CN101464944A - Crowd density analysis method based on statistical characteristics - Google Patents

Abstract

The invention discloses a method for analyzing crowd density base on statistical characteristics. The method comprises the following steps: video input and frame extraction are carried out; the mosaic image difference MID characteristics are extracted from the video-frequency frame sequence, and the subtle movements in the crowd are detected; the uniform distribution of the sequence time of the mosaic image difference MID characteristics is checked; the geometric correction is performed to the crowd and the scene with obvious perspective phenomenon, and a contribution factor of each picture element to the crowd density is obtained on the image plane; and the weighting process is performed to the crowd space area, so as to obtain the crowd density. Compared with the prior method, the method has no need of the reference background, also has no need of the background modeling and can self-adapt change of either morning or evening light, the algorithm is quite robust, and the application is convenient; the mathematical model is simple and effective, the spatial distribution and the size of the crowd can be accurately located, and the vivacity is strong; the calculation amount is small, and the method is suitable for real-time visual monitoring. The invention can be widely applied to the monitoring and the management of the public places with detained crowd density such as the public transportation, the subway, the square and the like.

Description

A kind of crowd density analysis method based on statistical nature

Technical field

The present invention relates to mode identification technology, particularly based on the crowd density analysis method of the robust of statistical nature.

Background technology

Along with computer technology and digital electric technology rapid development, the vision monitoring technology has obtained application more and more widely.Traditional visual monitor system often can only provide the function of video acquisition and storage, need the monitor staff to watch display screen to come unusual circumstance at Control Room, not only increased monitoring cost, and the visual fatigue of watching display screen to cause for a long time can make monitor staff's vigilance reduce, and makes supervisory system can not bring into play due effect at some crucial moment.At present, the intelligent vision monitoring technology is risen, and more and more receives the concern of medium.The intelligent vision monitoring technology is exactly to allow computer generation replace people's brain, allow camera replace people's eyes, analyze the image sequence that from camera, obtains by computer intelligence ground, content in the monitored scene is understood, compared remarkable advantages with traditional visual monitor system: on the one hand, the quantity that can significantly reduce the monitor staff reduces cost; On the other hand, can build the information of being concerned about (anomalous event, break in traffic rules and regulations, passenger flow etc.) that large-scale distributed intelligent monitor system obtains large area region in real time, make things convenient for resource (security personnel, police strength, public transport etc.) scheduling by networking.The crowd density analytical technology is a kind of intelligent vision monitoring technology that has very much using value, it can be applied to monitoring scenes such as subway, public transport platform, square, obtain crowd density information in real time, report to the police for overcrowding scene, the public transport platform crowd density analytic system of networking can also provide foundation for the scheduling of public transport resource.

A main difficult point of crowd density analytical technology is position and the size that how to obtain the crowd.People such as S.A Velastin propose to obtain the crowd by the background subtraction method, and this method disadvantage is the scene background picture that needs to take in advance width of cloth nobody; People such as Sheng-Fuu Lin propose to utilize machine learning, number of people detection technique to come the location crowd, and this method exists the scene of seriously blocking to tend to lose efficacy for too blocking up; People such as NiKos Paragios propose a kind of based on Markov Random Field crowd detection method, this method mathematical model complexity, and constraint condition is many, and parameter is many, uses inconvenience; People such as Xinyu Wu propose a kind of method based on texture analysis and study and estimate the scene number, this method need be to each monitored scene sampling, the artificial number of each sample of specifying is trained, but also requires the texture of monitoring area not abundant, uses inconvenience.

Another difficult point of crowd density analytical technology is after location crowd's position and size, how to estimate crowd's number or density.Number is often estimated relatively difficulty, under the situation that scene is blocked up especially, almost can not finish.Some actual application requirements wish that often intelligent monitor system provides the block up index of degree of a scene, and such as a decimal between 0～1,0 represents almost no one, almost has been full of the people in the 1 expression scene; When scene crowding index is that system can give the alarm automatically after time of 1 surpassing certain hour.

Summary of the invention

Difficult point is after location crowd's position and size in the prior art in order to solve, how to estimate crowd's the number or the problem of density, the purpose of this invention is to provide for this reason and a kind ofly can accurately locate the crowd, rationally estimate the crowd density analysis method of crowded degree.

To achieve these goals, the crowd density analysis method that the present invention is based on the robust of statistical nature comprises step:

Step 1: adopt video camera to catch video, the mode that original input video takes every several frames to get 1 frame is done the frame extraction, reduce frame per second;

Step 2: from sequence of frames of video, extract mosaic image difference MID feature, be used for the crowd's of detecting trickle motion;

Step 3: the check mosaic image difference MID time evenly distributes, and is used for determining whether the motion in the corresponding mosaic blockage is that stable crowd produces;

Step 4: utilize motion that gridding method calculates the crowd in spatial distributions, and will feed back to step 3 with the relevant parameter that is used for the time distribution inspection of space distribution;

Step 5: the crowd's scene with obvious perspective phenomenon is done geometry correction, and each pixel is to the contribution factor of crowd density on the acquisition plane of delineation;

Step 6: utilize contribution factor that the crowd's spatial area that obtains in the step 4 is done weighted, the crowd's spatial area after the weighting is crowd density with the ratio of the area of area-of-interest.

According to the embodiment of the invention, described mosaic image difference MID is characterized as:

${\mathrm{MID}}_t(m,n)=\left\{\begin{array}{cc}1,& \mathrm{if}{\left|\right|M_t(m,n)-M_{t-1}(m,n)\left|\right|}_{\∞}>T_t\\ 0,& \mathrm{else}\end{array}\right.$

Wherein: $M_t(m,n)=\frac{1}{{L_M}^2}\underset{i=m{L}_M}{\overset{(m+1)L_M-1}{\mathrm{\Σ}}}\underset{j={\mathrm{nL}}_M}{\overset{(n+1)L_M-1}{\mathrm{\Σ}}}{I}_t(i,j)$

‖ ‖ _∞Represent each component absolute value the maximum of a vector; I _t(i, j) the RGB vector of the capable j row of expression t two field picture i; L _MThe length of side of expression mosaic square; M _t(m, n) capable n the mosaic zone of expression m is at the average color of t frame, MID _t(m n) has characterized the difference of the average color of capable n mosaic zone t frame of m and t-1 frame, and this difference is quantized into two ranks: if the component of the absolute value maximum of difference value vector is greater than certain threshold value T _tThen note does 1, otherwise note does 0.

According to the embodiment of the invention, described threshold value T _tBy being calculated as:

T _t＝T ₀α _t，

Wherein: α _t=1+ (B _t-127)/255,

B _tBe the mean flow rate of t two field picture, T ₀Get default value.

According to the embodiment of the invention, the described mosaic image difference MID time evenly distributes, adopt three statistical parameters of the MID sequence in check a period of time: the MID eigenwert is that average, the MID eigenwert of time of 1 is that the variance of time of 1 and MID eigenwert are whether the sheet number of 1 probability non-zero timeslice satisfies certain condition and determine that whether this MID sequence obeys temporal even distribution, uses indicative function U _t(m n) is expressed as:

$U_t(m,n)=\left\{\begin{array}{cc}1& \mathrm{if}{N}_{\mathrm{Nz}}>N_s/2\mathrm{and}|\mathrm{Mean}-(1+N_s)/2|<N_s/5\mathrm{and\; Var}>{\mathrm{\&sigma;}}_t(m,n)\\ 0& \mathrm{else}\end{array}\right.$

Wherein: N _sThe sheet number of the timeslice that is divided for the time period of being analyzed;

N _NZ=| { P _l≠ 0|l=1...N _s|, expression MID eigenwert is the sheet number of 1 probability non-zero timeslice;

$\mathrm{Mean}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{\mathrm{lP}}_l,$ Expression MID eigenwert is the average of time of 1;

$\mathrm{Var}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{(l-\mathrm{Mean})}^2{P}_l,$ Expression MID eigenwert is the variance of time of 1;

P _lBe after the MID sequence is divided into timeslice, the MID eigenwert is 1 probability in each timeslice;

The threshold value σ of time variance _t(m, n) relevant with this mosaic blockage location in space; Zonule according to this mosaic blockage place all is that people or no one are arranged in the piecewise analysis time in history, and the size of variance threshold values is set.

According to the embodiment of the invention, the zone whether capable q the grid of described grid computing p finally is included into the crowd is its G _t(p, q) whether value is 1, G _t(p, q) by following formula decision:

$G_t(p,q)=\left\{\begin{array}{cc}1& \mathrm{if}\left|\left\{M(m,n)\right|U_t(m,n)=1,M(m,n)\&Subset;G(p,q)\}\right|>0\\ 0& \mathrm{else}\end{array}\right.$

Wherein | { } | the number of a set element of expression, M (m, n) capable n the mosaic zone of expression m, G (p, q) capable q the net region of expression p; G _t(p is that 1 implication is q): if the t frame constantly grid G (p has the zone that is marked as the people in the mosaic step in front of and above quantity in q), and then whole grid all is marked as people's zone at the t frame.

According to the embodiment of the invention, described geometry correction is that two straight lines of delimitation obtain the end point in the scene, calculates the contribution factor S of each pixel then _C(x, y):

$S_C(x,y)={\left(\frac{y_r-y_v}{y-y_v}\right)}^2$

Wherein (x y) is pixel coordinate, y _rThe y coordinate of a horizontal reference line of choosing arbitrarily on the presentation video plane, y _vThe y coordinate of expression disappearance point.

According to the embodiment of the invention, the net result of described crowd density

For the ratio of the area of the area-of-interest after the foreground area after the weighting and the weighting is:

$D_t^P=\frac{\underset{\mathrm{FG}}{\mathrm{\&Sigma;}}{S}_C(x,y)}{\underset{\mathrm{BG}}{\mathrm{\&Sigma;}}{S}_C(x,y)}$

Wherein: FG={ (x, y) | G _t(p, q)=1, (x, y) ∈ (G (p, q) ∩ D _ROI), (x, y) remarked pixel coordinate, D _ROIThe interested image-region that expression delimited, the physical meaning of FG be expression all not only in interesting areas but also the set of all pixel coordinates in being marked as people's grid; BG={ (x, y) | (x, y) ∈ D _ROI, represent the pixel coordinate set of all area-of-interests.

The present invention compares with the up-to-date method of delivering both at home and abroad at present has several obvious advantages:

1) do not need reference background image, also do not need background modeling, have nothing to do with scene texture, and can adapt to the slow variation of illumination, relatively robust is used conveniently.

2) mathematical model is simple, can accurately locate crowd's space distribution and size, and intuitive is strong; Calculated amount is little, is fit to the real-time vision monitoring, and the computation process of crowd density has meaning directly perceived.

3) calculated amount is little, can satisfy the requirement that real-time video is handled.

The present invention can be widely used in monitoring and the management that the public place that the crowd is dense is detained on public transport, subway and square etc.

Description of drawings

Fig. 1 technical solution of the present invention process flow diagram

Fig. 2 geometry correction synoptic diagram of the present invention

Bus stop, Fig. 3 Beijing scene synoptic diagram

Fig. 4 t ₁Moment platform image

Fig. 5 t ₁The MID characteristic sequence is obeyed the mosaic block that the even time distributes constantly

Fig. 6 t ₁The crowd's that moment gridding method is calculated space distribution (crowd density is 0.6 after the geometry correction)

Fig. 7 t ₂Moment platform image

Fig. 8 t ₂The MID characteristic sequence is obeyed the mosaic block that the even time distributes constantly

Fig. 9 t ₂The crowd's that moment gridding method is calculated space distribution

Embodiment

Describe each related detailed problem in the technical solution of the present invention in detail below in conjunction with accompanying drawing.Be to be noted that described embodiment only is intended to be convenient to the understanding of the present invention, and it is not played any qualification effect.

Thought main points of the present invention are:

1) stable crowd necessarily has various trickle motions, such as whispering to each other, turn round, move etc.

2) from the occupied space of time and crowd, these trickle motions are the even distribution on presentative times and the space substantially.

3) on the basis that detects these trickle motions, check they whether the obedience time go up and the space on even distribution then can determine crowd's distribution range.

4) according to the perspective principle detected crowd's area is made weighted, obtain more accurate crowd density.

Whole technical proposal process flow diagram of the present invention as shown in Figure 1.Below involved ins and outs in the invention are illustrated, provided the experimental result under the public transport platform scene at last.

1. video input and frame extract

Adopt video camera to catch video, after importing computer or flush bonding processor into, the mode that original input video takes every several frames to get 1 frame is done the frame extraction, and the digital image sequence that transmits from camera was generally for 25 frame/seconds, and the movable information among the detection crowd does not need so high frame per second.Before doing the crowd density analysis, can sample, extract a frame such as per 5 frames to frame sequence.Frame extracts not only can reduce frame per second, reduces calculated amount, and helps motion detection, because the movement velocity of the little motion among the crowd often relatively slowly, too fast frame per second can cause the difference between frame and the frame too little, is unfavorable for motion detection.Generally, it is proper that frame per second was reduced to for 5 frame/seconds.

(2.MID mosaic image difference) feature extraction

From sequence of frames of video, extract mosaic image difference MID feature, be used for the crowd's of detecting trickle motion.

MID (Mosaic Image Difference) is the effective ways of the motion that produces among a kind of crowd of sign.

At first to be divided into a series of length of side be L to the t two field picture that will import _MThe mosaic square, and calculate the RGB color value in each mosaic square zone, formula is as follows:

$M_t(m,n)=\frac{1}{{L_M}^2}\underset{i=m{L}_M}{\overset{(m+1)L_M-1}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{nL}}_M}{\overset{(n+1)L_M-1}{\mathrm{\&Sigma;}}}{I}_t(i,j)---\left(1\right)$

Wherein, I _t(i, j) the RGB vector of the capable j row of expression t two field picture i; L _MThe length of side of expression mosaic square.M _t(m, n) average color in capable n the mosaic zone of m of expression t frame mosaic image.M, n, i, the span of j is relevant with picture size.

The computing formula of mosaic image difference MID then is:

${\mathrm{MID}}_t(m,n)=\left\{\begin{array}{cc}1,& \mathrm{if}{\left|\right|M_t(m,n)-M_{t-1}(m,n)\left|\right|}_{\&infin;}>T_t\\ 0,& \mathrm{else}\end{array}\right.---\left(2\right)$

Wherein, ‖ ‖ _∞Represent each component absolute value the maximum of a vector; MID _t(m n) has characterized the difference of the average color of capable n mosaic zone t frame of m and t-1 frame, and this difference is quantized into two ranks: if the component of the absolute value maximum of difference value vector greater than certain threshold value T _tThen note does 1, otherwise note does 0.

Threshold value T _t=T ₀α _t, T ₀Be constant, generally get 15, α _tBe that an adaptive threshold is adjusted the factor, it changes with the overall brightness of t frame:

α _t＝1+(B _t-127)/255， (3)

B _tBe the mean flow rate of t two field picture, T ₀Get default value.

3.MID even distribution inspection of time

The check mosaic image difference MID time evenly distributes, and is used for determining whether the motion that produces in the corresponding mosaic blockage is the motion that stable crowd produces.If not the motion that stable crowd produces, then can not think that the crowd is arranged in this mosaic square., but do not stay through a mosaic zone such as a people, also can to produce some values be 1 MID feature in this mosaic zone at short notice, and this situation just need evenly distribute by the time of check MID and get rid of.In addition, it can all not be 1 at every frame also that the MID eigenwert of the mosaic blockage of stablizing the crowd is arranged, because it is motionless that the people among the crowd may keep in a period of time, need the time distribution character of the MID sequence by checking this mosaic blockage to predict whether this mosaic is that the crowd is arranged at present frame in this case.

Supposing has stable crowd to exist in the t frame mosaic blockage that the capable n of m is listed as constantly, and according to hypothesis, the MID feature of capable n the mosaic of m is at t-N _MFrame is to the interior during this period of time sequence { MID of t frame _k(m, n) | t-N _M＜k≤t} should be that a string value is 0 and 1 equally distributed random number, wherein N _MBe the length of the sequence analyzed (it is different to look concrete scene, about general 75 frames, general 15 seconds).In order to check this MID sequence whether to obey temporal even distribution, this sequence is divided into N by the time _sIndividual timeslice (visual field scape and different, general every period general lasting 3 seconds), and check three statistics of this MID sequence: the MID eigenwert is the sheet number of the timeslice of 1 probability non-zero, and the MID eigenwert is 1 the average moment, and the MID eigenwert is the variance in moment of 1.Order:

$S=\underset{k=i-N_M}{\overset{t}{\mathrm{\&Sigma;}}}{\mathrm{MID}}_k(m,n)---\left(4\right)$

$P_l=\frac{1}{S}\underset{k=t-N_M+(l-1)\&times;N_M/N_s}{\overset{k=t-N_M+l\&times;N_M/N_s-1}{\mathrm{\&Sigma;}}}{\mathrm{MID}}_k(m,n),l=1...N_s---\left(5\right)$

N _NZ＝|{P _l≠0|l＝1...N _s}| (6)

$\mathrm{Mean}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{\mathrm{lP}}_l---\left(7\right)$

$\mathrm{Var}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{(l-\mathrm{Mean})}^2{P}_l---\left(8\right)$

Wherein | { } | the number of a set element of expression, S represents t-N _MFrame to the t frame during this period of time in, the MID eigenwert of the capable n row of m mosaic is 1 total degree that occurs, P _lL timeslice MID eigenwert in time period of being analyzed of expression is 1 probability, and the MID eigenwert is 1 the average moment in the time period that Mean represents to be analyzed, and Var represents that interior MID eigenwert of the time period of being analyzed is the variance in moment of 1, N _NZThe sheet number of the probability non-zero timeslice that mosaic changes in the time period that expression is analyzed, P in theory _l=1/N _s, Mean=(1+N _s)/2, $\mathrm{Var}=(N_s^2-1)/12.$

Define an indicative function U _t(m, n) (in the mosaic blockage of the capable n row of value 1 expression t frame m the people is arranged, otherwise the expression no one):

$U_t(m,n)=\left\{\begin{array}{cc}1& \mathrm{if}{N}_{\mathrm{Nz}}>N_s/2\mathrm{and}|\mathrm{Mean}-(1+N_s)/2|<N_s/5\mathrm{and\; Var}>{\mathrm{\&sigma;}}_t(m,n)\\ 0& \mathrm{else}\end{array}\right.---\left(9\right)$

Wherein: N _sThe sheet number of the timeslice that is divided for the time period of being analyzed;

N _NZ=| { P _l≠ 0|l=1...N _s|, expression MID eigenwert is the sheet number of 1 probability non-zero timeslice;

$\mathrm{Mean}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{\mathrm{lP}}_l,$ Expression MID eigenwert is the average of time of 1;

$\mathrm{Var}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{(l-\mathrm{Mean})}^2{P}_l,$ Expression MID eigenwert is the variance of time of 1;

P _lBe after the MID sequence is divided into timeslice, the MID eigenwert is 1 probability in each timeslice.

The threshold value σ of time variance _t(m, n) relevant with this mosaic blockage location in space, when all there is the people zonule at a mosaic square place (grid of mentioning later) in a period of time in history, then the time variance threshold value can be smaller; When in a period of time in history during the no one, the time variance threshold value can be provided with bigger; Provide concrete computing formula in the back.

4. gridding method is calculated crowd's space distribution

Utilize motion that gridding method calculates the crowd in spatial distributions, and will feed back to step 3 with the relevant parameter that is used for the time distribution inspection of space distribution; Indicative function U _t(m, mosaic n)=1 is a sampling in the occupied zone of crowd only, because crowd's shape is unpredictable, estimate crowd's spatial area, and a kind of feasible scheme is a gridding method.According to concrete scene, it is W that the plane of delineation is divided into a series of sizes _g* H _gGrid, W generally speaking _g* H _g=12 * 12.

If capable q the grid of p has the people then to be designated as at the t frame: G _t(p, q)=1, otherwise G _t(p, q)=0.With M (m, n) expression m capable n mosaic zone, with G (p q) represents capable q the net region of p, then:

$G_t(p,q)=\left\{\begin{array}{cc}1& \mathrm{if}\left|\left\{M(m,n)\right|U_t(m,n)=1,M(m,n)\&Subset;G(p,q)\}\right|>0\\ 0& \mathrm{else}\end{array}\right.---\left(10\right)$

Wherein | { } | the number of a set element of expression, M (m, n) capable n the mosaic zone of expression m, G (p, q) capable q the net region of expression p; G _t(p is that 1 implication is q): if the t frame constantly grid G (p has the zone that is marked as the people in the mosaic step in front of and above quantity in q), and then whole grid all is marked as people's zone at the t frame.

Suppose that interested image-region is D _ROI, the area of a grid is designated as S _G, the t frame total area A that on the plane of delineation, occupies of crowd constantly so _tFor:

$A_t=S_G\underset{G(p,q)\&Subset;D_{\mathrm{ROI}}}{\mathrm{\&Sigma;}}{G}_t(p,q)---\left(11\right)$

When not needing to consider the influencing of perspective projection, crowd density D _tMay be calculated:

$D_t=\frac{A_t}{S_D}---\left(12\right)$

S wherein _DBe region D _ROIArea.

Above-mentioned parameter σ _t(m is n) with preface { G _k(p, q) | t-N _G≤ k＜t, M (m, n) ∈ G (p, q) } relevant:

${\mathrm{\&sigma;}}_t(m,n)={\mathrm{\&sigma;}}_0+{\mathrm{\&sigma;}}_1(1-\frac{1}{N_G}\underset{k=t-N_G}{\overset{t-1}{\mathrm{\&Sigma;}}}{G}_k(p,q))---\left(13\right)$

N wherein _G=2N _M, σ ₀, σ ₁Be constant, visual field scape and deciding is general ${\mathrm{\&sigma;}}_0=\frac{(N_s^2-1)/12}{6},$ ${\mathrm{\&sigma;}}_1=\frac{(N_s^2-1)/12}{3}.$

5. geometry correction

Crowd's scene with obvious perspective phenomenon is done geometry correction, and each pixel is to the contribution factor of crowd density on the acquisition plane of delineation.

When there is more serious perspective phenomenon (same object in the projection of scene on the plane of delineation, closely seem big from video camera, far seem little from video camera) time, formula (12) can not characterize crowd's density well, therefore need make weighted to the contribution of different pixels on the plane of delineation.Suppose that ground is the plane, the people is perpendicular to ground.

Shown in accompanying drawing 2 geometry correction synoptic diagram of the present invention, suppose at horizontal reference line y _rThe plane P at place ₅P ₁P ₄P ₈On a size is arranged is Δ W _r* Δ H _rObject O _A, the object O of identical size _BBe put into the plane P at horizontal line y place ₆P ₂P ₃P ₇The size of correspondence is Δ W * Δ H when last, when these two object areas level off to zero the time O _AWith O _BThe ratio of area is the pixel and the reference line y at horizontal line y place _rThe pixel at place is to the ratio of the contribution of crowd density.According to the principle of perspective imaging, establish the some P that disappears _vCoordinate be (x _v, y _v), reference line is y=y _r=H/2 can release O by the simple geometric relation _AWith O _BThe ratio of area is:

$S_C(x,y)={\left(\frac{y_r-y_v}{y-y_v}\right)}^2---\left(14\right)$

This is any one pixel I (x, contribution factor y) on the plane of delineation just.(x y) represents any one pixel coordinate, y _rThe y coordinate of a selected horizontal reference line on the presentation video plane, y _vThe y coordinate of expression disappearance point.

X _WY _WZ _WExpression three-dimensional world coordinate system.

6. crowd density calculates

Utilize contribution factor that the crowd's spatial area that obtains in the step 4 is done weighted, the crowd's spatial area after the weighting is crowd density with the ratio of the area of area-of-interest.The note set

BG={ (x, y) | (x, y) ∈ D _ROI, then have

Crowd's area after the geometry correction

For:

$A_t^P=\underset{\mathrm{FG}}{\mathrm{\&Sigma;}}{S}_C(x,y)$

Crowd density after the geometry correction

For:

$D_t^P=\frac{A_t^P}{\underset{\mathrm{BG}}{\mathrm{\&Sigma;}}{S}_C(x,y)}---\left(16\right)$

Wherein: FG={ (x, y) | G _t(p, q)=1, (x, y) ∈ (G (p, q) ∩ D _ROI), (x, y) remarked pixel coordinate, D _ROIThe interested image-region that expression delimited.The physical meaning of FG be expression all not only in interesting areas but also the set of all pixel coordinates in being marked as people's grid;

BG={ (x, y) | (x, y) ∈ D _ROI, represent the pixel coordinate set of all area-of-interests.

More than being exactly the detailed description of the invention process step, is example with bus stop, Beijing scene below, provides experimental result.As shown in Figure 3, certain bus stop is a scene that typically has the perspective phenomenon, and point coordinate is (144 ,-10) by obtaining disappearing after the projection straight line of ground level parallel lines on the plane of delineation in the artificial setting three dimensions.

After frame extracted, the actual frame per second of this experiment was about for 5 frame/seconds, and some parameters in addition are as follows:

Threshold parameter T in the formula of calculating MID feature ₀=15;

The mosaic blockage length of side: L _M=4;

The grid square length of side: L _G=12;

The length of the sequence of MID even distribution inspection of characteristic time: N _M=75;

The number of the timeslice of MID even distribution inspection of characteristic time: N _S=5;

Fig. 4 is t ₁The image of platform constantly, this moment, people's distribution approximately occupied more than whole public transport platform general; Fig. 5 is t ₁Detected MID characteristic sequence is obeyed the mosaic blockage set that the even time distributes constantly; Fig. 6 is t ₁The crowd's that moment grid gridding method obtains space distribution, the crowd density that obtains after the geometry correction is 0.6, meets the observations of human eye substantially.

Fig. 7 is t ₂The image of moment platform, this moment, people's distribution occupied almost whole public transport platform; Fig. 8 is t ₂Detected MID characteristic sequence is obeyed the mosaic blockage set that the even time distributes constantly; Fig. 9 is t ₂The crowd's that moment gridding method obtains space distribution, the crowd density after the geometry correction is 0.95, meets the eye-observation result.

Experiment showed, that crowd density analysis method proposed by the invention can locate crowd's space distribution exactly, and reasonably calculate crowd density; Compare with other existing crowd density computing method, this method intuitive is strong, and restricted few, calculated amount is little, and the crowd who is fit to public places such as public transport, subway and square monitors.

The above; only be the embodiment among the present invention; but protection scope of the present invention is not limited thereto; anyly be familiar with the people of this technology in the disclosed technical scope of the present invention; can understand conversion or the replacement expected; all should be encompassed in of the present invention comprising within the scope, therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.

Claims (7)

1, a kind of crowd density analysis method based on statistical nature is characterized in that, comprises step:

Step 1: adopt video camera to catch video, the mode that original input video takes every several frames to get 1 frame is done the frame extraction, reduce frame per second;

Step 2: from sequence of frames of video, extract mosaic image difference MID feature, be used for the crowd's of detecting trickle motion;

Step 3: the check mosaic image difference MID time evenly distributes, and is used for determining whether to have in the corresponding mosaic blockage stable crowd's motion;

Step 4: utilize motion that gridding method calculates the crowd in spatial distributions, and will feed back to step 3 with the relevant parameter that is used for the time distribution inspection of space distribution;

Step 5: the crowd's scene with obvious perspective phenomenon is done geometry correction, and each pixel is to the contribution factor of crowd density on the acquisition plane of delineation;

Step 6: utilize contribution factor that the crowd's spatial area that obtains in the step 4 is done weighted, the crowd's spatial area after the weighting is crowd density with the ratio of the area of area-of-interest.

According to right 1 described crowd density analysis method, it is characterized in that 2, described mosaic image difference MID is characterized as:

${\mathrm{MID}}_t(m,n)=\left\{\begin{array}{cc}1,& \mathrm{if}{\left|\right|M_t(m,n)-M_{t-1}(m,n)\left|\right|}_{\&infin;}>T_t\\ 0,& \mathrm{else}\end{array}\right.$

Wherein:

$M_t(m,n)=\frac{1}{{L_M}^2}\underset{i={\mathrm{mL}}_M}{\overset{(m+1)L_M-1}{\mathrm{\&Sigma;}}}\underset{j={\mathrm{nL}}_M}{\overset{(n+1)L_M-1}{\mathrm{\&Sigma;}}}{I}_t(i,j)$

‖ ‖ _∞Represent each component absolute value the maximum of a vector; I _t(i, j) the RGB vector of the capable j row of expression t two field picture i; L _MThe length of side of expression mosaic square; M _t(m, n) capable n the mosaic zone of expression m is at the average color of t frame, MID _t(m n) has characterized the difference of the average color of capable n mosaic zone t frame of m and t-1 frame, and this difference is quantized into two ranks: if the component of the absolute value maximum of difference value vector is greater than certain threshold value T _tThen note does 1, otherwise note does 0.

3, according to right 2 described crowd density analysis methods, it is characterized in that described threshold value T _tBy being calculated as:

T _t＝T ₀α _t，

Wherein: α _t=1+ (B _t-127)/255,

B _tBe the mean flow rate of t two field picture, T ₀Get default value.

4, according to right 2 described crowd density analysis methods, it is characterized in that, the described mosaic image difference MID time evenly distributes, adopt three statistical parameters of the MID sequence in check a period of time: the MID eigenwert is that average, the MID eigenwert of time of 1 is that the variance of time of 1 and MID eigenwert are whether the sheet number of 1 probability non-zero timeslice satisfies certain condition and determine that whether this MID sequence obeys temporal even distribution, uses indicative function U _t(m n) is expressed as:

$U_t(m,n)=\left\{\begin{array}{cc}1& \mathrm{if}{N}_{\mathrm{Nz}}>N_s/2\mathrm{and}|\mathrm{Mean}-(1+N_s)/2|<N_s/5\mathrm{and\; Var}>{\mathrm{\&sigma;}}_t(m,n)\\ 0& \mathrm{else}\end{array}\right.$

Wherein: N _sThe sheet number of the timeslice that is divided for the time period of being analyzed;

N _NZ=| { P _l≠ 0|l=1...N _s|, expression MID eigenwert is the sheet number of 1 probability non-zero timeslice;

$\mathrm{Mean}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{\mathrm{lP}}_l,$ Expression MID eigenwert is the average of time of 1;

$\mathrm{Var}=\underset{l=1}{\overset{N_s}{\mathrm{\&Sigma;}}}{{(l-\mathrm{Mean})}^{2}P}_l,$ Expression MID eigenwert is the variance of time of 1;

P _lBe after the MID sequence is divided into timeslice, the MID eigenwert is 1 probability in each timeslice;

The threshold value σ of time variance _t(m, n) relevant with this mosaic blockage location in space; When the zonule at this mosaic blockage place all is that people or no one are arranged in the piecewise analysis time in history, the size of variance threshold values is set.

5, according to right 1 described crowd density analysis method, it is characterized in that: the zone whether capable q the grid of described grid computing p finally is included into the crowd is its G _t(p, q) whether value is 1, G _t(p, q) by following formula decision:

$G_t(p,q)=\left\{\begin{array}{cc}1& \mathrm{if}\left|\right\{M(m,n)|U_t(m,n)=1,M(m,n)\&Subset;G(p.q)\}|>0\\ 0& \mathrm{else}\end{array}\right.$

Wherein | { } | the number of a set element of expression, M (m, n) capable n the mosaic zone of expression m, G (p, q) capable q the net region of expression p; G _t(p is that 1 implication is q): if the t frame constantly grid G (p has the zone that is marked as the people in the mosaic step in front of and above quantity in q), and then whole grid all is marked as people's zone at the t frame.

6, according to right 1 described crowd density analysis method, it is characterized in that: described geometry correction is that two straight lines of delimitation obtain the end point in the scene, calculates the contribution factor S of each pixel then _C(x, y):

$S_C(x,y)={\left(\frac{y_r-y_v}{y-y_v}\right)}^2$

Wherein (x y) is pixel coordinate, y _rThe y coordinate of a horizontal reference line of choosing arbitrarily on the presentation video plane, y _vThe y coordinate of expression disappearance point.

7, according to right 1 described side's crowd density analytic approach, it is characterized in that the net result of described crowd density

For the ratio of the area of the area-of-interest after the foreground area after the weighting and the weighting is:

$D_t^P=\frac{\underset{\mathrm{FG}}{\mathrm{\&Sigma;}}{S}_C(x,y)}{\underset{\mathrm{BG}}{\mathrm{\&Sigma;}}{S}_C(x,y)}$

Wherein: FG={ (x, y) | G _t(p, q)=1, (x, y) ∈ (G (p, q) ∩ D _ROI), (x, y) remarked pixel coordinate, D _ROIThe interested image-region that expression delimited, FG represent that all are not only in interesting areas but also the set of all pixel coordinates in being marked as people's grid; BG={ (x, y) | (x, y) ∈ D _ROI, represent the pixel coordinate set of all area-of-interests.

Images