CN102044073A - Method and system for judging crowd density in image - Google Patents

Abstract

The invention provides a method and system for judging the crowd density in an image. The method comprises the following steps: 1. selecting a target region from a video image sample acquired in an image acquisition device by utilizing a block analysis unit, and carrying out block analysis of image blocks in the target region; 2. determining the composite form of two-classifiers by a coding unit; 3. selecting a confidence training sample by a training unit, and respectively training each two-classifier; and 4. obtaining the crowd density grade category of the maximum posteriori probability through a decoding unit by means of a channel transmission model. The method can be suitable for obtaining the credible crowd density grade in different scenes and can provide the basis for crowd monitoring and safety guarantee of important regions.

Description

Judge the method and system of crowd density in the image

Technical field

The invention belongs to image processing field, relate to crowd density determination methods in a kind of image.

Background technology

Traditional crowd's monitoring is to realize that by a certain scene of supervision it needs user self that scene image is made judgement by closed-circuit television.This method subjectivity is very strong, can not carry out quantitative test.The development of modern digital image processing is for the approach that provides is provided.Technology such as Flame Image Process, pattern-recognition, computer vision are applied in crowd's monitoring, can reach automatic, objective, real-time analysis the crowd.

Crowd density is judged based on computer vision and mode identification technology in the image, by monitoring image or video are carried out analytical calculation, draws the quantization level of crowd density in the image.Crowd density information is the strong foundation of crowd's monitoring management on a large scale, the crowd density information that it can provide inside, market to distribute by the different periods to the market, assist management laminated reason distribution services and management resource; It also can be widely used in crowd's monitoring of the Vomitory and the important area of communal facilitys such as bus stop, rail vehicles for transporting passengers station, airport, to obtain the accurate data of passenger number and distribution in real time, prevention is owing to the crowded potential safety hazard that causes of client, for science scheduling, safety guarantee provide foundation.

The crowd density determination methods mainly can be divided into two big classes in the image at present commonly used: based on the method for pedestrian detection with based on the method for statistical learning.

Utilize existing pedestrian detection and people's face to detect operator based on the method for pedestrian detection, from image, detect single pedestrian, in conjunction with motion detection and tracking technique, the variation of analyst's population density.This method depends on the effect of pedestrian detection operator,, serious shielding big or lower, the single pedestrian of picture quality at crowd density can not the situation of reliable detection under performance can't guarantee.

By correlated characteristic that extracts crowd in the picture frame and the study that supervision is arranged, obtain the relation between characteristic quantity and the crowd density based on the method for statistical learning.Application the earliest be Davies etc. by extracting foreground area area and edge length and setting up linear relationship with crowd density, crowd's image in Liverpool Street railway station is analyzed.This method is calculated simple, is easy to realize, but can't solves occlusion issue, along with crowd density increases, will increase owing to block the evaluated error that causes.Periodical ISIBGRAPHI ' the 98:Proceedings of the International Symposium onComputer Graphics that Marana etc. published in 1998, Image Processing, article On the efficacy oftexture analysis for crowd monitoring among the and Vision points out, have corresponding relation between textural characteristics and the crowd density: highdensity crowd shows as thin pattern on texture, and low-density crowd shows as the roughcast formula on texture.Method based on statistical learning does not rely on pedestrian detection, directly sets up the corresponding relation of crowd density and feature by statistical learning, by selecting suitable feature and sorter model, can obtain effect preferably.

When crowd density is quantified as density rating, the problem that crowd density is judged just is converted into the problem of multicategory classification: its input is the characteristic quantity relevant with the crowd, and output is the density rating of limited number.The common method that solves this multicategory classification problem has neighbour's analytic approach, polynomial fitting method, neural network model etc.Wherein, neighbour's analytic approach is the training sample cluster in feature space with same density rating, form M representative point (representative point of each density rating needn't be identical), for the sample of a certain unknown classification, utilize the classification of its arest neighbors representative point or K neighbour's representative point to be classification according to definite this sample.This method is simple, but depends critically upon choosing of cluster number, needs rule of thumb or experimental verification selection parameter.Polynomial fitting method and neural network model are the methods that minimizes the experience error, but are absorbed in local minimum easily or produced the study phenomenon.

Statistical Learning Theory is present best theoretical at small sample statistical estimate and prediction study, and it has systematically been studied under the condition that the empiric risk minimization principle is set up, the limited sample relation of empiric risk and expected risk theoretically and has reached and how to utilize these theories to find problems such as new learning principle and method.Support vector machine (SVM) has been widely used in a lot of fields such as pattern-recognition and data mining as a kind of implementation method of Statistical Learning Theory.Illustrate among the A Tutorial on Support Vector Machines forPattern Recognition that Christopher J.C.Burges delivers in the Data Mining andKnowledge Discovery that published in 1998, the essence of svm classifier device is linear two sorters, it is based on the structural risk minimization theory, construction optimum segmentation lineoid in feature space, make learner obtain global optimization, and the expected risk value in whole sample space satisfy certain upper bound with certain probability.

When utilizing SVM to handle the multiclass problem, just need the suitable multicategory classification device of structure.At present, the method of structure SVM multicategory classification device mainly contains two classes: a class is a direct method, directly on objective function, make amendment, the parametric solution of a plurality of classifying faces is merged in the optimization problem, realize multicategory classification by finding the solution this optimization problem " disposable ".This method seems simply, but its computation complexity implements the comparison difficulty than higher, only is suitable in the small scale problem.Another kind of is indirect method, mainly is to realize the structure of multi-categorizer by making up a plurality of two sorters, method commonly used comprises one to one, a pair of surplus, decision tree and error correction output code method (error correcting output codes, ECOC) etc.

Summary of the invention

Technical matters to be solved by this invention provides a kind of method of judging crowd density in the image, this method is at first determined the array configuration of two sorters to the zone in the video image sample, analysis is selected and is put letter training sample and respectively each two sorter being trained, and obtains maximizing the density rating of posterior probability by the Channel Transmission model.This method obtains corresponding density rating at the crowd characteristic of input, goes for different scenes and obtains believable crowd density grade judgement.

The invention provides crowd density determination methods in a kind of image, this method comprises: step 1, by select target zone in the video image sample of gathering, stroke block analysis unit, and in described target area, carry out the block analysis of drawing of image block by image collecting device; Step 2 is determined the array configuration of two sorters by coding unit; Step 3 is selected by training unit and to be put the letter training sample and each two sorter is trained respectively; Step 4 is obtained maximizing the crowd density graded category of posterior probability by the Channel Transmission model by decoding unit.

Further, the method for described judgement crowd density also comprises: carry out the block analysis of drawing of image block in the step 1 in described target area according to perspective model.

Further, the method for described judgement crowd density also comprises: have the roughly the same number that can hold at most in the different image block according to described perspective model division.

Further, the method for described judgement crowd density also comprises: described perspective model is used to be implemented in the unification of the image block under different scenes, different cameras angle, the diverse location and determining of clear and definite density rating standard.

Further, the method for described judgement crowd density also comprises: after the step 1, by computing unit described each image block is calculated multiple dimensioned local binaryzation operator, and add up the texture description symbol of the normalized frequency histogram of its correspondence as the crowd.

Further, the method for described judgement crowd density also comprises: ask for the multi-dimension texture feature at described each image block.

Further, the method for described judgement crowd density also comprises: on average each image block is calculated multiple dimensioned local binaryzation operator based on piece.

Further, the method for described judgement crowd density also comprises: extract the texture description symbol by extraction unit again, for each image block generating feature space, obtain training sample and test sample book.

Further, the method for described judgement crowd density also comprises: by coding unit error correction output code matrix is set in the step 2, to determine the array configuration of two sorters.

Further, the method for described judgement crowd density also comprises: described matrix design step is as follows:

(1) for classification set Q, calculates per two classification subclass S _iAnd S _jBetween expense snow be the fisher distance:

$f(S_i,S_j)=\frac{{(m_i-m_j)}^T(m_i-m_j)}{{\mathrm{\Σ}}_i+{\mathrm{\Σ}}_j}$

Wherein, S _i, S _jBe two target classifications, m _i, m _jBe average in the target classification, ∑ _i, ∑ _jBe target classification internal variance.

(2) establish one two classification, the fisher distance after make dividing between two subclass be a maximum.

(3) subclass repeating step (1), (2 of dividing for step (2)) are till each subclass only comprises a classification.

Further, the method for described judgement crowd density also comprises: in the described step 3 by training unit in each two sorter, utilize the posterior probability analysis from training sample, to select and put the letter training sample, and train respectively.

Further, the method for described judgement crowd density also comprises: adopt the k nearest neighbor method to add up described posterior probability P (w ⁺| x), wherein x is a training sample.

Further, the method for described judgement crowd density also comprises: put the letter sample and select: if | P (w ⁺| x)-and 05|＜ε, then x is the non-letter sample of putting, weight is 0; Otherwise for putting the letter sample, weight is 1.

Further, the method for described judgement crowd density also comprises: the input end of the model of Channel Transmission described in the step 4 represents that the true classification of sample, output terminal represent that sample is through this sorter sorted output classification.

The present invention also provides a kind of system of judging crowd density in the image, and this system comprises: draw the block analysis unit, select a zone in the video image sample of being gathered by image collecting device, and carry out the block analysis of drawing of image block according to perspective model in the zone; Computing unit calculates based on the average multiple dimensioned local binaryzation operator of piece each image block, and adds up the texture description symbol of the normalized frequency histogram of its correspondence as the crowd, and asks for the multi-dimension texture feature at described each image block; Extraction unit extracts the texture description symbol, for each image block generating feature space, obtains training sample and test sample book; Coding unit is provided with error correction output code matrix, to determine the array configuration of two sorters; Training unit in each two sorter, utilizes the posterior probability analysis to select from training sample and puts the letter training sample, and train respectively; Decoding unit obtains maximizing the crowd's of posterior probability density rating classification from test sample book by the Channel Transmission model.

Beneficial effect of the present invention have following some:

1, set up in different scenes, different cameras angle, the unification of diverse location hypograph piece density rating, the clear and definite criteria for classifying, made this method can adapt to different scenes, provided believable density rating and estimate.

2, this method adopts and selects based on the opposed letter sample of the principle of posterior probability estimation, whole categorizing system can be brought under the classical framework of Bayesian inference, can be effectively the performance difference of different sorters be embodied in the reasoning process by the structure mode, thereby under the prerequisite of fast prediction, have improved predictablity rate.

3, this method adopts based on the decision-making of decoding of Channel Transmission model, can effectively simulate the information delivery format of sorter, thereby obtain the transition probability of single two sorters from the input state to the output state.Simultaneously, this Channel Transmission model can be connected, thereby handles the situation of a plurality of two sorter cascades.

4, this method adopts the design of error correction output encoder matrix, effectively selects and makes up two sorters, reduces redundant sorter, and can avoid occurring inseparable situation.

This method goes for different scenes and obtains believable crowd density grade, can provide foundation for the crowd's monitoring and the safety guarantee of important area.

Description of drawings

Fig. 1 judges the process flow diagram of the method for crowd density in the image for the present invention;

Fig. 2 judges the image block sample demarcation synoptic diagram of the method for crowd density for the present invention;

Fig. 3 judges two sorter Channel Transmission models of the method for crowd density for the present invention.

Embodiment

For purpose of the present invention, feature and advantage can be become apparent more, the present invention is further detailed explanation below in conjunction with the drawings and specific embodiments.

Basic scheme of the present invention is: based on perspective model video image sample and target video image zone are carried out the image block division, determine the array configuration of two sorters, analysis is selected and is put letter training sample and respectively each two sorter being trained, and obtains maximizing the density rating of posterior probability by the Channel Transmission model.

The invention provides an embodiment.With reference to figure 1, show the process flow diagram of crowd density estimation method in video image of one embodiment of the invention, specifically can may further comprise the steps.

Step 1:, and in the zone, carry out the block analysis of drawing of image block according to perspective model by select target zone in the video image sample of gathering, stroke block analysis unit by image collecting device.

The crowd density of video image sample is judged it is the relation of setting up between local grain feature and the density rating.Present embodiment is realized determining for the unification of the image block under different scenes, different cameras angle, diverse location and clear and definite density rating standard based on perspective model.So-called perspective model is when utilizing two dimensional image to represent three-dimensional spatial information, can produce the visual angle distortion, and promptly distant objects ratio object nearby seems smaller.This three-dimensional can be similar to a linear transformation usually to the perspective model of two dimension:

$\left(\begin{array}{c}u\\ \mathrm{<figure-callout\; id="1"\; label="v"\; filenames="H2009100934405E0000022.png"\; state="\{\{state\}\}">v</figure-callout>}\\ 1\end{array}\right)=H\&times;\left(\begin{array}{c}\mathrm{<figure-callout\; id="1"\; label="x\; y"\; filenames="H2009100934405E0000022.png"\; state="\{\{state\}\}">x</figure-callout>}& \\ y\\ 1\end{array}\right)$

Wherein, H is a perspective transformation matrix.H accurately finds the solution the information that need use camera calibration, in practical operation, can then intermediate mass be carried out interpolation and obtain the piece size by specifying the most nearby the piece size with the farthest.This step can be regarded as the rectification to the visual angle distortion, guarantees that the area of the actual area that Same Scene hypograph piece is represented is basic identical.

In the present embodiment, the principle of dividing the image block of different densities grade based on perspective model is exactly by the size of image block is set, make different image blocks have roughly the same " capacity ", promptly different image blocks has the roughly the same number that can hold at most.During calculating, at first determine the size of smallest blocks and largest block image block; Be similar to the size that the perspective scale model is tried to achieve intermediate mass by linear difference again.

The criteria for classifying of present embodiment density rating is as shown in table 1, mainly considers two factors: the one, and number in the image block, the 2nd, the ratio of crowd's area occupied in the image block.There is association between the two, and all directly affects the subjective perception of people for density.The density rating of dividing is the 1-5 Pyatyi, and wherein the crowd density the difference representative image is very low, low, medium, high, very high from 1 grade to 5 grades.Fig. 2 is the synoptic diagram that present embodiment image block sample is demarcated.

Table 1

Density rating	1	2	?3	?4	5
						Number	0-0.5	0.5-2	?2-4	?4-6	＞6
Crowd's area occupied	＜10％	10％-30％	?30％-60％	?60％-90％	＞90％

Step 2: by computing unit each image block is calculated based on the average multiple dimensioned local binaryzation operator of piece, and add up the texture description symbol of the normalized frequency histogram of its correspondence as the crowd; Ask for the multi-dimension texture feature at described each image block.

This texture description symbol more is applicable to the description of crowd's texture than gray level co-occurrence matrixes, edge orientation histogram and general LBP histogram commonly used.By calculating above-mentioned feature, each image block generates the proper vector of 48 dimensions.Multicategory classification just carries out in this feature space.

Step 3: extract the texture description symbol by extraction unit,, obtain training sample and test sample book for each image block generating feature space.

Step 4: by coding unit error correction output code matrix is set, to determine the array configuration of two sorters.

The design that error correction output code matrix is set is the ECOC matrix is the suitable classification coding of design, makes coding of all categories that the maximum property of distinguishing be arranged, and the tool separability (helping improving the performance of two sorters) of each corresponding two sorter.In general, long coding can improve the ability of error correction, but can increase the complexity of calculating and training, classification time, and if introduce the sorter of classification performance difference, can reduce classification accuracy rate on the contrary.Can be any 1,0 ,-1 combination of satisfying following condition in the ECOC matrix:

1) matrix row, column full rank;

2) row are not 1,0 or-1 entirely;

3) row is not 0 entirely.

ECOC coding commonly used has one to one, a pair of surplus and generated code etc. at random.Generated code promptly at random: set the sorter number, generate the classification chain code at random, reject the chain code that does not meet above-mentioned condition, generate the ECOC matrix.

Binary tree thought is incorporated in the ECOC matrix design, has avoided redundant two sorters and irrational two sorters to occur.

Concrete steps are as follows:

(1), for classification set Q, calculate per two classification subclass S _iAnd S _jBetween expense snow be the fisher distance:

$f(S_i,S_j)=\frac{{(m_i-m_j)}^T(m_i-m_j)}{{\mathrm{\&Sigma;}}_i+{\mathrm{\&Sigma;}}_j}---\left(1\right)$

Wherein, S _i, S _jBe two target classifications, m _i, m _jBe average in the target classification, ∑ _i, ∑ _jBe target classification internal variance.

(2), establish one two classification, make divide after fisher distance between two subclass be maximum.

(3), subclass repeating step (1), (2) divided for step (2), till each subclass only comprises a classification.

Step 5: in each two sorter, utilize the posterior probability analysis from training sample, to select and put the letter training sample, and train respectively by training unit.

Crowd density grade separation problem has the not available characteristics of general many classification problems: 1) adjacent similar, promptly adjacent density rating is the most similar; 2) there is fuzzy sample in gradual change boundary between the adjacent classification, form a gradual change watershed area.

" litura " between the adjacent classification makes that the classification boundary is indeterminate.The training process of SVM is subjected to the litura influence easily fully by data-driven, causes classification accuracy to descend and the structure risk increase.The present invention proposes the SVM (post probability SVM) based on posterior probability, estimates that by nearest neighbour method or parzen window training sample x belongs to the conditional probability P (w of positive classification ⁺| x).For P (w ⁺| near the x) point 0.5, think that its degree of confidence is lower, give less weight; And for P (w ⁺| near the x) point 0 or 1, think that its degree of confidence is higher, give bigger weight, thereby reduced the influence of noise and exceptional value point effectively for classifying face.

In order to simplify application, propose following based on the SVM training method of putting the letter sample:

Posterior probability P (w ⁺| statistics x): adopt the k nearest neighbor method, get near K sample of x, statistics wherein belongs to the K that counts of positive classification ⁺, then:

$P\left(w^{+}\right|x)=\frac{K^{+}}{K}---\left(2\right)$

Putting the letter sample selects: if | P (w ⁺| x)-and 0.5|＜ε, then x is the non-letter sample of putting, weight is 0; Otherwise for putting the letter sample, weight is 1.

Utilize all to put the letter sample point, according to traditional SVM coaching method training svm classifier device.

Wherein, threshold epsilon and neighbour count the selection of K, can obtain by cross validation, in this article, are taken as 0.1 and 5 respectively.Experiment showed, that owing to removed the ambiguous training sample this employing put SVM that the letter sample training obtains than traditional svm classifier accuracy height, and the support vector number reduces significantly, extensive performance strengthens.

Step 6: the density rating classification that from test sample book, obtains maximizing the crowd of posterior probability by decoding unit by the Channel Transmission model.

The output vector of test sample book is decoded, ask the capable coding vector the most similar according to predetermined similarity measurement, and sample is assigned to this row corresponding class to it.

From the angle analysis of Channel Transmission, decoding should be the process of a reduction true value from the observation signal that has error.

Fig. 3 has represented the mode of two sorters, and its input end is the true classification of sample, and on behalf of this sample, 1 should belong to positive classification in this sorter, and-1 this sample of expression belongs to negative classification in this sorter, and 0 this sample of expression does not relate in this sorter.Its output terminal is represented this sample through the sorted output classification of this sorter, and 1 this sample of expression is divided into positive sample, and-1 this sample of expression is divided into negative sample, and the classification that does not relate to also will be divided into respectively with certain probability in the middle of this two class.Transmission parameter in the model is suc as formula shown in (3)～(5):

$a=P(y_O=1|y_I=1)=\frac{{\mathrm{\&Sigma;}}_{i=1}^{n}I(y_{\mathrm{iI}}=1,y_{\mathrm{iO}}=1)}{{\mathrm{\&Sigma;}}_{i=1}^{n}I(y_{\mathrm{iI}}=1)}---\left(3\right)$

$b=P(y_O=0|y_I=1)=\frac{{\mathrm{\&Sigma;}}_{i=1}^{n}I(y_{\mathrm{iI}}=0,y_{\mathrm{iO}}=1)}{{\mathrm{\&Sigma;}}_{i=1}^{n}I(y_{\mathrm{iI}}=0)}---\left(4\right)$

$c=P(y_O=-1|y_I=-1)=\frac{{\mathrm{\&Sigma;}}_{i=1}^{n}I(y_{\mathrm{iI}}=-1,y_{\mathrm{iO}}=-1)}{{\mathrm{\&Sigma;}}_{i=1}^{n}I(y_{\mathrm{iI}}=-1)}---\left(5\right)$

Wherein, y _I, y _ORepresent the sample class of input and output respectively, n represents the number of training sample.I () represents indicative function.Suppose that test sample book x is through M the two sorters output vector S (s that obtains classifying _j∈ [1 ,-1], j=1 ... M), then sample x belongs to classification w _iPosterior probability be:

$P\left(w_i\right|x)=P\left(M_i\right|S)=\frac{P\left(S\right|M_i)P\left(M_i\right)}{P\left(S\right)}---\left(6\right)$

Wherein, P (M _i) represented the prior probability of classification i, under the situation of no priori, can be made as 1/N, N is the classification number, P (S) is the probability that chain code S occurs, P (S|M _i) be illustrated in the probability that occurs chain code S under the condition that sample belongs to the i class.Suppose between each two sorter separately, then have:

$P\left(S\right|M_i)=\underset{j=1}{\overset{M}{\mathrm{\&Pi;}}}P\left(s_j\right|m_{\mathrm{ij}})---\left(7\right)$

Wherein, P (s _j| m _Ij) being the transmission parameter of single two sorters, its value is one of six parameters in the mode { a, 1-a, b, 1-b, c, 1-c}.Finally, select the affiliated classification of the classification of posterior probability maximum as sample x:

$w=\arg \underset{j}{\max }P\left(w_j\right|x)---\left(8\right)$

The present invention also provides another embodiment, and a kind of method of judging crowd density in the image comprises the steps:

Step 1: in video image sample, select an area-of-interest, and carry out the block analysis of drawing of image block according to the perspective model in the zone;

Step 2:(1), each image block is calculated based on the average multiple dimensioned local binaryzation operator of piece, and add up the texture description symbol of the normalized frequency histogram of its correspondence as the crowd; (2), ask for the multi-dimension texture feature at described each image block;

Step 3: extract the texture description symbol,, obtain training sample and test sample book for each image block generating feature space;

Step 4: design error correction output code matrix, to determine the array configuration of two sorters;

Step 5: in each two sorter, utilize the posterior probability analysis from training sample, to select and put the letter training sample, and train respectively;

Step 6: the density rating classification that from test sample book, obtains maximizing the crowd of posterior probability by the Channel Transmission model.

The error correction output code matrix of described step 4, form by following steps:

(1) for classification set Q, calculates per two classification subclass S _iAnd S _jBetween fisher apart from f (S _i, S _j) (as the formula (1)).

(2) ask one two classification, the fisher distance after make dividing between two subclass be a maximum

(3) subclass repeating step (1), (2 of dividing for step (2)) are till each subclass only comprises a classification.

Posterior probability P (the w of described step 5 ⁺| estimation x): adopt the k nearest neighbor method, get near K sample of x, wherein belong to counting of positive classification and be K ⁺, then: P (w ⁺| x) draw by formula (2).

Wherein, the neighbour counts K and obtains by cross validation.

The mid-letter training sample of described step 5 is selected: if | P (w ⁺| x)-and 0.5|＜ε, then x is the non-letter sample of putting, weight is 0; Otherwise for putting the letter sample, weight is 1; Wherein, threshold epsilon obtains by cross validation.

The input end of Channel Transmission model is represented the true classification of sample in the described step 6, and output terminal is represented sample through the sorted output classification of this sorter, and transmission parameter a, b in the model and c are suc as formula shown in (3), (4), (5).

The test sample book by the Channel Transmission model in the described step 6 obtains the output vector of classifying through a plurality of two separate sorters, and then this test sample book belongs to the posterior probability of described classification, selects the affiliated classification of the classification of posterior probability maximum as sample for use.

Although illustrated and described embodiments of the invention, but it will be appreciated by those skilled in the art that, on the basis of not departing from spirit of the present invention and principle, can make a change this embodiment, scope of the present invention is limited by claims and their equivalents.

Claims (15)

1. method of judging crowd density in the image, this method comprises:

Step 1 by select target zone in the video image sample of being gathered by image collecting device, stroke block analysis unit, and is carried out the block analysis of drawing of image block in described target area;

Step 2 is determined the array configuration of two sorters by coding unit;

Step 3 is selected by training unit and to be put the letter training sample and each two sorter is trained respectively;

Step 4 is obtained maximizing the crowd density graded category of posterior probability by the Channel Transmission model by decoding unit.

2. the method for claim 1 is characterized in that, carries out the block analysis of drawing of image block in the step 1 in described target area according to perspective model.

3. method as claimed in claim 2 is characterized in that, has the roughly the same number that can hold at most in the different image block according to described perspective model division.

4. method as claimed in claim 2 is characterized in that, described perspective model is used to be implemented in the unification of the image block under different scenes, different cameras angle, the diverse location and determining of clear and definite density rating standard.

5. the method for claim 1 is characterized in that, after the step 1, by computing unit described each image block is calculated multiple dimensioned local binaryzation operator, and adds up the texture description symbol of the normalized frequency histogram of its correspondence as the crowd.

6. method as claimed in claim 5 is characterized in that, asks for the multi-dimension texture feature at described each image block.

7. method as claimed in claim 5 is characterized in that, on average each image block is calculated multiple dimensioned local binaryzation operator based on piece.

8. method as claimed in claim 5 is characterized in that, also comprises, extracts the texture description symbol by extraction unit, for each image block generating feature space, obtains training sample and test sample book.

9. the method for claim 1 is characterized in that, by coding unit error correction output code matrix is set in the step 2, to determine the array configuration of two sorters.

10. method as claimed in claim 9 is characterized in that, the described step that error correction output code matrix is set comprises:

(1) for classification set Q, calculates per two classification subclass S _iAnd S _jBetween take snow distance:

$f(S_i,S_j)=\frac{{(m_i-m_j)}^T(m_i-m_j)}{{\mathrm{\&Sigma;}}_i+{\mathrm{\&Sigma;}}_j}$

Wherein, S _i, S _jBe two target classifications, m _i, m _jBe average in the target classification, ∑ _i, ∑ _jBe target classification internal variance.

(2) establish one two classification, take the snow distance for maximum after make dividing between two subclass.

(3) subclass repeating step (1), (2 of dividing for step (2)) are till each subclass only comprises a classification.

11. the method for claim 1 is characterized in that, in the described step 3 by training unit in each two sorter, utilize the posterior probability analysis from training sample, to select and put the letter training sample, and train respectively.

12. method as claimed in claim 11 is characterized in that, adopts the k nearest neighbor method to add up described posterior probability P (w ⁺| x), wherein x is a training sample.

13. method as claimed in claim 11 is characterized in that, puts the letter sample and selects: if | P (w ⁺| x)-and 0.5|＜ε, then x is the non-letter sample of putting, weight is 0; Otherwise for putting the letter sample, weight is 1.

14. the method for claim 1 is characterized in that, the input end of the model of Channel Transmission described in the step 4 represents that the true classification of sample, output terminal represent that sample is through this sorter sorted output classification.

15. a system of judging crowd density in the image comprises:

Draw the block analysis unit, be used for, and in described zone, carry out the block analysis of drawing of image block according to perspective model in video image sample select target zone by the image collecting device collection;

Computing unit is used for each image block is calculated based on the average multiple dimensioned local binaryzation operator of piece, and adds up the texture description symbol of the normalized frequency histogram of its correspondence as the crowd, and asks for the multi-dimension texture feature at described each image block;

Extraction unit is used to extract the texture description symbol, for each image block generating feature space, obtains training sample and test sample book;

Coding unit is used to be provided with error correction output code matrix, to determine the array configuration of two sorters;

Training unit is used for utilizing the posterior probability analysis to select from training sample and putting the letter training sample, and respectively it is trained at each two sorter;

Decoding unit is used for obtaining maximizing from test sample book by the Channel Transmission model crowd's of posterior probability density rating classification.

Images