CN102831409A - Method and system for automatically tracking moving pedestrian video based on particle filtering - Google Patents

Abstract

The invention discloses a method and a system for automatically tracking a moving pedestrian video based on particle filtering. The method comprises the following steps of: inputting one frame of images, and carrying out detection through an HOG (Histogram of Oriented Gradient) feature vector set and an SVM (Support Vector Machine) vector machine; in order to realize particle filtering tracking based on double HOG and color features, firstly obtaining an initial rectangular area of target pedestrian, sampling a plurality of particles from a target rectangular area, extracting an HOG feature and a color feature, computing the weight of the particles after the double HOG and color features are fused, obtaining the final state estimation through a minimum mean square error estimator, outputting an estimation target and then resampling; and closely locking the tracked target pedestrian. The method extracts the double HOG and color features to increase the robustness of a particle filtering likelihood model and eliminate the unstable situation in the tracking process, the method combines the HOG feature to build the better likelihood model through a fusion strategy of weighted mean, the robustness of the tracking algorithm is greatly increased, and the stable tracking is completed.

Description

Motion pedestrian video automatic tracking method and system based on particle filter

Technical field

The application relates to technical field of computer vision, relates in particular to a kind of motion pedestrian video automatic tracking method and system based on particle filter.

Background technology

The tracking of moving object is exactly on the continuous images sequence, the corresponding matching problem of features relevant such as the position that the pedestrian of motion is occurred, size, shape.Numerous scholars have proposed many algorithms.Existing moving body track algorithm mainly contains following four kinds: based on the tracking of template matches, based on the tracking of profile, based on the tracking and the particle filter tracking method of motion prediction.

Wherein, the particle filter tracking method is in the real world applications scene, and the noise of image is disobeyed Gaussian distribution usually, and therefore, Kalman filtering can not obtain tracking effect preferably, in order to be applied to the scene of real world applications.Prior art is incorporated into the vision track field with particle filter algorithm.

The main thought of particle filter is based on monte carlo method; It is to utilize the particle collection with weight to represent posterior probability; Can be applied on any type of state-space model, be a kind of order importance sampling method (Sequential Importance Sampling, SIS).In simple terms, the particle filter method is meant through seeking one group of random sample of propagating at state space probability density function is similar to, and replaces integral operation with sample average, thus the process that the state of acquisition minimum variance distributes.

Particle filter is to nonlinear motion, the situation that multi-mode distributes.Posterior probability distribution estimating value through to former frame is sampled, and propagates the posterior probability estimation value that these sampled values form present frame then.The shortcoming of particle filter is for guaranteeing that the required sampled point of correctness that current state is carried out maximal possibility estimation causes calculated amount excessive too much.

Another weak point of particle filter is the particle degradation phenomena, and the main point of penetration that present numerous scholars improve particle filter also is to start with from solving the particle degradation phenomena.Prior art is adjusted the particle filter sampling policy through introducing average drifting, and uses integration histogram to accelerate the histogrammic calculating of each particle, improves the speed and the precision of track algorithm; Wu Tao etc. have proposed to improve algorithm based on the particle filter of MCMC method, use the MCMC method to choose the performance that sampling policy preferably improves track algorithm; And carry out the improvement of track algorithm through abandoning little weight and making full use of the principle that the meaning of particle weight size representative duplicates.Though these improvement are arranged, in pedestrian's tracking, do not see relevant report, also there are many technical barriers to capture.

In sum, be necessary to provide a kind of motion pedestrian video automatic tracking method and system based on particle filter to address the above problem.

Summary of the invention

In view of this, the present invention provides a kind of motion pedestrian video automatic tracking method and system based on particle filter, effectively raises the robustness of track algorithm, accomplishes stable tracking.

To achieve these goals, the technical scheme that provides of the application embodiment is following:

A kind of motion pedestrian video automatic tracking method based on particle filter said method comprising the steps of:

S1, input one two field picture detect through HOG set of eigenvectors and SVM vector machine, have judged whether the pedestrian, if, execution in step S2, if not, input next frame image detects again;

S2, realize particle filter tracking based on HOG and color double characteristic; At first obtain target pedestrian's initial rectangular zone; And the plurality of particles of from the target rectangle zone, sampling; Extract HOG characteristic and color characteristic, calculate the weight that HOG and color double characteristic merge the back particle, resample after getting state estimation to the end and export estimating target through the least mean-square error estimator;

S3, judge whether image is last frame, if, then finish to follow the tracks of, if not, return step S2.

As further improvement of the present invention, " detecting through the HOG set of eigenvectors " among the said step S1 is specially:

Evaluate image pixel characteristic in gray space or color space is carried out the gamma correction standardization to image;

According to the extraction of HOG gradient, calculate the gradient of each pixel;

The computing unit gradient magnitude is specially the gradient orientation histogram that comprises pixel in each unit that adds up, and each gradient orientation histogram is mapped on definite angle again, obtains the HOG proper vector;

A piece is formed in adjacent unit, carries out piece normalization;

Select unit and piece for use, obtain the HOG set of eigenvectors and detect.

As further improvement of the present invention, the SVM vector machine among the said step S1 comprises the svm classifier device, and said svm classifier device comprises traditional svm classifier device and linear inseparable svm classifier device.

As further improvement of the present invention, the classification function in said traditional svm classifier device is:

Classification function in the linear inseparable svm classifier device is: $f\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_{i}K(x\·x_i)+b^{*}.$

As further improvement of the present invention, said step S2 is specially:

Obtain target pedestrian's initial rectangular zone, and each particle weight of the plurality of particles of from the target rectangle zone, sampling initialization is 1/N;

Through state equation x _t=Ax _T-1+ Bw _T-1Prediction obtains the t state of particle constantly

Calculate the weight that HOG and color double characteristic merge the back particle:

$w^i=\frac{\mathrm{colcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_c^i+\frac{\mathrm{hogcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_h^i;$

Get state estimation to the end through the least mean-square error estimator:

The output estimating target also resamples.

As further improvement of the present invention, the extraction of color characteristic is specially among the said step S2:

Calculate color histogram at rgb space, the probability distribution of target area is: $q\left(x\right)=C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}k\left(\right|\left|\frac{x-x_i}{h}\right|\left|\right)\mathrm{\&delta;}\left(b(x_i-u)\right),u=1\&CenterDot;\&CenterDot;\&CenterDot;m,$ Wherein, the state variable of said target candidate target area is X, and the center is x, and δ is a Kronecker delta function, and n is the number of said target area pixel, function b:R ²→ { 1...m} distributes the bin of a correspondence for each pixel, and C is a constant, and the normalization formula does. $C=1/{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|x_i\left|\right|}^2\right),$ K representes kernel function, and $k\left(x\right)=\left\{\begin{array}{c}1-x^2:x<1\\ 0:\mathrm{Otherwise}\end{array}\right.;$

To Template p is according to candidate template

Calculate, calculate similarity apart from doing

Then according to the observed reading of similarity distance calculation particle $p\left(z\right|x_t)=\frac{1}{\sqrt{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}}\mathrm{Exp}\left(\frac{-{\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="HDA00002078203700021.png,HDA00002078203700022.png"\; state="\{\{state\}\}">d</figure-callout>}}^2}{2{\mathrm{\&sigma;}}^2}\right).$

As further improvement of the present invention, the HOG Feature Extraction is specially among the said step S2:

Use single order center operator [1,0,1], the Grad G of each pixel level of computed image and vertical direction _hAnd G _v, wherein:

$G_h(x,y)=f(x+1,y)-f(x-1,y),\&ForAll;x,y,$ $G_v(x,y)=f(x,y+1)-f(x,y-1),\&ForAll;x,y;$

Calculate each pixel intensity M (x, y) with gradient direction θ (x, y), wherein:

$M(x,y)=\sqrt{G_h{(x,y)}^2+G_v{(x,y)}^2},$ θ(x，y)＝tan ^-1(G _h(x,y)/G _v(x，y))；

Gradient intensity is carried out normalization, obtain the histogram of gradients of entire image.

As a further improvement on the present invention, the sampling among the said step S2 is specially:

At the hypercube D of unit _s: [0,1) the last QMC point set { u that produces ⁱ, i=1,2 ... N} utilizes formula x ⁱ=[a+ (b-a) ο u ⁱ] convert the QMC point set into particle collection { x ⁱ, 1,2 ... N}, wherein a and b are x ⁱThe interval in space, place.

Correspondingly, a kind of motion pedestrian video automatic tracking system based on particle filter, said system comprises:

Pre-processing module, said module are utilized HOG and SVM detection algorithm, detect the pedestrian who comprises in the initial frame, and provide pedestrian's rectangular area;

Video capture module is used to accomplish the video capture of video file or camera, and the video of catching is shown to system interface, provides intuitively and understands;

Parameter is provided with module, and said module is utilized the MFC control, provides the user that the interface of track algorithm parameter is set;

Pedestrian's tracking module is used to the method that adopts HOG and color double characteristic to combine, uses the tracking framework of particle filter, accomplishes the tracking to the motion pedestrian.

Can see that by above technical scheme motion pedestrian video automatic tracking method and system based on particle filter of the present invention accomplish the tracking to the motion pedestrian under based on the framework of particle filter.By extracting HOG and color double characteristic; Improve the robustness of particle filter likelihood model; Eliminate unsettled situation in the tracing process; To solid color feature tracking effect under complex background the requirement of problem such as losing and can't satisfy real-time tracking appears easily following the tracks of particularly; Make up better likelihood model in conjunction with the HOG feature by average weighted convergence strategy; Thereby improve the robustness of track algorithm greatly, accomplish stable tracking;

Simultaneously, through studying the characteristic that the particle filter method of sampling obtains the particle collection, we find the particle that produces based on the MC method because its random character causes occurring easily between particle gap and range upon range of; Influence the precision of track algorithm; What is more, causes wave filter to restrain, and can't accomplish tracking.The present invention is through the QMC method of sampling, produces the particle collection that the uniform particle collection of detailed rules and regulations more replaces the MC stochastic sampling to obtain, and eliminates interparticle gap and range upon range of, finally improves the precision of track algorithm.

Description of drawings

In order to be illustrated more clearly in the application embodiment or technical scheme of the prior art; To do to introduce simply to the accompanying drawing of required use in embodiment or the description of the Prior Art below; Obviously, the accompanying drawing in describing below only is some embodiment that put down in writing among the application, for those of ordinary skills; Under the prerequisite of not paying creative work, can also obtain other accompanying drawing according to these accompanying drawings.

Fig. 1 is the particular flow sheet that the present invention is based on the motion pedestrian video automatic tracking method of particle filter;

Fig. 2 is HOG characteristic schematic diagram among the present invention;

Fig. 3 is the synoptic diagram of this traditional svm classifier device optimal interval lineoid;

Fig. 4 (a) is MC method of sampling distribution of particles synoptic diagram among the present invention, and Fig. 4 (b) is QMC method of sampling distribution of particles synoptic diagram among the present invention;

Fig. 5 is the motion pedestrian video automatic tracking system module diagram that the present invention is based on particle filter;

Fig. 6 handles every frame elapsed time comparison diagram for two kinds of trackings in prior art and an embodiment of the present invention;

Fig. 7 is based on the synoptic diagram of the maximum particle weight of color characteristic in the prior art;

Fig. 8 in the embodiment among the present invention based on the synoptic diagram of the maximum particle weight of HOG and color characteristic;

Fig. 9 is the comparison diagram of pursuit path and actual path in prior art and another embodiment of the present invention;

Figure 10 is the comparison diagram in each frame processing time in prior art and another embodiment of the present invention.

Embodiment

In order to make those skilled in the art person understand the technical scheme among the application better; To combine the accompanying drawing among the application embodiment below; Technical scheme among the application embodiment is carried out clear, intactly description; Obviously, described embodiment only is the application's part embodiment, rather than whole embodiment.Based on the embodiment among the application, those of ordinary skills are not making the every other embodiment that is obtained under the creative work prerequisite, all should belong to the scope of the application's protection.

Join shown in Figure 1, a kind of motion pedestrian video automatic tracking method of the present invention based on particle filter, this method may further comprise the steps:

S1, input one two field picture detect through HOG set of eigenvectors and SVM vector machine, have judged whether the pedestrian, if, execution in step S2, if not, input next frame image detects again;

S2, realize particle filter tracking based on HOG and color double characteristic; At first obtain target pedestrian's initial rectangular zone; And the plurality of particles of from the target rectangle zone, sampling; Extract HOG characteristic and color characteristic, calculate the weight that HOG and color double characteristic merge the back particle, resample after getting state estimation to the end and export estimating target through the least mean-square error estimator;

S3, judge whether image is last frame, if, then finish to follow the tracks of, if not, return step S2.

" detecting through the HOG set of eigenvectors " in step S1 is specially:

Evaluate image pixel characteristic in gray space or color space is carried out the gamma correction standardization to image.The evaluate image pixel characteristic is generally at gray space or color space.Choose which kind of realistic space,, in pedestrian detection, to use acquired information can reduce the collection of the information of image, therefore can cause the decline of accuracy of detection for whether accurately display image is most important.Obviously, color space can provide more Useful Informations.In the application scenarios of reality, detected image generally can be hypographous and the interference of exposure, in order to eliminate these disturbing factors, image carried out the gamma correction standardization.Usually, gamma correction has two kinds of different functions: power time function (n sum of powers n th root) and logarithm (logarithm and inverse logarithm) function.Having improved verification and measurement ratio through 2 th root is roughly: in 10-4FPPW (False Positives Per Window), error rate is 1%; And under the logarithm situation, then reduced verification and measurement ratio: at 10-4, error rate is 2%;

According to the extraction of HOG gradient, calculate the gradient of each pixel.In this embodiment, because single order center operator calculated amount is little, calculate simply, therefore, we select for use single order center operator [1,0,1] to come the Grad of calculating pixel;

The computing unit gradient magnitude is specially the gradient orientation histogram that comprises pixel in each unit that adds up, and each gradient orientation histogram is mapped on definite angle again, obtains the HOG proper vector.The statistics of unit cell gradient magnitude is the target information that is used for describing regional area, can reduce the pedestrian because clothing outward appearance and attitude change the influence that brings change in shape.Shown in Fig. 2 (a), the basic thought that gradient magnitude calculates is, comprises the gradient orientation histogram of pixel in each unit that adds up, and each gradient orientation histogram is mapped on definite angle again, obtains proper vector;

A piece is formed in adjacent unit, carries out piece normalization.Adjacent unit cell forms a piece, is called block.Shown in Fig. 2 (b).Can find out that it is overlapping to understand appearance between the adjacent block, and can the raising classifying quality.Will occur a local feature in the unit like this has different results in each block, ginseng Fig. 2 (c) appears in the proper vector with different values through after the normalization.And then the information of description entire image, promptly be the HOG set of eigenvectors;

Select unit and piece for use, obtain the HOG set of eigenvectors and detect.In the testing process, select different cell cell and piece block for use, the detection effect that obtains has very big difference.This paper adopts and proposes in the Dalal literary composition, and the cell size is 8 * 8, and each cell of 2 * 2 constitutes a block, and detection window is 64 * 128, and then obtain the HOG set of eigenvectors.

SVM vector machine among the step S1 comprises the svm classifier device, and the svm classifier device comprises traditional svm classifier device and linear inseparable svm classifier device.

Tradition svm classifier device is meant sensu lato sorter, is under the situation of linear separability, to propose.Supposing has H ₁And H ₂Two training samples, wherein H1 is positive sample, H ₂It is negative sample.Under the condition of linear separability, can find two such lineoid, make between these two lineoid, to have no sample point.And optimum linear classification is exactly to make the distance between these two lineoid maximum, is optimum lineoid, and is specifically as shown in Figure 3.

Wherein the lineoid equation is: wx _i+ b=0;

Simultaneously in order to make the sample number strong point all outside the spacer region of lineoid, we need guarantee for all x _iSatisfy one of them condition: y _i(wx _i+ b)>=1, i=1 ... N;

Suppose x _iDistance to lineoid is D, then has:

Then spacing distance is:

$m\arg \mathrm{in}=\underset{\{x_i,y_i=1\}}{\min }D(w,b:x_i)+\underset{\{x_i,y_i=-1\}}{\min }D(w,b:x_i)$

$=\underset{\{x_i,y_i=1\}}{\min }\frac{|w\&CenterDot;x_i+b|}{\left|\right|w\left|\right|}+\underset{\{x_i,y_i=-1\}}{\min }\frac{|w\&CenterDot;x_i+b|}{\left|\right|w\left|\right|}$

$=\frac{2}{\left|\right|w\left|\right|}$

Can find out from following formula; Want to make spacing maximum; Then making || w|| is minimum; The best lineoid of demand makes

minimum under the condition of lineoid equation.Utilize the optimized solution of quadratic programming can draw classification function to be: $f\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{a}_i(x\&CenterDot;x_i)+b^{*},$ Wherein, a is the Lagrangian factor.

The disadvantage of tradition svm classifier device is exactly the situation that only is applicable to the training sample linear separability, and in the practical application scene, generally all is linear inseparable training data.To this situation, numerous scholars have been developed linear inseparable svm classifier device on the basis of traditional svm classifier device.To inseparable this problem of linearity; Great majority are with in sample space DUAL PROBLEMS OF VECTOR MAPPING to a high-dimensional feature space; Be about to the linear problem that non-linear inseparable problem converts new high-dimensional feature space into, ask the problem of optimal interval lineoid then at high-dimensional feature space, according to finding in the above-mentioned linearly separable function process; We only need the utilization kernel function in the space, to carry out inner product operation; Just can realize linear classification, wherein kernel function satisfies the Mercer condition, and kernel function commonly used mainly contains three kinds:

One, be maximum RBF RBF of usefulness:

$K(x,y)=\exp (-\frac{{\left|\right|x-y\left|\right|}^2}{2{\mathrm{\&sigma;}}^2})$

Two, be the polynomial kernel function:

K(x,y)=(x·y+1) ^d

Three, be to use the Sigmoid function:

K(x,y)=tanh(w·(x·y)+b)

Select suitable kernel function according to application scenarios, can obtain classification function and be:

$f\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{\mathrm{\&alpha;}}_{i}K(x\&CenterDot;x_i)+b^{*}.$

In the prior art, the motion propagation model is the communication process of system state just, and what it was described is the motion process of motion pedestrian along with the time.For any target tracking algorism, all need carry out modeling to the motion of target.Usually tracking field mathematical model commonly used has autoregressive model (Auto regressive), migration model (Random walk) immediately.Because the randomness and the uncertainty of pedestrian's movement locus, so this paper adopts migration model immediately to describe the pedestrian of motion, and equation is: s _t=As _T-1+ w _tWherein, A is a constant, w _tFor separate between system's multivariate Gaussian noise and each variable, s _tExpression target pedestrian is at t location status constantly.

Target pedestrian's motion propagation model is described the pedestrian along with the movement locus of time, is reflected in the particle filter tracking framework, is to describe each particle with which kind of mode of motion to shift and propagate.But the state after propagating also needs observation model to examine its similarity degree.Like this, the design of observation model with regard to the foundation that is equivalent to similar function with found the solution problem.When the estimated value that observed reading and motion are propagated similar more, the state of approaching the motion of the dbjective state of then estimating pedestrian reality.This paper chooses the foundation as motion pedestrian likelihood model of HOG and color characteristic, and the foundation of two kinds of characteristic likelihood model is following.

The foundation of similar function

Suppose that q (x) is the reference model of system model, x representes the center of object module; P representes object module, and we select for use Pasteur's coefficient to weigh two similaritys between the model, and wherein Pasteur's coefficient equation is following:

$\mathrm{\&rho;}(p,q)=\underset{u=1}{\overset{m}{\mathrm{\&Sigma;}}}\sqrt{p^{*}{q}_u\left(x\right)};$

The similarity distance is:

$d=\sqrt{1-\mathrm{\&rho;}(p,q)};$

After obtaining the similarity distance, calculate the observed reading of particle:

$p\left(z\right|x_t)=\frac{1}{\sqrt{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}}\exp \left(\frac{-{\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="HDA00002078203700021.png,HDA00002078203700022.png"\; state="\{\{state\}\}">d</figure-callout>}}^2}{2{\mathrm{\&sigma;}}^2}\right),$

Wherein, σ ²It is scale factor.

Color characteristic extracts

In the video tracking field, color characteristic is a characteristic that is widely used, and mainly is because color characteristic has certain robustness to target deformation, rotation and partial occlusion.Usually, the distribution of color of target area uses the color histogram of discretize to represent that constitutional diagram lattice (bin) is taken as m=16*16*16.This paper calculates color histogram at rgb space, and the state variable of hypothetical target candidate target region is X, and the center is x, and then the probability distribution in the target area is:

$q\left(x\right)=C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}k\left(\right|\left|\frac{x-x_i}{h}\right|\left|\right)\mathrm{\&delta;}\left(b(x_i-u)\right),$ u=1…m，

Wherein δ is a Kronecker delta function, and n is the number of this target area pixel.Function b:R ²→ 1...m} distributes the bin of a correspondence for each pixel, and C is a constant, and the normalization formula is: K representes kernel function, defines as follows:

$k\left(x\right)=\left\{\begin{array}{c}1-x^2:x<1\\ 0:\mathrm{otherwise}\end{array}\right.,$

In like manner, To Template p calculates according to candidate template, and then calculates the similarity distance, obtains the observed reading of particle.

The HOG feature extraction

Histogram of gradients HOG is a kind of characteristic that is used to describe motion for line people shape profile, is often used in pedestrian detection, and obtains quite good detecting effectiveness.Owing to have the better performance of describing pedestrian's characteristic, and single characteristic can't be accomplished the stable tracking to the motion pedestrian, so this chapter introduces robustness and accuracy that the HOG characteristic improves track algorithm on the basis of solid color characteristic.At first target pedestrian zone is divided into each unit that equates and is called cell on the gray scale yardstick; All pixels make up corresponding histogram of gradients according to its Grad and direction in each cell then; At last each histogram is carried out normalization, obtain the histogram of gradients of entire image.

Use single order center operator [1,0,1], come the Grad G of each pixel level of computed image and vertical direction _hAnd G _v:

$G_h(x,y)=f(x+1,y)-f(x-1,y),\&ForAll;x,y,$ $G_v(x,y)=f(x,y+1)-f(x,y-1),\&ForAll;x,y;$

Calculate each pixel intensity M (x, y) with gradient direction θ (x, y):

$M(x,y)=\sqrt{G_h{(x,y)}^2+G_v{(x,y)}^2},$ θ(x，y)＝tan ^-1(G _h(x,y)/G _v(x，y))；

Gradient intensity is carried out normalization, obtain the histogram of gradients of entire image.

The convergence strategy of HOG and color characteristic

According to the aforementioned calculation process; Can obtain in the target pedestrian zone following based on the weighted value of each characteristic, wherein the weight based on color characteristic is

based on the HOG characteristic weighted value is

After obtaining weighted value, we propose the final weights value that a kind of new convergence strategy obtains each pixel, carry out the iteration execution of track algorithm then according to the final weights value.New convergence strategy is at first set a threshold value k, adds up the validity feature number under each characteristic information according to following formula then.

$\mathrm{colcount}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{count}\left(w_c^i\right),$ $\mathrm{hogcount}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{count}\left(w_h^i\right),$

Wherein, $\mathrm{Count}\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="k"\; filenames="HDA00002078203700022.png,HDA00002078203700031.png"\; state="\{\{state\}\}">k</figure-callout>}\\ 0,x<k\end{array}\right..$

According to effective characteristic number, can obtain the final weights value:

$w^i=\frac{\mathrm{colcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_c^i+\frac{\mathrm{hogcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_h^i,$

After obtaining final weight, just can carry out the renewal of particle weight, accomplish the particle filter tracking algorithm.

The ultimate principle of particle filter algorithm is to adopt the imparametrization Monte Carlo simulation to find the solution the integration problem in the Bayesian Estimation according to law of great number, is applicable to the estimation problem under non-linear, the non-Gauss's scene under the reality.

The ultimate principle of Bayesian Estimation is to come the prior probability of tectonic system state with known all information from 1 ~ t-1 moment, again according to t observed reading z constantly _{1: t}Revise, obtain the posterior probability p (x of system _t| z _{1: t}).Usually, the recursive process of Bayes's filtering is respectively prediction and upgrades two steps and accomplish.

The first step, prediction.According to the metastasis model of system state, can be by the state p (x of system state etching system when t-1 probability density function is constantly predicted t _t| z _{1: t-1}).

Get by the Chapman-Kolmogorov equation:

p(x _t|z _1:t-1)=∫p(x _t|x _t-1)p(x _t-1|z _1:t-1)dx _t-1；

Second step: upgrade.At t constantly, calculate corresponding observed reading, utilize Bayesian formula to accomplish prior probability p (x by the observation model of system _t| z _{1: t-1}) to posteriority Probability p (x _t| z _{1: t}) derivation.

$p\left(x_t\right|z_{1:t})=\frac{p\left(z_{1:t}\right|x_t)p\left(x_t\right)}{p\left(z_{1:t}\right)},$

The posterior probability that particle filter comes the approximation system state through one group of particle collection that has a weight.Based on the law of large numbers, when population is very big, this approximate posterior probability that is equal to.P (x then _t| z _{1: t}) can convert into:

$p\left(x_t\right|z_{1:t})\&ap;\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_t^i\mathrm{\&delta;}(x_{0:t}-x_{0:t}^i),$

Wherein N is a population, and

is the weight of particle.Usually the particle collection can't directly be sampled from posterior probability and obtained, and the general using importance sampling obtains from another importance density function of sampling easily.Then the weight correction formula is:

$w_t=w_{t-1}\frac{p\left(z_t\right|x_t)p\left(x_t\right|x_{t-1})}{q\left(x_t\right|x_{0:t-1},z_{1:t})},$

For convenience of calculation, importance function commonly used usually is q (x _t| x _{0: t-1}, z _{1: t})=p (x _t| x _T-1), bring into $w_t=w_{t-1}\frac{p\left(z_t\right|x_t)p\left(x_t\right|x_{t-1})}{q\left(x_t\right|x_{0:t-1},z_{1:t})}$ Formula gets: w _t=w _T-1P (z _t| x _t).

Can get thus, the particle filter density function of making a difference obtains the particle collection, and along with the observed reading iteration is tried to achieve corresponding weights, the mode of final weighted sum is represented the posterior probability of state to obtain final estimated value.

The particle collection of supposing to have weight is X={ (x ⁿ, w ⁿ) | n=1 ... N}, X represent the state of target, and w is the particle corresponding weights, and then posterior probability can obtain through following formula: The value of w promptly is the weights after HOG and color double characteristic merge.

In sum, step S2 can for:

Obtain target pedestrian's initial rectangular zone, and each particle weight of the plurality of particles of from the target rectangle zone, sampling

initialization is 1/N;

Through state equation x _t=Ax _T-1+ Bw _T-1Prediction obtains the t state of particle constantly

Calculate the weight that HOG and color double characteristic merge the back particle:

$w^i=\frac{\mathrm{colcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_c^i+\frac{\mathrm{hogcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_h^i;$

Get state estimation to the end through the least mean-square error estimator:

The output estimating target also resamples.

Point out in the preceding text that the particle filter theory is widely used in the estimation problem under non-linear, non-Gauss's scene, be highly suitable under the environment of complex background and follow the tracks of, and to block, phenomenon such as interference has good robustness.It is based on Monte Carlo (MC) method stochastic sampling and obtains particle, the posterior probability of approaching state through a particle collection that has a weight.Though particle filter has been obtained widely and has been used, the last accuracy of following the tracks of of the de-stabilising effect that when the dealing with complicated problem, possibly become; This mainly is owing to the random character based on Monte Carlo sampling causes differing very big between the particle collection, and the accumulation of exponentially level in the process of estimating, What is more, causes the inefficacy of wave filter.The main performance of this random character is that particle is at the excesssive gap of part or range upon range of in a large number.Compared to monte carlo method, intend the opposite of Monte Carlo (QMC) method as MC, it is formed approximate with random series that basic thought is to use the ultra even distribution series of determinacy of detailed rules and regulations distribution more to make up in the MC method.This Deterministic Methods can access the optimum distributed points in sample space and that is to say the optimal particle collection, avoids the excesssive gap, range upon range of of particle in the MC stochastic sampling process, improves sampling efficiency and precision.

Intend Monte Carlo (QMC) and be in order to seek the error that a kind of point set reduces the Monte Carlo integration, its integrated form is consistent with the Monte Carlo, and only the random number of usefulness is different.QMC mainly is devoted to construct average error will good point set than MC, and the QMC method is the point set immediately that replaces MC with deterministic point set.According to the difference that obtains N the point set method of sampling, just the integration problem is divided into Monte Carlo (MC) method and intends Monte Carlo (QMC) method simultaneously.

MC and the QMC basic thought in being applied to the particle filter tracking algorithm all is to adopt the posteriority sample particles of weighting to approach the posterior probability distribution, will calculate finding the solution with the form of summation of integration.Yet, formal because the random character of the MC method of sampling causes sampling interparticle gap and range upon range of, produce the error of estimating, influenced the effect of particle filter, cause the problem of tracking accuracy decline.The QMC method of sampling then uses deterministic point to eliminate these influences, and it is very little to make that approximate error becomes.Difference according to production method.These deterministic points are called low diversity sequence (low discrepancy sequences) again; On mathematics, mainly can represent: Halton sequence, Sobel sequence and Niederreiter sequence etc. through following several kinds of forms; With respect to other two kinds of sequences; The Halton sequence is easier to understand and realizes, so this paper is selected from the Halton sequence and produces the QMC point set.Shown in Fig. 4 (a) and 4 (b), be the distribution of MC and QMC sampled point, detailed rules and regulations are even more can to find out the distribution of QMC sampling point set, and the point set of MC then has to be assembled and the gap.

Adopt the particle filter tracking algorithm after the QMC sampling improves to be:

Initialization: at the hypercube D of unit _s: [0,1) the last QMC point set { u that produces ⁱ, i=1,2 ... N} utilizes formula x ⁱ=[a+ (b-a) ο u ⁱ] convert the QMC point set into particle collection { x ⁱ, 1,2 ... N}, wherein a and b are x ⁱThe interval in space, place;

The evolution of particle collection: each particle collection that develops, and carry out the resampling and the generation of QMC point set;

Weight is upgraded: w _t=w _T-1P (z _t| x _t);

The average of estimation particle collection

is accomplished and is followed the tracks of.

Correspondingly, join shown in Figure 5ly, a kind of motion pedestrian video automatic tracking system based on particle filter of the present invention comprises:

Pre-processing module 10, module are utilized HOG and SVM detection algorithm, detect the pedestrian who comprises in the initial frame, and provide pedestrian's rectangular area.Concrete implementation procedure is that the HOG detection built-in function that utilizes OpenCV to provide is accomplished;

Video capture module 20 is used to accomplish the video capture of video file or camera, and the video of catching is shown to system interface, provides intuitively and understands.The realization that this module is concrete is to utilize MFC+OpenCV technology, converts the IPLImage data structure of OpenCV among the MFC cvvImage data structure, and is shown in the IDC_STATIC_IMAGE control;

Parameter is provided with module 30, and module is utilized the MFC control, the interface that provides the user that the track algorithm parameter is set, and like population, track button etc.;

Pedestrian's tracking module 40 is used to the method that adopts HOG and color double characteristic to combine, uses the tracking framework of particle filter, accomplishes the tracking to the motion pedestrian.

The present invention's one preferred implementation, exploitation under the VS2008 environment, used PC is configured to the internal memory of intel pentium double-core 2.0CPU and 2G, mainly is to use the support of the built-in function OpenCV that increases income.The data set of selecting for use is PETS2001, and the sampling population is 100, and the primary weight is 0.01, and choosing of threshold value k can draw according to experiment, generally sets the value of k according to population, and k is 0.02 here.

The present invention compares with the track algorithm of traditional solid color characteristic in the prior art, and the tracking effect of double characteristic obviously will be higher than the tracking effect of solid color characteristic.

Joining shown in Figure 6 is the time result that two kinds of track algorithms consume when handling each frame; Owing to introduced the HOG characteristic; Cause at the track algorithm of the calculated amount double characteristic of feature extraction phases bigger than single color characteristic; The difference average out to 10ms of time when as can be seen from the figure two kinds of track algorithms are handled each frame was 25 frame/seconds according to the frame per second of video, and the double characteristic Processing Algorithm that exceeds 10ms still can satisfy the requirement of real-time follow-up.

The weight limit value change profile of particle can reflect the situation of change of particle weight in communication process, from this changes, can find out the number of effective particle, and effectively speak more the more performance of bright track algorithm of particle number is good more, and the precision of tracking is accurate more.Fig. 7 and Fig. 8 are the change profile of the maximum weight of particle, and wherein Fig. 7 is the weight limit variation of single characteristic, and Fig. 8 is that the weight limit of double characteristic changes; Can find out based on many than single characteristic of the particle weighted value of the maximum of HOG and color double characteristic, just effectively the single characteristic of population ratio want many, so the track algorithm of double characteristic has more performance.

In another embodiment, select for use actual image sequence to test the track algorithm that improves particle filter, the main information of image sequence is 320*240, and frame rate was 25 frame/seconds.Algorithm is realized under the Vs2008+Opencv2.1 environment; Come the auto-initiation correlation parameter through manually selecting the pedestrian that will follow the tracks of; Population is 300; And compare with traditional particle filter algorithm, when taking place to disturb and blocking, traditional particle filter algorithm occurs following the tracks of skew easily and follows the tracks of the phenomenon of losing; And the track algorithm that this chapter proposes is because the improvement of particle sampler method obtains careful more uniform particle, and to blocking with complex background good adaptability arranged all, and still tracking target accurately has certain robustness.

Fig. 9 has provided the comparison diagram that the present invention improves track algorithm and traditional particle filter tracking algorithm and target pedestrian real motion track, can compare the precision of two kinds of track algorithms more intuitively.Wherein X representes the position on target's center's directions X, and Y representes the position of target's center on the Y direction; The pursuit path that improves track algorithm more approaches real pursuit path than traditional track algorithm.Because what following the tracks of had appearred in the influence of disturbing and blocking, green traditional track algorithm loses, cause following the tracks of failure, and improvement track algorithm of the present invention, receive after the small influence tracking target pedestrian again, continue to accomplish target pedestrian's tracking.

Figure 10 weighs the performance of track algorithm from the track algorithm processing needed time of each frame.Because the present invention uses based on the QMC method of sampling and obtains more effectively particle, makes that in the particle filter iterative process, the efficient of resampling has improved, and finally embodies the minimizing in processing time, more traditional particle filter tracking algorithm is faster, and efficient is better.

In sum, motion pedestrian video automatic tracking method and system based on particle filter of the present invention accomplish the tracking to the motion pedestrian under based on the framework of particle filter.By extracting HOG and color double characteristic; Improve the robustness of particle filter likelihood model; Eliminate unsettled situation in the tracing process; To solid color feature tracking effect under complex background the requirement of problem such as losing and can't satisfy real-time tracking appears easily following the tracks of particularly; Make up better likelihood model in conjunction with the HOG feature by average weighted convergence strategy; Thereby improve the robustness of track algorithm greatly, accomplish stable tracking;

Simultaneously, through studying the characteristic that the particle filter method of sampling obtains the particle collection, we find the particle that produces based on the MC method because its random character causes occurring easily between particle gap and range upon range of; Influence the precision of track algorithm; What is more, causes wave filter to restrain, and can't accomplish tracking.The present invention is through the QMC method of sampling, produces the particle collection that the uniform particle collection of detailed rules and regulations more replaces the MC stochastic sampling to obtain, and eliminates interparticle gap and range upon range of, finally improves the precision of track algorithm.

The above only is the application's a preferred implementation, makes those skilled in the art can understand or realize the application.Multiple modification to these embodiment will be conspicuous to one skilled in the art, and defined General Principle can realize under the situation of spirit that does not break away from the application or scope in other embodiments among this paper.Therefore, the application will can not be restricted to these embodiment shown in this paper, but will meet and principle disclosed herein and features of novelty the wideest corresponding to scope.

Claims (9)

1. the motion pedestrian video automatic tracking method based on particle filter is characterized in that, said method comprising the steps of:

S1, input one two field picture detect through HOG set of eigenvectors and SVM vector machine, have judged whether the pedestrian, if, execution in step S2, if not, input next frame image detects again;

S2, realize particle filter tracking based on HOG and color double characteristic; At first obtain target pedestrian's initial rectangular zone; And the plurality of particles of from the target rectangle zone, sampling; Extract HOG characteristic and color characteristic, calculate the weight that HOG and color double characteristic merge the back particle, resample after getting state estimation to the end and export estimating target through the least mean-square error estimator;

S3, judge whether image is last frame, if, then finish to follow the tracks of, if not, return step S2.

2. method according to claim 1 is characterized in that, " detecting through the HOG set of eigenvectors " among the said step S1 is specially:

Evaluate image pixel characteristic in gray space or color space is carried out the gamma correction standardization to image;

According to the extraction of HOG gradient, calculate the gradient of each pixel;

The computing unit gradient magnitude is specially the gradient orientation histogram that comprises pixel in each unit that adds up, and each gradient orientation histogram is mapped on definite angle again, obtains the HOG proper vector;

A piece is formed in adjacent unit, carries out piece normalization;

Select unit and piece for use, obtain the HOG set of eigenvectors and detect.

3. method according to claim 1 is characterized in that, the SVM vector machine among the said step S1 comprises the svm classifier device, and said svm classifier device comprises traditional svm classifier device and linear inseparable svm classifier device.

4. method according to claim 3 is characterized in that, the classification function in said traditional svm classifier device is:

Classification function in the linear inseparable svm classifier device is: $f\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{\mathrm{\&alpha;}}_{i}K(x\&CenterDot;x_i)+b^{*}.$

5. method according to claim 1 is characterized in that, said step S2 is specially:

Obtain target pedestrian's initial rectangular zone, and each particle weight of the plurality of particles of from the target rectangle zone, sampling

initialization is 1/N;

Through state equation x _t=Ax _T-1+ Bw _T-1Prediction obtains the t state of particle constantly

Calculate the weight that HOG and color double characteristic merge the back particle:

$w^i=\frac{\mathrm{colcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_c^i+\frac{\mathrm{hogcount}}{\mathrm{colcount}+\mathrm{hogcount}}*w_h^i;$

Get state estimation to the end through the least mean-square error estimator:

The output estimating target also resamples.

6. method according to claim 5 is characterized in that, the extraction of color characteristic is specially among the said step S2:

Calculate color histogram at rgb space, the probability distribution of target area is:

U=1 ... M, wherein, the state variable of said target candidate target area is X, and the center is x, and δ is a Kronecker delta function, and n is the number of said target area pixel, function b:R ²→ { 1...m} distributes the bin of a correspondence for each pixel, and C is a constant, and the normalization formula does. $C=1/{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|x_i\left|\right|}^2\right),$ K representes kernel function, and $k\left(x\right)=\left\{\begin{array}{c}1-x^2:x<1\\ 0:\mathrm{Otherwise}\end{array}\right.;$

To Template p is according to candidate template

Calculate, calculate similarity apart from doing Then according to the observed reading of similarity distance calculation particle $p\left(z\right|x_t)=\frac{1}{\sqrt{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}}\mathrm{Exp}\left(\frac{-d^2}{2{\mathrm{\&sigma;}}^2}\right).$

7. method according to claim 6 is characterized in that, the HOG Feature Extraction is specially among the said step S2:

Use single order center operator [1,0,1], the Grad G of each pixel level of computed image and vertical direction _hAnd G _v, wherein:

$G_h(x,y)=f(x+1,y)-f(x-1,y),\&ForAll;x,y,$ $G_v(x,y)=f(x,y+1)-f(x,y-1),\&ForAll;x,y;$

Calculate each pixel intensity M (x, y) with gradient direction θ (x, y), wherein:

$M(x,y)=\sqrt{G_h{(x,y)}^2+G_v{(x,y)}^2},$ θ(x，y)＝tan ^-1(G _h(x,y)/G _v(x，y))；

Gradient intensity is carried out normalization, obtain the histogram of gradients of entire image.

8. method according to claim 7 is characterized in that, the sampling among the said step S2 is specially:

At the hypercube D of unit _s: [0,1) the last QMC point set { u that produces ⁱ, i=1,2 ... N} utilizes formula x ⁱ=[a+ (b-a) ο u ⁱ] convert the QMC point set into particle collection { x ⁱ, 1,2 ... N}, wherein a and b are x ⁱThe interval in space, place.

9. motion pedestrian video automatic tracking system based on particle filter as claimed in claim 1 is characterized in that said system comprises:

Pre-processing module, said module are utilized HOG and SVM detection algorithm, detect the pedestrian who comprises in the initial frame, and provide pedestrian's rectangular area;

Video capture module is used to accomplish the video capture of video file or camera, and the video of catching is shown to system interface, provides intuitively and understands;

Parameter is provided with module, and said module is utilized the MFC control, provides the user that the interface of track algorithm parameter is set;

Pedestrian's tracking module is used to the method that adopts HOG and color double characteristic to combine, uses the tracking framework of particle filter, accomplishes the tracking to the motion pedestrian.

Images