CN102750550A - Multi-target tracking method and device based on video - Google Patents

Abstract

The invention provides a multi-target tracking method and device based on video. The method comprises the following steps of: extracting a target module, initiating target parameters, detecting Adaboost, carrying out dynamic prediction on a particle set according to a motion model, updating a weight value of each mixing component, updating a motion state of each target, and updating the module. A multi-mode problem in multi-target tracking can be effectively solved, the effectiveness of calculation is maintained, the multi-target tracking can be realized, and simultaneously, the requirement of instantaneity can be ensured.

Description

Multi-object tracking method and device based on video

Technical field

The present invention relates to computer vision and pattern analysis field, particularly a kind of multi-object tracking method and device based on video.

Background technology

Along with the development of video technique, video frequency object tracking becomes popular research topic.It is a main direction in computer vision research field, is the basis that advanced videos such as behavior identification, intelligent video monitoring, human motion analysis are used.

Target following is not a simple question; Because sensation target itself and surrounding enviroment are complicated and changeable; Such as target can receive illumination, block, close background interference, staggered between the target, and the influence of disturbing factor such as the outward appearance of target itself, attitude, form, motion be irregular; Finally cause following the tracks of failure or having than large deviation, therefore setting up stable tracker is to have challenging research topic in the computer vision.And wherein has bigger difficulty for the foundation of multiple-target system.

Face two subject matters during multiple target tracking.At first observation model and Target Assignment are highly non-linear non-Gausses, and secondly, a large amount of tracking targets that change exist because overlapping and uncertain and generation complicated interaction.Under the multiple goal complex environment, follow the tracks of visual signature and face many uncertainties mostly as a whole.

In order to catch the uncertainty of these factors comprehensively; Usually require algorithm to have the multi-mode search capability; Particle filter is as a kind of approximation method of common application; Itself can realize the multi-mode search, and can in global space, find optimal result, therefore the search strategy than monotype has better robustness.

But particle filter is not enough below existing again in the multi-mode search: at first it is more weak on the various modes of keeping Target Assignment continuously.Secondly, with noise or be uncertain when deriving from multiple goal, this causes the various modes growth to dbjective state as the measurement of deficiency.In the particle filter of reality was realized, all particles often took place and merge to a kind of pattern fast, and abandoned every other pattern in the target that will follow the tracks of all existence in such cases, cause particle sampler barren too early, can not realize multiobject tracking.

In order to handle these deficiencies, essential method is exactly to propose novel particle filter.Require novel particle filter can handle the various modes problem well.Calculated amount is little simultaneously, and the validity of calculating is eager to excel.But any simple any particle filter is all not enough to the multiple target tracking processing power of number of variations.

Summary of the invention

To the problem that prior art exists, the present invention aims to provide multi-object tracking method and the device based on video that the advantage that on the basis of novel particle filter, combines Adaboost to detect makes up a kind of ability study, detection and tracking interesting target.

First aspect present invention provides a kind of multi-object tracking method based on video; Its particle filter that distributes formula to change into a mixing multiple goal distributes; Go on foot the novel particle filter of Recursive Implementation by the Monte Carlo through predicting and upgrading two; Again novel particle filter and Adaboost are detected fusion constructs multiple target tracking device, comprise the steps:

The parameter of A, extraction To Template and initialization target;

B, Adaboost detect;

C, according to motion model, the particle collection is carried out performance prediction;

D, the weights of each mixed components are upgraded;

E, upgrade the motion state of each target;

F, template renewal;

G, end.

Preferably, said steps A specifically comprises the steps:

A1, obtain tracing area;

The motion state of A2, initialization target and particle assembly.

Preferably, said step B specifically comprises the steps:

B1, detect through cascade Adaboost detecting device and whether to have fresh target to occur, and the Gaussian distribution that detects to the center in order to Adaboost generates particle.

B2, from Adaboost detects the abstract image piece, initialization exchange probability principal component analysis (PCA) template renewal device.

Preferably, dynamic model described in the said step C uses Chang Su and random walk to mix the processing that dynamic model adapts to motion, rotation, target sizes variation and blocks mutually.

Preferably, said step D specifically comprises the steps:

D1, mixing Bayes sequential filtering process;

D2, novel particle filter tracker right value update;

Choosing of D3, suggestion density;

The calculating of D4, observation likelihood.

Preferably, said step e specifically comprises the steps:

E1, resample, produce not weighted sample according to importance weight.

E2, unweighted sample is averaged.

E3, resampling.

Preferably, said step F specifically comprises study, upgrades and predicts three key steps.

Second aspect present invention provides a kind of multiple target tracking device based on video, comprises the target deriving means, object initialization device, Adaboost pick-up unit, particle filter tracker device, template renewal apparatus.

A kind of multi-object tracking method and device provided by the invention based on video; Its particle filter that distributes formula to change into a mixing multiple goal distributes; Go on foot the novel particle filter of Recursive Implementation by the Monte Carlo through predicting and upgrading two; Again novel particle filter and Adaboost are detected fusion constructs multiple target tracking device, have following beneficial effect:

1, novel particle filter can effectively be handled multi-mode problem in the multiple target tracking, and only in the calculating of mixed weight-value, this just makes computation amount to the interaction between stuff and other stuff, and has kept the validity of calculating.When realizing multiple target tracking, also can guarantee the requirement of real-time.

2, novel particle filter and Adaboost detect fusion constructs multiple target tracking device has two important feature: at first, it uses Adaboost to rebuild the suggestion density function, has improved the robustness of algorithm.The algorithm performance that merges the suggestion density function of up-to-date observation is compared with the suggestion density of only using the transfer prior density, has good superiority.The second, Adaboost provides a framework that obtains and keep mixing description.Particularly, it can detect target effectively and leave and get into scene.The final multiple target tracking of realizing number of variations.

Description of drawings

Fig. 1 is the process flow diagram of a kind of multi-object tracking method based on video of the present invention;

Fig. 2 is the process flow diagram that extracts the parameter of To Template and initialization target among Fig. 1;

The process flow diagram of Fig. 3 for the weights of each mixed components being upgraded among Fig. 1;

Fig. 4 is the likelihood distribution curve comparison diagram of suggestion density function among Fig. 3;

Fig. 5 is the process flow diagram that upgrades the motion state of each target among Fig. 1;

Fig. 6 is the structural representation of a kind of multiple target tracking device based on video of the present invention.

Embodiment

Further specify technical scheme of the present invention below in conjunction with accompanying drawing and through embodiment:

Please with reference to Fig. 1; The present invention provides a kind of multi-object tracking method based on video; Its particle filter that distributes formula to change into a mixing multiple goal distributes; Through predicting and upgrading the novel particle filter of two step Recursive Implementation, more novel particle filter and Adaboost are detected fusion constructs multiple target tracking device by the Monte Carlo, comprise the steps:

The parameter of A, extraction To Template and initialization target;

B, Adaboost detect;

C, according to motion model, the particle collection is carried out performance prediction;

D, the weights of each mixed components are upgraded;

E, upgrade the motion state of each target;

F, template renewal;

G, end.

Said steps A mainly may further comprise the steps:

A1, obtain tracing area;

The motion state of A2, initialization target and particle assembly.

Said step B mainly comprises the steps:

B1, detect through cascade Adaboost detecting device and whether to have fresh target to occur, and the Gaussian distribution that detects to the center in order to Adaboost generates particle.

B2, from Adaboost detects the abstract image piece, initialization exchange probability principal component analysis (PCA) template renewal device.

Dynamic model described in the said step C uses Chang Su and random walk to mix the processing that dynamic model adapts to motion, rotation, target sizes variation and blocks mutually.

Said step D specifically comprises the steps:

D1, mixing Bayes sequential filtering process;

D2, novel particle filter tracker right value update;

Choosing of D3, suggestion density;

The calculating of D4, observation likelihood.

Said step e specifically comprises the steps:

E1, resample, produce not weighted sample according to importance weight.

E2, unweighted sample is averaged.

E3, resampling.

Said step F specifically comprises study, upgrades and predicts three key steps.

Concrete implementation is following:

Please refer to Fig. 2, the parameter of steps A, extraction To Template and initialization target:

Wherein, said steps A 1, obtain tracing area and be specially:

Tracking target can adopt interactive means appointed area in video, also can obtain tracing area automatically according to the moving object detection of video.Hypothetical target is a center for (x, y), length and width are respectively (sx, rectangular area sy), but do not limit the shape of rectangular area.Set up the clarification of objective template according to these prioris.

The state parameter of initialization target is: the state vector of target is represented by a six-vector: it is turned to by parameter:

wherein; X and y represent the center-of-mass coordinate of rectangle;

and

is the speed of target along x and y, and Sx and Sy are target width and height.

The motion state and the particle collection of said steps A 2, initialization target are specially:

With N sample set

Expression P (X0), wherein $X_{m,0}^{\left(i\right)}~q\left(X_0\right),w_{m,0}^{\left(i\right)}=1/N,$ M=0.

It is particularly important how to obtain, propagate and keep new hybrid representation in the novel particle filter, in the ideal case, for each pattern in target distribution a mixing portion will be arranged.Yet in fact the quantity of pattern is to know seldom in advance in target distribution.And the quantity of pattern can not keep immobilizing, maybe be because uncertain the growth, be decomposed or target appearing and subsiding and fluctuating.Therefore to consider that these fluctuations must constantly recomputate hybrid representation.With (C _t, M)=F (X _t, C _t, M) space process of reconstruction of expression as input particle or current hybrid representation, is rebuild function F and is used initial sample set P (X ₀) the k mean cluster obtain initial hybrid representation, hybrid representation recomputates those because separate the particle collection become and to have disperseed through merging obvious lap at every turn in iterating.

Step B, Adaboost detect:

Step B1 is specially: detect target through cascade Adaboost detecting device, if M is arranged _NewIndividual fresh target occurs, then for m from 1 to M _New, i.e. M=M+M _New, the Gaussian distribution that detects to the center through Adaboost produces N particle

Step B2: abstract image piece y from Adaboost detects _{M, t}Initialization SPPCA template renewal device.

Step C, according to motion model, the particle collection is carried out performance prediction:

Hybrid dynamic model is described as:

where

zero-mean Gaussian white noise.

Please refer to Fig. 3, step D, the weights of each mixed components upgraded:

Please, detect the suggestion density that merges mutually with particle filter according to Adaboost with reference to Fig. 3 and Fig. 4.

In the right value update step of novel particle filter, comprise mix Bayes's sequential filtering process D1, novel particle filter tracker right value update D2, suggestion density choose D3, and the several committed steps of calculating D4 of observation likelihood.

Wherein mixing Bayes's sequential filtering D1 is specially:

Suppose x _tThe state vector of expression target, y _1:t=(y ₁Y _t) be observation vector up to time t.The distribution of target is exactly to realize filtering distribution p (x for tracking problem _t| y _1:t).In Bayes's sequence estimation, this distribution uses two step autoregressions to calculate:

Prediction: p (x _t| y _1:t-1)=∫ D (x _t| x _T-1) p (dx _T-1| y _1:t-1) (1)

Upgrade: $p\left(x_t\right|y_{1:t})=\frac{L\left(y_t\right|x_t)p\left(x_t\right|y_{1:t-1})}{\∫L\left(y_t\right|S_t)p\left({\mathrm{Ds}}_t\right|y_{1:t-1})}---\left(2\right)$

Wherein, prediction distribution is followed marginalisation and is distributed, and it is the direct result of Bayes rule that new filtering distributes.

Autoregression requires the dynamic model D (x of concrete description state transitions _t| x _T-1) and according to current observation L (y _t| x _t) provide the model of free position likelihood, and according to some initial distribution p (x ₀) come initialization.

In order to follow the tracks of multiple goal, filtering is resolved into the imparametrization mixture model of M component, its posteriority distribution p (x _t| y _1:t) be:

$p\left(x_t\right|y_{1:t})=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\π}}_{m,t}{p}_m\left(x_t\right|y_{1:t})---\left(3\right)$

Wherein

is assumed that individual mixed components for non-parametric model.The autoregression of imparametrization mixture model under same form upgraded derivation and is divided into prediction and upgrades the realization of two steps.

Suppose mixed filtering distribution p (x _T-1| y _1:t-1) known, new prediction distribution substitution formula (1), derive:

$p\left(x_t\right|y_{1:t-1})=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\π}}_{m,t-1}\∫D\left(x_t\right|x_{t-1})p_m\left({\mathrm{dx}}_{t-1}\right|y_{1:t-1})$

$=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\π}}_{m,t-1}{p}_m\left(x_t\right|y_{1:t-1})---\left(4\right)$

P wherein _m(x _t| y _1:t-1)=D (x _t| x _T-1) p _m(dx _T-1| y _1:t-1) be the prediction distribution of m component.Therefore new prediction distribution directly obtains through the prediction distribution of each independent component, on the basis that keeps the original component weights, merges these distributions then.

Obtain new filtering and distribute, new prediction distribution substitution formula (2), derive:

$p\left(x_t\right|y_{1:t})=\frac{{\mathrm{\Σ}}_{m=1}^M{\mathrm{\π}}_{m,t-1}L\left(y_t\right|x_t)p_m\left(x_t\right|y_{1:t-1})}{{\mathrm{\Σ}}_{n=1}^M{\mathrm{\π}}_{n,t-1}\∫L\left(y_t\right|s_t)p_n\left({\mathrm{ds}}_t\right|y_{1:t-1})}$

$=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\left[\frac{{\mathrm{\π}}_{m,t-1}\∫L\left(y_t\right|s_t)p_m\left({\mathrm{ds}}_t\right|y_{1:t-1})}{{\mathrm{\Σ}}_{n=1}^M{\mathrm{\π}}_{n,t-1}\∫L\left(y_t\right|s_t)p_n\left({\mathrm{ds}}_t\right|y_{1:t-1})}\right]\×\left[\frac{L\left(y_t\right|x_t)p_m\left(x_t\right|y_{1:t-1})}{\∫L\left(y_t\right|s_t)p_m\left({\mathrm{ds}}_t\right|y_{1:t-1})}\right]$

$=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\π}}_{m,t}{p}_m\left(x_t\right|y_{1:t})---\left(5\right)$

Second the new filtering that can regard m component as in the second row bracket distributes.That is:

$p_m\left(x_t\right|y_{1:t})=\frac{L\left(y_t\right|x_t)p_m\left(x_t\right|y_{1:t-1})}{\∫L\left(y_t\right|s_t)p_m\left({\mathrm{ds}}_t\right|y_{1:t-1})}.$

First is state vector x in the bracket _tBe independently, its new weights are:

${\mathrm{\π}}_{m,t}=\frac{{\mathrm{\π}}_{m,t-1}\∫L\left(y_t\right|s_t)p_m\left({\mathrm{ds}}_t\right|y_{1:t-1})}{{\mathrm{\Σ}}_{n=1}^M{\mathrm{\π}}_{n,t-1}\∫L\left(y_t\right|s_t)p_n\left({\mathrm{ds}}_t\right|y_{1:t-1})}=\×\frac{{\mathrm{\π}}_{m,t-1}{p}_m\left(y_t\right|y_{1:t-1})}{{\mathrm{\Σ}}_{n=1}^M{\mathrm{\π}}_{n,t-1}{p}_n\left(y_t\right|y_{1:t-1})}---\left(6\right)$

Formula (3) shows that it is again the mixing that each individual component filtering distributes that new filtering distributes.Promptly, the weights that mix just can obtain correct target distribution as long as upgrading according to formula (6).New weights are standardized component weights likelihood functions, and M different likelihood distribution { p promptly arranged _m(y _1:t| x _t) m=1...M, when one or more fresh targets appeared in the scene, these likelihoods distributed and can obtain through detection or observation model auto-initiation.

Step D2, novel particle filter tracker right value update are specially:

Following is the derivation of detailed novel particle filter tracker right value update recurrence.

Suppose P _t={ N, M, Π _t, X _t, W _t, C _tMixed filtering distributes in the representation formula (3) particle describes.N is the number of particle; M is that the number of mixed components is the multiple goal number;

is the weights of mixed components;

representes particle state; expression particle weights; The mark of

expression component; Promptly

if particle i belongs to m mixed components, and then the Monte Carlo that distributes of

stuff and other stuff filtering is approximately:

$\stackrel{\‾}{p}\left(x_t\right|y^t)=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\π}}_{m,t}\underset{i\∈I_m}{\mathrm{\Σ}}{w}_t^{\left(i\right)}{\mathrm{\δ}}_{x_t^i}\left(x_t\right)---\left(7\right)$

δ wherein _a() is that dirac measures,

It is the particle indexed set that particle belongs to m mixed components.The weights of mixed components are 1 with each particle weights summation, promptly

With $\underset{i\∈I_m}{\mathrm{\Σ}}{w}_t^{\left(i\right)}=1,m=1...M.$

According to p (x _T-1| y _1:t-1) the likelihood given particle collection P that distributes _T-1, purpose is to calculate new particle collection P _tMake it from p (x _t| y _1:t) middle sample.By knowing that each mixed components is independently launched in the general combined tracking recurrence of a last joint, the interaction of mixed components is only calculated through mixed weight-value.Mixed components is also used identical method representation particle for particle filter.With regard to m component, sample

Be from p _m(x _T-1| y _1:t-1) the weights sample set that obtains.New sample will produce through suitably selecting suggestion to distribute; And the selection that suggestion distributes and old state and new measurement are relevant; Promptly

in order to keep weighted sample collection rightly, new particle weight setting is:

$w_t^{\left(i\right)}=\frac{{\tilde{w}}_t^{\left(i\right)}}{{\mathrm{\Σ}}_{J\∈I_m}{\tilde{w}}_t^{\left(J\right)}},{\tilde{w}}_t^{\left(i\right)}=\frac{w_{t-1}^{\left(i\right)}L\left(y_t\right|x_t^{\left(i\right)})D\left(x_t^{\left(i\right)}\right|x_{t-1}^{\left(i\right)})}{q\left(x_t^{\left(i\right)}\right|x_{t-1}^{\left(i\right)},y_t)}---\left(8\right)$

Stuff and other stuff filtering begins to carry out from the prior density sampling, and deriving through formula (8) obtains more new particle of importance weight.

Step D3, choosing of density of suggestion are specially:

The process of choosing of suggestion density

is: stuff and other stuff filtering merges the Adaboost detection in suggestion density framework in right value update.The expression formula of suggestion density function is following:

$q\left(x_t^{\left(i\right)}\right|x_{t-1}^{\left(i\right)},y_t)={\mathrm{\α}}_{\mathrm{ada}}{q}_{\mathrm{ada}}\left(x_{m,t}^{\left(i\right)}\right|y_t)+(1-{\mathrm{\α}}_{\mathrm{ada}})D\left(x_t^{\left(i\right)}\right|x_{t-1}^{\left(i\right)})---\left(9\right)$

Wherein, q _AdaBe to be the Gaussian distribution at center with the Adaboost detection, variance is changeless.Parameter alpha _AdaCan dynamically set and not influence the convergence of particle filter.

Wherein likelihood Distribution calculation 244 is specially:

The calculating of step D4, observation likelihood is specially:

The calculation procedure that said likelihood distributes is:

(1) color model

Because HSV isolates brightness from color (being color (Hue) and purity (Saturation)) (be lightness (Value); So its relative RGB color histogram is to illumination effect and insensitive; Therefore adopt based on the histogrammic color observation model of hsv color, please with reference to Fig. 4.The equal number of subspace distribute to(for) each color component.Be N _h=N _s=V _vThe HSV histogram is by N=N _hN _s+ N _vIndividual histogram is formed, N _h, N _s, N _vBe respectively form and aspect, purity and lightness subspace number.Use b _k(u) ∈ 1 ..., N} representes and at the color vector y of k frame position u _kThe histogram index of (u) being correlated with.

Suppose R (x _t) be the center in the position x _tTwo-dimensional rectangle zone.Then at x _tCandidate region R (the x of position _t) color histogram of a l model is by Q (x _t)={ q ^l(n; x _t) _{N=1 ... N}Expression, wherein:

$q^l(n;x_t)=C^l{\mathrm{\Σ}}_{k\∈R\left(x_t\right)}\mathrm{\δ}[b_t\left(u\right)-n]$

${\mathrm{\Σ}}_{n=1}^N{q}^l(n;x_t)=1---\left(10\right)$

δ is that the Crow Buddhist nun restrains function, C in the formula ^lBe to make

Generalized constant, u is region R (x _t) interior any pixel.Through standardization color histogram Q (x _t) be a discrete probability distribution.Based on the tracking of color be search with the reference target model in the candidate region that belongs to of the most similar histogram.The object module of l model is:

Q ^*＝p ^l(n；x ₀) _n＝1...N

Represent that with the Bhattacharyya coefficient two similaritys between the single histogram measure:

$\mathrm{\ρ}[Q\left(x_t\right),Q^{*}]=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{p(n;x_0)q(n;x_t)}---\left(12\right)$

Distance table between two histograms is shown:

$d(x_t,x_0)=\sqrt{1-\mathrm{\ρ}[Q\left(x_t\right),Q^{*}]}---\left(13\right)$

Obtain about after the color histogram distance B, the definition similarity is distributed as:

$p_{\mathrm{hsv}}\left(y_t\right|x_t)\∝\mathrm{ex}{p}^{-\mathrm{\λ}{d}^2}(x_t,x_0)---\left(14\right)$

Get empirical value λ=20 in the experiment.

(2) edge orientation histogram

Edge orientation histogram is a kind of statistical nature that is used for describing picture shape, edge and texture, and the edge and the texture information of reflection target that can be concrete can be with reference to Fig. 4.It is shape, edge and a texture features of describing target through the edge of statistical objects and texture information.Under the tracking situation, the characteristics of edge orientation histogram are following: a: it comprises the profile information of target, has embodied the structural information of target.Use edge statistics can tackle the variation of field color and illumination certainly rightly;

B: this characteristic under background interference and the illumination variation, has robustness preferably at partial occlusion;

C: two dimension moved with convergent-divergent have unchangeability.These characteristics of edge orientation histogram can be used to remedy the deficiency of solid color information.

The computation process of edge orientation histogram observation likelihood is: the two-value edge point diagram that at first adopts Canny boundary operator computed image; Calculate the edge direction gradient of each marginal point in the binary map again, the proportion that counts each edge direction at last obtains its histogram.(i is that histogrammic horizontal ordinate generates histogram H j) with θ _uThe edge orientation histogram expression formula is:

$H_u=\underset{i,j}{\mathrm{\Σ}}A(i,j)\mathrm{\δ}[\mathrm{\θ}(i,j)-u]---\left(15\right)$

In the formula, $\mathrm{\δ}[x-u]=\left\{\begin{array}{cc}1& x=u\\ 0& x\≠u\end{array}\right..$

Wherein (i j) is gradient magnitude to A, and (i j) is gradient direction to θ.(i, span j) is divided into the u equal portions and quantizes with θ.The value of u is big more, and calculated amount is big more, and computational accuracy is also high more simultaneously.

With after the histogram H normalization as a distribution of probability density, and suppose that it has m gradient direction interval, then the definition edge orientation histogram object module expression formula that is used to follow the tracks of is:

$\hat{H}=\left\{{\hat{H}}_u\right\}u=1...m,$ ${\hat{H}}_u=C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}k\left({\left|\right|x_i^{*}\left|\right|}^2\right)\mathrm{\δ}[b\left(x_i^{*}\right)-u]---\left(16\right)$

The candidate target model:

${\hat{H}}^{*}=\{{\hat{H}}_u^{*}{\}}_{u=1...m}$ ${\hat{H}}_u^{*}=C_h\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)\mathrm{\δ}[b\left(x_i^{*}\right)-u]---\left(17\right)$

C and C _hThe same C of definition ^l

Order

Expression H and H ^*Similarity, then the distance between two edge orientation histograms is: $D\left(y\right)=\sqrt{1-H[\hat{H}\left(y\right),{\hat{H}}^{*}]}---\left(18\right)$

Obtain about after the edge orientation histogram distance B, the definition similarity is distributed as:

$P_{\mathrm{EOH}}\left(y_t\right|x_t)\∝{\mathrm{Exp}}^{-\mathrm{\λ}{D}^2[\hat{H}\left(y\right),{\hat{H}}^{*}]}$ (19) finally obtain Fusion of Color through the weighting rule and be with the observation likelihood function of shape:

$L\left(y_t\right|x_t^i)\∝P_{\mathrm{hsv}}\left(y_t\right|x_t^i)P_{\mathrm{EOH}}\left(y_t\right|x_t^i)---\left(20\right)$

Please refer to Fig. 5, step e, upgrade the motion state of each target:

Step E1: According importance weights

resampling

generate unweighted sample

Step e 2, unweighted sample is averaged:

E3, resampling:

For fear of the degeneration of weights however the time the resampling particle.The standard weight sampling is insensitive to particle position, and this just causes target in distribution, to be lost.Because method with mixed model allows each mixed components independently to resample according to component particle weights among the present invention, reservation posteriority that so just can be very natural distributes.Follow this process, new particle weights become

Step F, template renewal:

(A) to

is the center of the image block extraction

(2) SPPCA template renewal device U _mThrough

Use RBPF to upgrade { z _{M, t}, s _{M, t}.

Step G: finish

When the AdaBoost degree of belief is lower than certain given threshold values, remove M ₁Individual target., two targets merge M when overlapping each other ₂To target, this has M=M-M ₁-M ₂

Please refer to Fig. 6, the target tracker based on video provided by the invention mainly comprises with the lower part:

Object initialization device 10 is set up the color histogram template of target, edge orientation histogram template and unique point template, and the motion state of initialization target and particle assembly respectively.

1) the Adaboost detector means 20, i.e. the target deriving means.Mainly form by three parts: one of which, by color model and edge orientation histogram character representation target, and realize the quick calculating of character numerical value with integrogram; Its two, pick out the rectangular characteristic that some can represent target by the AdaBoost algorithm, i.e. Weak Classifier removes to construct a strong classifier according to the mode of weighting ballot through Weak Classifier then; Its three, some strong classifiers series connection that is obtained by training constitutes the range upon range of sorter of a cascade structure.

2) novel particle filter apparatus 30, comprising: dbjective state prediction, right value update, resampling.

Whether 3) the dbjective state updating device 40, confirm the position of target particle assembly to be resampled with weighted criterion, and judge to follow the tracks of and finish.If follow the tracks of to finish then withdraw from, if do not finish then next frame is proceeded to handle.

4) the template renewal apparatus 50, with exchanging probability principal component analysis (PCA) template renewal device through study, prediction and the realization template renewal of more newly arriving.

A kind of multi-object tracking method and device provided by the invention based on video; Its particle filter that distributes formula to change into a mixing multiple goal distributes; Go on foot the novel particle filter of Recursive Implementation by the Monte Carlo through predicting and upgrading two; Again novel particle filter and Adaboost are detected fusion constructs multiple target tracking device, have following beneficial effect:

1, novel particle filter can effectively be handled multi-mode problem in the multiple target tracking, and only in the calculating of mixed weight-value, this just makes computation amount to the interaction between stuff and other stuff, and has kept the validity of calculating.When realizing multiple target tracking, also can guarantee the requirement of real-time.

2, novel particle filter and Adaboost detect fusion constructs multiple target tracking device has two important feature: at first, it uses Adaboost to rebuild the suggestion density function, has improved the robustness of algorithm.The algorithm performance that merges the suggestion density function of up-to-date observation is compared with the suggestion density of only using the transfer prior density, has good superiority.The second, Adaboost provides a framework that obtains and keep mixing description.Particularly, it can detect target effectively and leave and get into scene.The final multiple target tracking of realizing number of variations.

Combine accompanying drawing that the present invention has been carried out exemplary description above; Obvious realization of the present invention does not receive the restriction of aforesaid way; As long as the various improvement of having adopted technical scheme of the present invention to carry out; Or design of the present invention and technical scheme are directly applied to other occasion without improving, all in protection scope of the present invention.

Claims (8)

1. multi-object tracking method based on video; Its particle filter that distributes formula to change into a mixing multiple goal distributes; Go on foot the novel particle filter of Recursive Implementation by the Monte Carlo through predicting and upgrading two; Again novel particle filter and Adaboost are detected fusion constructs multiple target tracking device, it is characterized in that, comprise the steps:

The parameter of A, extraction To Template and initialization target;

B, Adaboost detect;

C, according to motion model, the particle collection is carried out performance prediction;

D, the weights of each mixed components are upgraded;

E, upgrade the motion state of each target;

F, template renewal;

G, end.

2. the multi-object tracking method based on video according to claim 1 is characterized in that said steps A specifically comprises the steps:

A1, obtain tracing area;

The motion state of A2, initialization target and particle assembly.

3. according to right 1 described multi-object tracking method, it is characterized in that said step B specifically comprises the steps: based on video

B1, detect through cascade Adaboost detecting device and whether to have fresh target to occur, and the Gaussian distribution that detects to the center in order to Adaboost generates particle.

B2, from Adaboost detects the abstract image piece, initialization exchange probability principal component analysis (PCA) template renewal device.

4. according to right 1 described multi-object tracking method, it is characterized in that dynamic model described in the said step C uses Chang Su and random walk to mix the processing that dynamic model adapts to motion, rotation, target sizes variation and blocks mutually based on video.

5. according to right 1 described multi-object tracking method, it is characterized in that said step D specifically comprises the steps: based on video

D1, mixing Bayes sequential filtering process;

D2, novel particle filter tracker right value update;

Choosing of D3, suggestion density;

The calculating of D4, observation likelihood.

6. according to right 1 described multi-object tracking method, it is characterized in that said step e specifically comprises the steps: based on video

E1, resample, produce not weighted sample according to importance weight.

E2, unweighted sample is averaged.

E3, resampling.

7. according to right 1 described multi-object tracking method, it is characterized in that said step F specifically comprises study, upgrades and predict three key steps based on video.

8. the multiple target tracking device based on video is characterized in that, comprises the target deriving means, object initialization device, Adaboost pick-up unit, particle filter tracker device, template renewal apparatus.

Images