CN102800108B - Based on the visual target tracking method of local restriction least-squares estimation - Google Patents

Abstract

The invention provides a kind of visual target tracking method based on local restriction least-squares estimation.1) the present invention utilizes local restriction least-squares estimation, provides the description of the topological structure of target.Fully inquired into the relation between target patch area information, this contributes to processing target and blocks and similar purpose interference problem.2) the present invention is based on the measuring similarity that Pasteur's coefficient defines two topological structures.Target following result is provided under particle filter framework.Experimental result shows that the method that method proposed by the invention compares has better tracking accuracy.<!-- 2 -->

Description

Based on the visual target tracking method of local restriction least-squares estimation

Technical field

What the present invention relates to is visual target tracking method in a kind of computer vision field.Specifically, what relate to is a kind of visual target tracking method based on local restriction least-squares estimation.

Background technology

Current tracking technique mainly divides two large classes: determinacy and statistically method.Deterministic Methods is as MeanShift (MS) tracker [D.Comaniciu, V.Ramesh, andP.Meer.Real-timetrackingofnon-rigidobjectsusingmeansh ift.CVPR, pp.142-149,2000.], the position of target is obtained by the similarity iteratively maximizing To Template and candidate.Although the method is computationally very effective, its to background interference, clutter, the sensitivity such as to block.Once lose objects, rely on and self be difficult to recover to follow the tracks of.This problem is eased by statistical method.Statistical method is as particle filter [K.Nummiaroa, E.Koller-Meierb, andL.V.Gool.AnAdaptiveColor-BasedParticleFilter [J] .ImageandVisionComputing, vol.21, pp.99-110,2003.] maintain multiple hypothesis at state space, obtain more robustness with this.These document utilization histogram features describe the apparent of a region.Because for histogram, it is little that the spatial information of target utilizes, thus cause tracking performance inaccurate and can not process and block.

People propose various method and solve spatial information shortage problem.Literary composition [S.Birchfield, S.Rangarajan.Spatiogramsversushistogramsforregion-basedt racking.CVPR, pp.1158-1163,2005.] single order of the position of color and second moment information are added in histogrammic each chest.[K.Okuma, A.Taleghani, N.D.Freitas, J.LittleandD.G.Lowe.Aboostedparticlefilter:multitargetde tectionandtracking.ECCV, pp.28-39,2004.], [E.MaggioandA.Cavallaro.Multi-parttargetrepresentationfor colourtracking.ICIP, pp.729-732,2005.] adopt multiple histogram to carry out compensation space information to the different piece of target.[H.Wang, D.SuterandK.Schindler.Effectiveappearancemodelandsimilar itymeasureforparticlefilteringandvisualtracking.ECCV, pp.606-618,2008.] the apparent model of Gaussian Mixture of spatio-color is given by adding spatial information.Adam [A.Adam, E.RivlinandI.Shimshoni.Robustfragments-basedtrackingusin gtheintegralhistogram.CVPR, vol.1, pp.798-805,2006.] target is divided into nonoverlapping panel region, each panel region is by comparing template image corresponding part to give a mark, and then the measuring similarity of all panel region combines the likelihood score obtaining assumed position.

Above-mentioned all methods, do not consider the topological relation between part or panel region.Topological relation can reflect object construction information, and this contributes to processing target and blocks and similar purpose interference problem.Therefore, based on local restriction least-squares estimation [J.Wang, J.Yang, K.Yu, F.Lv, T.Huang, andY.Gong.Locality-constrainedlinearcodingforimageclassi fication.CVPR, pp.3360-3367,2010.], the present invention provides a description of object construction.Tracing process is carried out under particle filter framework.The weight that each particle is corresponding is measured by Pasteur's coefficient.

Summary of the invention

The object of the invention is for existing methodical deficiency, the new method being applied to target following based on local restriction least-squares estimation proposed, main innovate point comprises: 1) utilize half-tone information to set up a topological structure to the panel region of multiple overlaps of target, the corresponding summit of each panel region, sheet and sheet or the relation between summit and summit or the weight on limit utilize local restriction least-squares estimation to weigh.The weight corresponding with on the limit between the panel region that this panel region is more similar is larger.2) measuring similarity of two topological structures based on Pasteur's coefficient is provided.Because Pasteur's coefficient has extraordinary geometric interpretation, the similarity of measurement two topological structures can be used for.The method can effectively be blocked and similar purpose interference problem by processing target.

The importance of similarity is emphasized based on local restriction least-squares estimation, similarity is introduced in regularization, this estimation is used for the uniform enconding problem of other all panel region of each panel region herein, this coding gives the larger panel region of similarity larger weight, makes again this encoding error minimum simultaneously.Thus the topological structure matrix built between these panel region.

Because Pasteur's coefficient has extraordinary geometric interpretation, with the measuring similarity of its definition two topological structure matrixes.

The present invention is achieved through the following technical solutions,

According to an aspect of the present invention, a kind of visual target tracking method based on local restriction least-squares estimation is provided, it is characterized in that: comprise the steps:

The first step, to template image after resolution adjustment, utilizes a slip wicket, and each horizontal direction or vertical direction slip q pixel, obtain multiple panel region thus, and each panel region color characteristic of the vectorization of its correspondence describes, and is designated as S={s _i∈ R ^d| i=1 ..., M};

Second step, sets up a topological structure to the color characteristic set that the first step obtains, regards a summit as by each element, set up an adjacent map; Being calculated as follows of weight on limit in adjacent map:

The panel region of each vectorization represents by other all vectorization panel region, namely here coefficient a _ij(i ≠ j) provides by solving local restriction least-squares estimation problem;

Make A=(| a _ij|), its diagonal entry is 0, the topological structure in this matrix reflection template between panel region information;

3rd step, does the work same with template image to the object candidate area obtained by particle sampler, sets up the topological structure that this candidate region is corresponding; Order l is particle index;

4th step, utilizes Pasteur's coefficient to measure A and B ^(l)similarity;

5th step, obtains the weight of particle according to the measuring similarity of the 4th step, utilize MAP to estimate to obtain the tracking results of target.

Preferably, in a first step, described template image through resolution adjustment, utilize sub-moving window to obtain the panel region of multiple overlap, in panel region, gray scale or RGB vectorization feature describe the information of this panel region.

Preferably, the adjacent map set up between template panel region information described in second step, i.e. topological structure, refer to and utilize local restriction least-squares estimation to encode to other all panel region information of each panel region information, code coefficient is used for weighing other panel region information to the contribution degree of this panel region information.

Preferably, in the third step, propagated by particle filter and obtain multiple particle candidate target, set up its topological structure to each candidate target, specific practice is identical with the topological structure building template.That is, to particle the panel region of overlapping vectorization is extracted in corresponding candidate region sheet information can be expressed as in advance with the relation of all remaining information its corresponding sparse solving is obtained by local restriction least-squares estimation.Obtain matrix B ^(l).

Preferably, the definition of the measuring similarity in the 4th step.Be implemented as:

To two topological structure A and B ^(l), make A _colwith be respectively A and B ^(l)col row, and to each row to normalization, namely so, this two matrix Pasteur coefficient is defined as $\mathrm{sim}(A,B^{\left(l\right)})=\underset{\mathrm{col}}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}\sqrt{B_{\mathrm{col},j}^{\left(l\right)}{A}_{\mathrm{col},j}}.$

Preferably, in the 5th step, the estimation of dbjective state, utilizes MAP estimation to obtain.

Preferably, described color characteristic is RGB feature or gray feature.

The present invention proposes a kind of to block and background interference has the tracking of robustness, a kind of topological structure providing target apparent based on local restriction least-squares estimation describes, relation between this description energy target acquisition each several part, by the similarity weighing two topological structures based on Pasteur's coefficient, except to block and background interference has except robustness, to illumination variation to a certain degree, also there is certain robustness.Search framework is the multiple candidate samples obtaining target in particle filter framework.The present invention can be used in civilian and military target tracking and identifying system.

Accompanying drawing explanation

Fig. 1-4 and Fig. 5-8 is for the present invention is directed to tracking results instance graph and the Error Graph of different images sequence

Wherein:

Fig. 1 is the tracking results certain embodiments figure of image sequence ShopAssistant2cor: frame 1,160,180,210,250,300(is from top to bottom with left-to-right)

Fig. 2 is the tracking results certain embodiments figure of image sequence OneLeaveShopReenter2cor: frame 178,195,215,300,390,500(is from top to bottom with left-to-right)

Fig. 3 is the tracking results certain embodiments figure of image sequence pktest02: frame 87,110,130,150,170,200(is from top to bottom with left-to-right)

Fig. 4 is the tracking results certain embodiments figure of image sequence Race1: frame 1,50,100,180,230,300(is from top to bottom with left-to-right)

Fig. 5 represents the relative frame number figure of the relative error of image sequence ShopAssistant2cor

Fig. 6 represents the relative frame number figure of the relative error of image sequence OneLeaveShopReenter2cor

Fig. 7 represents the relative frame number figure of the relative error of image sequence pktest02

Fig. 8 represents the relative frame number figure of the relative error of image sequence Race1

Embodiment:

Below in conjunction with accompanying drawing, embodiments of the invention are elaborated: the present embodiment premised on technical solution of the present invention under implement, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

The object of the present embodiment be first test target exist block and background interference time tracking power, tracking power when secondly target scene being existed to a certain degree illumination variation.The approach taked is panel region object candidate area being divided into multiple overlap, the topological structure building these panel region describes, then provide the similarity of the topological structure of candidate and the topological structure of template based on Pasteur's tolerance, thus determine the weight of the corresponding particle in candidate region.The last tracking results obtaining target according to MAP estimation criterion (MAP).Before telling about concrete method and implementing, first provide local restriction the least square estimation method, and the topological structure how being applied to target describes.As follows:

Local restriction least-squares estimation

Many fields that rarefaction representation is comprising visual target tracking are used widely.Based on sparse hypothesis, in Setting signal y dictionary Φ, a small amount of base vector linear-apporximation table goes out, i.e. y ≈ Φ α, and wherein factor alpha is by l below ₁regularization least square optimization problem

$\underset{\mathrm{\&alpha;}}{\min }\left\{J\left(\mathrm{\&alpha;}\right)\right\}=\underset{\mathrm{\&alpha;}}{\min }\frac{1}{2}\{{\left|\right|y-\mathrm{\&Phi;\&alpha;}\left|\right|}_2^2+\mathrm{\&eta;}{\left|\right|\mathrm{\&alpha;}\left|\right|}_1\},$

Obtain, the openness of coefficient is controlled by parameter η here.

In tracking application, because similarity is more important than openness, when in regularization term similarity being incorporated into above formula, and do its norm and relax, the optimization problem of above formula just becomes

s.t.1 ^Tα=1,

This formula is exactly local restriction uniform enconding problem (LLC) [J.Wang, J.Yang, K.Yu, F.Lv, T.Huang, andY.Gong.Locality-constrainedlinearcodingforimageclassi fication.CVPR, pp.3360-3367,2010.].Here represents the product of Element-Level, for Local actuators, it is according to composing power with the similarity of input vector.Dist (y, Φ)=[|| y-φ ₁|| ₂..., || y-φ _m|| ₂] ^tand φ _ifor in the solution of the i-th row .LLC of Φ, a small amount of component value is important, be in this sense have openness.

LLC coding can by utilizing multiple base vector to obtain less reconstructed error.And for identical signal, its coding has similar pattern.In addition, it has analytic solution.Due to the effect of Local actuators, in fact LLC can choose sparse basis set incompatible formation Local coordinate system for y.One fast approximate solution can realize by solving a less linear system B, B forms with the k of y arest neighbors vector in Φ.

Local restriction least-squares estimation embody rule in this article

From To Template, extract the panel region of multiple overlap, each panel region RGB of the vectorization of its correspondence or gray scale (general designation color) feature describe, and are designated as S={s _i∈ R ^d| i=1 ..., M}.Relationship description between these panel region is as follows:

By other remaining panel region information, Linearly Representation is come to the vectorization information of each panel region here the quick approximation method acquisition of local restriction least-squares estimation is just selected to show out coefficient,

Make A=(| a _ij|), its diagonal entry is 0, the topological structure in this matrix reflection template between panel region information.

To particle same work is also done in corresponding candidate region, namely extracts the panel region obtaining overlapping vectorization sheet information can be expressed as in advance with the relation of remaining sheet information its corresponding sparse solving is obtained by local restriction least-squares estimation.Obtain matrix B ^(l).

Next, two matrix A and B is weighed based on Pasteur's coefficient ^(l)similarity.Pasteur's coefficient is the approximate measure of the degree of overlapping of two statistical samples, is used to the relative similarity degree determining two samples considered.Make A _colwith be respectively A and B ^(l)col row, and to each row to normalization, namely so, this two matrix Pasteur coefficient is defined as $\mathrm{sim}(A,B^{\left(l\right)})=\underset{\mathrm{col}}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}\sqrt{B_{\mathrm{col},j}^{\left(l\right)}{A}_{\mathrm{col},j}}.$

Matlab language is adopted to realize this method and test on the video sequence of reality.At initial frame, target area is manually chosen, and target window is adjusted to the image of 21 × 27 sizes, and 3 × 3 pieces that are then divided into multiple overlap is panel region, sets up the topological structure of these panel region.Tracking results utilizes matrix window to show.The comparative approach adopted is the traditional particle filter tracker based on color (PEREZ) [K.Nummiaroa, E.Koller-Meierb, andL.V.Gool.AnAdaptiveColor-BasedParticleFilter [J] .ImageandVisionComputing, vol.21, pp.99-110,2003.].Solid line and dotted line box represent method of the present invention and PEREZ tracker respectively.

First video sequence (ShopAssistant2cor) comes from http:// homepages.inf.ed.ac.uk/rbf/CAVIARDATA1/, as Fig. 1.In this sequence, the people that follow the tracks of certain section by another wear different people block.PEREZ tracker and tracker of the present invention can capture the human body of movement.But the former does not have on tracking performance, and the latter's is accurate, and this can be seen at the 60th frame.

Second video sequence comes from ttp: //homepages.inf.ed.ac.uk/rbf/CAVIARDATA1/, as shown in Figure 2.A women is blocked by the people wearing color clothes identical.PEREZ tracker has just started to follow the tracks of target.When blocking generation, this tracker has got on regard to the people navigating to this clothing color identical.This is because histogram feature is global characteristics, insensitive to the change of object construction.And tracker of the present invention can provide the tracking results of the degree of accuracy of robust.

3rd video sequence (pktest02) come from // www.ist.temple.edu/hbling/codedata.htm, as shown in Figure 3.Automobile in this video sequence experienced by illumination variation and blocks.PEREZ tracker can the position of roughly target acquisition, but there is moderate drifting problem.Tracker of the present invention well can follow the tracks of the automobile of this movement in whole sequence.

4th video sequence (Race1) comes from [F.Porikli, O.Tuzel.CovarianceTrackingusingModelUpdateBasedonLieAlge bra.CVPR, pp.728-735,2006.], as shown in Figure 4.The racing car followed the tracks of for a moment near video camera for a moment away from video camera, make its afterbody there is dimensional variation.PEREZ tracker well can not estimate the yardstick of vehicle, thus makes location estimation there is skew, and reason is histogram is not invariant to dimensional variation.And tracker of the present invention successfully can follow the tracks of the dimensional variation of vehicle.

Here tracking error utilizes the relative position error at the center of tracking results and actual position to portray, and is defined as ε _k=|| (x _k, y _k)-(x _gk, y _gk) || ₂/ d _gt, wherein { x _gk, y _gk, d _gtbe the position of target demarcated and catercorner length.Desirable tracking should be this alternate position spike is 0.Fig. 5-8 shows the tracking accuracy that institute's extracting method can have during following the tracks of.

Claims (5)

1., based on a visual target tracking method for local restriction least-squares estimation, comprise the steps:

The first step, to template image after resolution adjustment, utilizes a slip wicket, and each horizontal direction or vertical direction slip q pixel, obtain multiple panel region thus, and each panel region color characteristic of the vectorization of its correspondence describes, and is designated as S={s _i∈ R ^d| i=1 ..., M};

Wherein, s _irepresent i-th panel region, R ^drepresent color characteristic set, M represents panel region number;

Second step, sets up a topological structure to the color characteristic set that the first step obtains, regards a summit as by each element, set up an adjacent map; Being calculated as follows of weight on limit in adjacent map:

The panel region of each vectorization represents by other all vectorization panel region, namely here coefficient a _ijprovide by solving local restriction least square problem;

Make A=(| a _ij|), its diagonal entry is 0, the topological structure in this matrix reflection template between panel region information;

3rd step, does the work same with template image to the object candidate area obtained by particle sampler, sets up the topological structure that this candidate region is corresponding; Order l is particle index;

4th step, utilizes Pasteur's coefficient to measure A and B ^(l)similarity;

5th step, obtains the weight of particle according to the measuring similarity of the 4th step, utilize MAP to estimate to obtain the tracking results of target;

In the third step, the multiple object candidate area obtained by particle sampler, set up its topological structure to each candidate target, and specific practice is identical with the topological structure building template, specifically, to particle the panel region of overlapping vectorization is extracted in corresponding candidate region sheet information be expressed as in advance with the relation of all remaining information its corresponding sparse solving is obtained by local restriction least-squares estimation, obtains matrix B ^(l);

The definition of the measuring similarity in the 4th step, is implemented as:

To two topological structure A and B ^(l), make A _colwith be respectively A and B ^(l)col row, and to each row normalization, namely so, this two matrix Pasteur coefficient is defined as $sim(A,B^{\left(l\right)})=\underset{col}{\&Sigma;}\underset{j}{\&Sigma;}\sqrt{B_{col,j}^{\left(l\right)}{A}_{col,j}};$

Wherein, b _i,jrepresent coefficient, represent the coefficient of corresponding particle index l.

2. the visual target tracking method based on local restriction least-squares estimation according to claim 1, it is characterized in that, in a first step, described to template image through resolution adjustment, utilize sub-moving window to obtain the panel region of multiple overlap, in panel region, gray scale or RGB vectorization feature describe the information of this panel region.

3. the visual target tracking method based on local restriction least-squares estimation according to claim 1, feature is, the adjacent map set up between template panel region information described in second step, i.e. topological structure, refer to and utilize local restriction least-squares estimation to encode to other all panel region information of each panel region information, code coefficient is used for weighing other panel region information to the contribution degree of this panel region information.

4. the visual target tracking method based on local restriction least-squares estimation according to claim 1, is characterized in that, in the 5th step, the estimation of target, utilizes MAP estimation to obtain.

5. the visual target tracking method based on local restriction least-squares estimation according to claim 1, is characterized in that: described color characteristic is RGB feature or gray feature.