step 4) to.

Increment Bi-2DPCA algorithm flow is as follows:

Input: target subspace model before upgrading: average

left and right transformation matrix L _t, R _t.Covariance average G _tand F _t.The target image set of blocks B={A newly traced into _t+1, A _t+2..., A _t+Mand average

Output: new target subspace model: average left and right transformation matrix L _t+M, R _t+M.Covariance average G _t+Mand F _t+M.

1. $\stackrel{\&OverBar;}{A_{t+M}}=\frac{1}{t+M}(t\&CenterDot;\stackrel{\&OverBar;}{A_t}+M\&CenterDot;\stackrel{\&OverBar;}{B});$

2. have $G_{t+M}=\frac{1}{t+M}[t\&CenterDot;G_t+M\&CenterDot;G_M+\frac{t\&CenterDot;M}{t+M}{(\stackrel{\&OverBar;}{A_t}-\stackrel{\&OverBar;}{A_M})}^T({\stackrel{\&OverBar;}{A}}_t-\stackrel{\&OverBar;}{A_M})],$

Wherein $G_M=\frac{1}{M}\underset{i=t+1}{\overset{t+M}{\mathrm{\&Sigma;}}}{(A_i-\stackrel{\&OverBar;}{A_M})}^T(A_i-\stackrel{\&OverBar;}{A_M});$

3. to G _t+Mcarry out Eigenvalues Decomposition, G is arranged _t+M=R Σ _rr ^t.Its maximum q of district eigenwert characteristic of correspondence vector forms right transformation matrix R _t+M;

By element in B at R _t+Mupper projection, have P _i=A _ir _t+M(i=t+1, t+2 ..., t+M);

5. have $F_{t+M}\&ap;\frac{1}{t+M}[t\&CenterDot;F_t+M\&CenterDot;P_M+\frac{t\&CenterDot;M}{t+M}({\stackrel{\&OverBar;}{A}}_t-\stackrel{\&OverBar;}{A_M})R_{t+M}{R}_{t+M}^T{(\stackrel{\&OverBar;}{A_t}-\stackrel{\&OverBar;}{A_M})}^T],$

Wherein $P_M=\frac{1}{M}\underset{i=t+1}{\overset{t+M}{\mathrm{\&Sigma;}}}(A_i-\stackrel{\&OverBar;}{A_M})R_{t+M}{R}_{t+M}^T{(A_i-\stackrel{\&OverBar;}{A_M})}^T.$

To F _t+Mcarry out Eigenvalues Decomposition, F is arranged _t+M=L Σ _ll ^t.Its maximum p of district eigenwert characteristic of correspondence vector forms left transformation matrix L _t+M.

Claims (1)

1. the online method for tracking target based on increment Bi-2DPCA study and rarefaction representation is characterized in that step is as follows:

Step 1: target-marking x in the first frame ₁, x ₁be the affine transformation parameter of target image piece in the first frame, an initialization N particle and weights thereof

Step 2: front T two field picture is used classical particle filtering algorithm tracking target, obtains initial target sample set A={A ₁, A ₂..., A _t, A _imean target image block matrix in the i two field picture, its size criteriaization is to m * n;

Step 3: element in A is carried out to the Bi-2DPCA processing, obtain following target subspace initial model:

element average image in A;

L ∈ R ^m*p: left transformation matrix, its column vector quadrature;

R ∈ R ^n*q: right transformation matrix, its column vector quadrature;

Step 4: input a new two field picture as present frame, and the hypothesis present frame is the t frame.Particle in former frame is pressed to its weights

proportional relation resampled, then use Gaussian motion model, obtain particle in present frame

Step 5: obtain particle in present frame

the outward appearance of correspondence image piece means, both being normalized into size is the image array of m * n

Step 6: calculate particle in present frame

the outward appearance of correspondence image piece means

probability with the visual similarity of target appearance model (this model consists of under the rarefaction representation framework the subspace model insertion based on Bi-2DPCA)

using this value as particle

new weights, then use maximum posteriori criterion (MAP) criterion, obtain in present frame having the state estimation value of the particle of maximum weights as this frame target

be the tracking results to present frame.If present frame is last frame, finishes, otherwise continue to carry out;

Step 7: judge whether to have followed the tracks of the M frame, if perform step 8, otherwise forward step 4 to; Described M is renewal frequency, 2<M<10;

Step 8: by this M tracking results, obtain an Increment Matrix B={A _t+1, A _t+2..., A _t+M, in B, each element is that the outward appearance of target image piece in the two field picture newly traced into means, its size criteria turns to m * n;

Step 9: use increment Bi-2DPCA algorithm to upgrade by C={C, B}, upgrade seasonal C={A, B} for the first time; The target subspace based on Bi-2DPCA built means model; Forward step 4 to.

Images