CN102136060A - Method for detecting population density - Google Patents

Abstract

The invention discloses a method for detecting population density. The method is implemented under the support of a digital signal processing chip and a digital camera serving as a sensor, and is characterized by comprising the following steps of: (1) continuously acquiring a fixed point area by using the digital camera serving as the sensor to form a digital video; (2) extracting the angular point characteristic of the digital video obtained in the step (1) on the basis of a gradient method; (3) extracting the textural characteristic of the digital video obtained in the step (1) on the basis of scaling invariant local triple property (SILIP); and (4) estimating the population density of the fixed point area by using the angular point characteristic obtained in the step (2) and the regression relation between the characteristic combination of the textural characteristic obtained in the step (3) and the population density. By adopting the method for detecting the population density, the problem of density estimating error caused by an angle difference, an over dense population, illumination variation and the like existing in the conventional density detection method based on video analysis is solved. The method has the advantages of high accuracy, strong robustness and the like.

Description

A kind of population density detection method

Technical field

The invention belongs to the population density detection method, especially based on the population density detection method of video analysis.

Background technology

Colony (vehicle/crowd) density early warning comprises vehicle block up early warning and crowd massing early warning etc., and there is very important meaning in unimpeded and public safety strick precaution field in public transport.Rely on the artificial estimation approach can't the scale support application, use video analysis and realize that the on-line intelligence of population density detects and become current research focus.Based on the method for video intelligent analyzing and testing population density, mainly comprise the Target Signature Analysis method and the population density estimation technique at present.

The Target Signature Analysis method is added up on the accurate detection basis based on individuality.At first detect, follow the tracks of the target signature in the video sequence image, identify the target of existence, add up then.These class methods are better to the more sparse scene detection effect of target, but for scene intensive as wagon flows such as intersection or subway stations or that the crowd is dense, go out target because problem such as block is difficult to Accurate Analysis, so can't add up target.

The population density estimation technique is not discerned statistics to individuality, but integral body is analyzed, and has the dense degree of colony in mind, and the individual concrete quantity of out of true statistic mass.At present the population density method of estimation based on video analysis has: based on " foreground pixel sum/density " regression relation, " gray level co-occurrence matrixes/density " regression relation and add the technology that some supplementary meanss are carried out extensive diagnostic, comprise wavelet transformation technique etc.Analyze based on pixel, because problems such as angles, there are not linear relationship in number of pixels and target, and close-range target pixels statistics number is more, long-range objectives pixels statistics number is less, so analysis result has often caused than mistake owing to angle difference or population density are different.And the illumination instability is also bigger to the density analysis influence based on pixel scale.

In sum, the existing problem that exists based on the Density Detection method of video analysis is that aspect abilities such as opposing angle difference, colony's dense degree and illumination variation are relatively poor.

Summary of the invention

The invention provides a kind of solution of the above problems, a kind of population density detection method based on video analysis is provided.

Technical scheme of the present invention provides a kind of population density detection method, and this method is to realize under the support of digital camera as sensor and digital signal processing chip, and it is characterized in that: this method may further comprise the steps:

1) adopt digital camera as sensor, continuous acquisition fixed point zone forms digital video;

2) the angle point feature of digital video image frame extraction 1);

3) textural characteristics of digital video image frame extraction 1);

4) characteristics combination of resulting described textural characteristics and the estimation of the regression relation between the population density 1 resulting described angle point feature and 3 application 2))) the regional population density of described fixed point.

Preferably, extract the angle point feature of described digital video image frame step 2 wherein) based on gradient method.

Preferably, wherein extract the textural characteristics of described digital video image frame in the step 3) based on the constant local ternary feature SILTP of convergent-divergent.。

Preferably, described regression relation obtains by support vector regression algorithm SVR study great amount of samples.

Preferably, when estimating the population density in described fixed point zone in the step 4), the distance of the distance of the described digital video of foundation adopts different regression relation response functions.

Population density detection method of the present invention has solved effectively that existing Density Detection method based on video analysis exists because angle difference, colony is overstocked and the problem of the density Estimation error that illumination variation etc. is brought.Have advantages such as precision height, strong robustness.

Description of drawings

Fig. 1 is an algorithm principle block diagram of the present invention.

Embodiment

Below the specific embodiment of the present invention is described in further detail.

As shown in Figure 1, population density detection method of the present invention is at first carried out video acquisition, and the video information that collects through angle point feature extraction, edge texture feature extraction, the estimation of support vector regression SVR algorithm, finally draws testing result successively.

Scheme of the present invention is to realize under the support of digital camera as sensor, estimates according to the regression relation between coarse texture degree and colony (vehicle/crowd) density.Below in conjunction with the algorithm principle figure shown in the accompanying drawing 1, the embodiment of process in detail.

1. adopt digital camera to form digital video frequency flow as signal source as each kind of groups focusing field scape of sensor continuous acquisition, digital camera adopts A/D chips such as CCD or CMOS;

2. manually choose in 1. described each the kind of groups focusing field scape digital video of step and have the population density value parameter of the video frame image of typical meaning as training sample and this sample of respective markers;

3. at first be based on the angle point feature that gradient method extracts sample.

Angle point is the point that two dimensional image brightness changes curvature maximum value on violent point or the image border curve.Angle point is when keeping the image graphics key character, can reduce the data volume of information effectively, make the content of its information very high, improved the speed of calculating effectively, played important effect at computer vision fields such as 3 D scene rebuilding, estimation, target following, Target Recognition, image registration and couplings.Method based on gradient is to judge the existence of angle point by the curvature of edge calculation, and the size of angle point evaluation is not only relevant with edge strength, and relevant with the rate of change of edge direction.Step is as follows:

(1) computed image pixel gradient in the horizontal and vertical directions, and both products obtains the value of 4 elements among the M;

$M=\left[\begin{array}{cc}{I_x}^2& I_x{I}_y\\ I_x{I}_y& {I_y}^2\end{array}\right].$

Wherein, I _x ²=I _x* I _xI ² _y=I _y* I _y

(2) image is carried out gaussian filtering, obtain new M;

Discrete two-dimensional zero-mean Gaussian function is

$\mathrm{Gauss}=\exp (-\frac{(x^2+y^2)}{2{\mathrm{\σ}}^2})$

(3) interest value of each corresponding pixel, i.e. R value on the calculating original image;

R＝{I _x ²×I _y ²-(I _xI _y) ²}-k{I _x ²+I _x ²} ²

(4) choose Local Extremum.Unique point is the pixel of the very big interest value correspondence in the subrange;

(5) setting threshold is chosen a certain amount of angle point;

4. then be based on the textural characteristics that the constant local ternary feature SILTP of convergent-divergent extracts sample;

Given any location of pixels (xc, yc), SILTP coding is as follows:

${\mathrm{SILTP}}_{N,P}^{\mathrm{\τ}}(x_c,y_c)=\underset{k=0}{\overset{N-1}{\⊕}}{s}_{\mathrm{\τ}}(I_c,I_k),---\left(1\right)$

Ic is the center pixel gray-scale value, and Ik is to be that N of equal intervals on the circle of radius faces the territory pixel with R,

Expression string of binary characters operator, τ are the interval zoom factors of comparison, s _τBe piecewise function, be defined as follows:

$s_{\mathrm{\&tau;}}(I_c,I_k)=\left\{\begin{array}{c}01,\mathrm{if}{I}_k>(1+\mathrm{\&tau;})I_c,\\ 10,\mathrm{if}{I}_k<(1-\mathrm{\&tau;})I_c,\\ 00,\mathrm{otherwise}.\end{array}\right.---\left(2\right)$

The SILTP operator has the ability of very strong anti-local noise, has extraordinary robustness for shadow region, illumination sudden change etc. especially.

5. use the regression relation between support vector regression SVR algorithm learning sample proper vector and colony (vehicle/crowd) density at last, the textural characteristics that 4. angle point feature that 3. the sample characteristics vector is extracted by step and step are extracted combines.

Solve the new tool of machine learning problem by means of optimization method.Support vector machine is used for regression problem and is called support vector regression SVRM, and implementation algorithm is called support vector regression SVR algorithm.

When support vector regression SVR algorithm is realized returning estimation function, 3 characteristics are arranged: 1) utilize the linear function that defines in the higher dimensional space to estimate to return; 2) utilize linearity to minimize and realize returning estimation; 3) risk function of Cai Yonging partly is made up of regularization of being derived by structural risk minimization of experience sum of errors.

Given n group concern unknown sample data set H={ (x i, yi) } i=1 ... n, wherein xi is an input vector, yi is an expectation value, n is the sum of data point.SVM can utilize a Nonlinear Mapping, and data x is mapped to high-dimensional feature space H, and carries out linear proximity in this space.By Statistical Learning Theory as can be known, this function has following form:

Regression estimation problem is defined as offers as a tribute minimized problem to a loss function, and optimum regression function is by minimize the regularization risk functional under certain constraint condition:

${}_2{}^1{\left|\right|\mathrm{\&omega;}\left|\right|}^2+C_l^1\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{L}_{\mathrm{\&epsiv;}}(y_i,f\left(x_i\right))---\left(2\right)$

Wherein

Make function more smooth, be called regularization term; Second is the empiric risk functional, can determine by different loss functions, and constant C＞0, penalty is controlled the punishment degree to the sample that exceeds error.ε-insensitive loss function that this paper adopts,

L _ε(y _i，f(x _i))＝max(|y _i-f(x _i)-ε，0)

For L _ε(y _i, f (x _i)), if the absolute value of estimating output f (xi) and the deviation of desired output yi less than the time, it equals 0; Otherwise the absolute value that it equals deviation deducts ε, by introducing non-negative slack variable ξ _i, ξ _i* then top equation can be exchanged into:

$\min {}_2{}^1{\left|\right|\mathrm{\&omega;}\left|\right|}^2+C_l^1\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}({\mathrm{\&xi;}}_i+{{\mathrm{\&xi;}}_i}^{*})$

s.t.

ξ _i ^*≥0

Introduce Lagrangian function, we finally can obtain:

$\mathrm{\&omega;}-\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}(a_i-a_i^{*})x_i=0$

$f\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}(a_i-a_i^{*})j\left(x_i\right)\mathrm{gj}\left(x\right)+b$

$f\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}(a_i-a_i^{*})K(x_i,x)+b$

Can get thus,

Introduce kernel function K (xi, xj), above formula can become:

(xi xj) is vector x i to K, and xj is at feature space (x i) with the inner product (xj), i.e. K (xi, xj)=(xi) (xj).Can directly calculate on the input space by the computing that kernel function is all, the selection of kernel function is extremely important to support vector machine.

During by the regression relation between support vector regression SVR algorithm learning sample proper vector and the density, consider the perspective effect of video camera, the not of uniform size of picture near objects and distant objects causes, adopted different regression relation response functions in zones of different, made that colony (vehicle/crowd) density Estimation is more accurate.When carrying out the training of crowd massing detecting device, training set has 1000 images, picks up from the crowd zone at Tube and Train station, and test set comprises 500 images, comprises the crowd zone on subway and the square.Through test, population density detection method of the present invention has reached 89.63% correct classification rate on test set.

Above embodiment only is the present invention's a kind of embodiment wherein, and it describes comparatively concrete and detailed, but can not therefore be interpreted as the restriction to claim of the present invention.Should be pointed out that for the person of ordinary skill of the art without departing from the inventive concept of the premise, can also make some distortion and improvement, these all belong to protection scope of the present invention.Therefore, the protection domain of patent of the present invention should be as the criterion with claims.

Claims (5)

1. population density detection method, this method is to realize under the support of digital camera as sensor and digital signal processing chip, it is characterized in that: this method may further comprise the steps:

1) adopt digital camera as sensor, continuous acquisition fixed point zone forms digital video;

2) the angle point feature of digital video image frame extraction 1);

3) textural characteristics of digital video image frame extraction 1);

4) characteristics combination of resulting described textural characteristics and the estimation of the regression relation between the population density 1 resulting described angle point feature and 3 application 2))) the regional population density of described fixed point.

2. population density detection method according to claim 1 is characterized in that: step 2 wherein) in extract the angle point feature of described digital video image frame based on gradient method.

3. population density detection method according to claim 1 is characterized in that: the textural characteristics that wherein extracts described digital video image frame in the step 3) based on the constant local ternary feature SILTP of convergent-divergent.

4. population density detection method according to claim 1 is characterized in that: described regression relation draws by support vector regression algorithm SVR study great amount of samples.

5. population density detection method according to claim 1 is characterized in that: when estimating the population density in described fixed point zone in the step 4), the distance of the distance of the described digital video of foundation adopts different regression relation response functions.