CN104091198A - Pedestrian flow statistic method based on ViBe - Google Patents

Abstract

The invention belongs to the technical field of computer video image processing, and discloses a pedestrian flow statistic method based on the ViBe. The method comprises the steps of acquiring a video streaming image in a monitored area as an input image, sampling the movement area of the input image to obtain a sampling mask Mt of the movement area, carrying out contour priority detection on the sampling mask based on the ViBe algorithm to obtain a moving pedestrian area, and calculating the flow of pedestrians passing the monitored area based on the linear regression analysis method. According to the method, contour priority detection is carried out on the moving pedestrians included in the sampling mask based on the ViBe algorithm, then the flow of the pedestrians passing the monitored area is obtained based on the linear regression analysis method, and therefore the efficiency and the accuracy of counting the number of the pedestrians in a public area are effectively improved.

Description

People flow rate statistical method based on ViBe

Technical field

The invention belongs to Video Image processing technology field, particularly a kind of people flow rate statistical method based on ViBe, accurately adds up for the pedestrian's quantity to public domain.

Background technology

In a lot of industries, people information can provide crucial foundation for people's flow management, resource management, management decision.For example, at subway station, by people counting, can understand in real time stream of people's size of each website, flexible dispatching subway train density, implement people's current control, the crowded regional information of real-time release, is conducive to strengthen crowd's conevying efficiency, guarantees that metro operation is steadily effective.Flow of the people is also related to the safety problem in crowded place, and effective crowd's quantity in controlling filed can in emergency circumstances be dredged rapidly crowd in fire alarm etc., the situation such as avoids trampling, push and occurs.

ViBe (Visual background extractor) algorithm is visual background extraction algorithm, and it is the modeling of a kind of Pixel-level video background or foreground detection algorithm.Its occupancy to Computing resource is smaller, and can effectively suppress shade, camera or video camera and rock the impact that foreground detection is caused.But at present, in the prior art, pedestrian detection technology based on ViBe inevitably can be subject to the impact of background interference and noise, make the pedestrian detection technology based on ViBe algorithm can not detect exactly motion pedestrian target, thereby cause poor effect that the pedestrian's number in public domain is added up.

Summary of the invention

The object of the invention is to disclose a kind of people flow rate statistical method based on ViBe, in order to improve pedestrian being carried out in public domain efficiency and the accuracy of demographics.

For achieving the above object, the invention provides a kind of people flow rate statistical method based on ViBe, the method comprises the following steps:

S1, obtain guarded region video streaming image as input picture;

S2, input picture is carried out to moving region sampling, obtain the sampling mask M of moving region ^t;

S3, by ViBe algorithm, described sampling mask is carried out to profile and preferentially detect, pedestrian region obtains moving;

S4, utilize linear regression analysis method statistic by the flow of the people in described guarded region.

As a further improvement on the present invention, described step S1 is specially: the video streaming image that obtains guarded region by video camera is as input picture, described guarded region be positioned at video camera under.

As a further improvement on the present invention, described step S2 is specially:

According to former frame input picture being detected to temporal information, spatial information and the frequency information comprising in the result of resulting motion pedestrian region, calculate moving region probability graph, and by least one sampling policy to obtain the mask M that samples ^t.

As a further improvement on the present invention, described sampling policy comprises: Random Discrete sampling policy, spatial spread sampling policy and sudden change pixel sampling strategy.

As a further improvement on the present invention, " profile preferentially detects " in described step S3 is specially:

Calculate the order of contact of each pixel in described sampling mask, and according to the detection order of the large pixel of little pixel the second priority processing order of contact of the first priority processing order of contact, little by little from the peripheral profile of sampling mask, to its inside, carry out sequence detection.

Compared with prior art, the invention has the beneficial effects as follows: in the present invention, the motion pedestrian who based on the preferential method detecting of profile, sampling mask is comprised by Vibe algorithm detects, then utilize linear regression analysis method to obtain by the flow of the people in guarded region, effectively improved pedestrian being carried out in public domain efficiency and the accuracy of demographics.

Accompanying drawing explanation

Fig. 1 is the process flow diagram that the present invention is based on the people flow rate statistical method of ViBe;

Fig. 2 is the principle of work schematic diagram of performing step S1;

Fig. 3 is for to sample to obtain sampling mask M to moving region ^tprocess principle figure;

Fig. 4 is the sampling mask M of moving region ^tand the schematic diagram of the original order of contact of each pixel;

Fig. 5 a to Fig. 5 c is that the order of contact of pixel in the preferential detection computations process of profile changes schematic diagram;

Fig. 6 is that motion pedestrian contour detects complete schematic diagram;

Fig. 7 is motion for line people detection result schematic diagram.

Embodiment

Below in conjunction with each embodiment shown in the drawings, the present invention is described in detail; but should be noted that; these embodiments are not limitation of the present invention; those of ordinary skills are according to these embodiment institute work energy, method or structural equivalent transformation or alternative, within all belonging to protection scope of the present invention.

Shown in ginseng Fig. 1, Fig. 1 is the schematic flow sheet that the present invention is based on the pedestrian counting method of optical flow method.

In the present embodiment, should comprise the following steps by the pedestrian counting method based on optical flow method:

First perform step S1: obtain the video streaming image of guarded region as input picture.

About the method for the pedestrian's demographics in public domain, conventional have method, the method based on shape information, the method based on pedestrian dummy, structural element, the method for stereoscopic vision, methods such as the method for neural network, small echo and support vector machine based on kinetic characteristic.

Shown in ginseng Fig. 2, the pedestrian counting method that the present invention is based on optical flow method is based on the vertical shooting of video camera 10 and applicable to outdoor environment and indoor environment.In the present embodiment, this step S1 is specially: the video streaming image that obtains guarded region 30 by video camera 10 is as input picture, described guarded region 30 be positioned at video camera 10 under.

Concrete, video camera 10 be arranged on gateway 20 directly over, pedestrian can walk up and down in the direction of arrow 201 in gateway 20.The guarded region 30 that video camera 10 obtains can cover the Zone Full of gateway 20 completely.This gateway 20 can be arranged on the market that need to add up pedestrian's number, garage, bank etc. to be needed in the front door or corridor in key monitoring place.

It should be noted that, the best results of the present invention when video camera 10 vertically faces guarded region 30, can also face toward by video camera 10 region that need to carry out pedestrian's number counting statistics certainly obliquely, to cover whole guarded region 30 by video camera 10.

In the present embodiment, this guarded region 30 is rectangle; Can certainly be square or circular or other shapes.Video camera 10 be positioned at guarded region 30 central point 301 directly over, now this guarded region 30 be positioned at video camera 10 under.

Then, execution step S2: input picture is carried out to moving region sampling, obtain the sampling mask M of moving region ^t.

Shown in ginseng Fig. 3, I ^trepresent t original image constantly, represent moving region probability graph, M ^trepresent sampling masked areas, D ^trepresent present frame moving region testing result, postpone D ^t-1represent former frame moving region testing result.

In the present embodiment, the probability of motion figure of moving region sampling by following three kinds of information, obtained, that is: temporal information, spatial information and frequency information.

Concrete, the pixel of establishing in input picture is n:

Temporal information: if certain pixel belongs to moving region in former frame, this pixel more may belong to moving region at present frame so.

$M_T^t\left(n\right)=(1-{\mathrm{\α}}_T)M_T^{t-1}\left(n\right)+{\mathrm{\α}}_T{D}^t\left(n\right);$

Wherein, α _trefer to learning rate, more approach 1, this pixel is that the probability of moving region is larger.

Spatial information: if the surrounding pixel point of certain pixel is all pixels of moving region, this pixel more may belong to moving region so, its probability is directly proportional to the number of the pixel of moving region around.

$M_S^t\left(n\right)=(1-{\mathrm{\α}}_s)M_s^{t-1}\left(n\right)+{\mathrm{\α}}_s{S}^t\left(n\right);$

Wherein, $S^t\left(n\right)=\frac{1}{{\mathrm{\ω}}^2}\underset{i\∈N\left(n\right)}{\mathrm{\Σ}}{D}^t\left(i\right).$

Concrete, α _srefer to study, N (n) represents the vicinity points within the scope of ω * ω, more approach 1, this pixel is that the probability of moving region is larger.

Frequency information: if the probability that certain pixel belongs to moving region or belongs to background area changes too frequently, to belong to the probability of noise larger for this pixel so.

$M_F^t\left(n\right)=(1-{\mathrm{\α}}_F)M_F^{t-1}\left(n\right)+{\mathrm{\α}}_F{f}^t\left(n\right);$ Wherein,

Concrete, α _frefer to learning rate, f ^t(n) represent whether pixel n changes, more approach 0, this pixel is that the probability of moving region is larger.

M _t, M _sand M _fbe all the numerical value of [0,1], be used for respectively temporal information, spatial information and the frequency information of estimated image.Therefore, pixel n is the probability of moving region at moment t can be expressed as follows:

$P_{\mathrm{FG}}^t\left(n\right)=M_T^t\left(n\right)\×M_S^t\left(n\right)\×(1-M_F^t\left(n\right)).$

According to the moving region testing result of the moving region probability graph of present frame and former frame, the sampling mask of moving region by least one sampling policy to obtain the mask M that samples ^t, and more preferably combine three kinds of sampling policies, that is:

1. Random Discrete sampling policy: in the region-wide image of each frame, according to probability ρ (ρ=0.05～0.1) random acquisition pixel.Wherein, if the pixel that stochastic sampling is adopted in previous frame belongs to moving region after testing afterwards, in present frame, will this pixel of chosen in advance be the pixel of sampling, remaining pixel is chosen at random according to remaining stochastic sampling quota,

2. spatial spread sampling policy: because moving region has a certain size, and be communicated with, if some pixels belong to the probability of moving region very large, near the pixel this pixel is probably also moving region so.Each sampling pixel points of therefore, sampling for Random Discrete check the value of this pixel in the probability graph of moving region, expand a square area N (i) according to probable value centered by this sampling pixel points, its size is ξ ^t(i) * ξ ^t(i), all pixels in square area are sampled, foursquare size is directly proportional to the value of moving region probability graph.

3. pixel sampling strategy suddenlys change: the sampling that this strategy carries out while entering into scene suddenly from scene for foreign body.Now to belong to the probability of moving region very little the emergent position of this object, but in stochastic sampling process, if pixel n is detected as moving region, and the condition that meets following computing formula, is just judged as sudden change pixel by this pixel centered by this pixel, expand a square area simultaneously.

Wherein ${\mathrm{\θ}}_{\mathrm{th}}^{t-1}=\max (P_{\mathrm{FG}}^{t-1}/{\mathrm{\ω}}_t).$

Therefore, sudden change pixel sampling is the square that N (i) is ω for width.

In conjunction with above-mentioned three kinds of sampling policies (three kinds of sampling policies being carried out and computing), can obtain the sampling mask M of final moving region ^t, be shown below:

In sum, in the present embodiment, moving region sampling algorithm can detect resulting moving region testing result according to former frame input picture, in conjunction with the temporal information, spatial information and the frequency information that comprise in the testing result of above-mentioned moving region, obtain the moving region probability graph of current frame image; Then adopt Random Discrete strategy, spatial spread strategy and these three kinds of sampling policies of sudden change pixel strategy, calculate the sampling mask M of a moving region ^t.

Next, execution step S3: by ViBe algorithm, described sampling mask is carried out to profile and preferentially detect, pedestrian region obtains moving.

In the present embodiment, at the sampling mask M of moving region ^tbasis on, by adjusting detection order, can improve the detection speed of ViBe algorithm to motion pedestrian region.Because motion pedestrian region is a connected region, if the sampling mask M of moving region ^tperipheral profile by correct motion pedestrian region, the sampling mask M of current moving region of detecting ^tinside must belong to motion pedestrian region, the therefore sampling mask M to moving region again ^tinside carry out ViBe detection, it directly can be judged as motion pedestrian region.In the present embodiment, by changing the pixel processing sequence of target detection, from the sampling mask M of moving region ^tperipheral profile to its inner detection, be that profile preferentially detects gradually, reduce the number that needs the pixel that detects, reach the effect of acceleration.

First the preferential detection algorithm of profile needs to define a parameter order of contact: in four neighborhood territory pixel points (being neighbours territories) up and down of each pixel, if there be N pixel to belong to the sampling mask M of moving region ^tor moving region, the order of contact of this pixel is N.

Shown in ginseng Fig. 4, for the sampling mask M of moving region ^teach interior pixel, calculates its order of contact.Wherein, the sampling mask M of region 41Wei moving region ^t, the sampling mask M of moving region ^tin the order of contact of each pixel of numeral, the sampling mask part of 42Wei Fei moving region, region, region 42 is directly judged as background area.

Concrete, the concrete disposal route of described " profile preferentially detects " is as follows:

Calculate described sampling mask M ^tthe order of contact of each interior pixel, and according to the detection order of the large pixel of little pixel the second priority processing order of contact of the first priority processing order of contact, little by little from sampling mask M ^tperipheral profile to its inside, carry out sequence detection.

In the testing process of motion pedestrian region, first need to process the sampling mask M of moving region ^tcontour area, then process the sampling mask M of moving region ^tinterior zone.

In Fig. 4, first process the pixel of order of contact minimum, the pixel that the order of contact that upper right corner marks with dotted line is 0.Suppose that this pixel detects and is judged as background through ViBe algorithm, this pixel is labeled as to background, the order of contact of contiguous four pixels (if having) is subtracted to 1 simultaneously.Then process the pixel of remaining order of contact minimum, it is the pixel that order of contact that in Fig. 4, the lower right corner marks with dotted line is 1, through ViBe algorithm, detect and be judged as background, equally this pixel is labeled as to background, the order of contact of four pixels of this pixel vicinity is subtracted to 1 simultaneously, obtain Fig. 5 a.

Shown in ginseng Fig. 5 a, if there are a plurality of onesize minimal-contact degree pixels, as there are a plurality of orders of contact in Fig. 5 a, it is 2 pixel, choose at random one of them, the pixel that the order of contact that the upper right corner marks with dotted line as chosen is 2, detects and is judged as motion pedestrian region through ViBe algorithm, this pixel is labeled as to motion pedestrian region, and the order of contact of four contiguous pixels is constant, motion pedestrian region is as the dash area of corresponding pixel in Fig. 5 b.

Shown in ginseng Fig. 5 b, continue to process the pixel of next order of contact minimum, the pixel that the order of contact marking with dotted line as Fig. 5 b upper right corner is 2, through ViBe algorithm, detect and be judged as background, the order of contact of contiguous four pixels (if having) subtracts 1, and the order of contact that Fig. 5 c upper right corner marks with dotted line is the pixel that 2 pixels and dash area right side order of contact are 3.

Shown in ginseng Fig. 6, according to above principle processed pixels point one by one in order, when the order of contact of remaining pixel is 4, represent that this pixel is the sampling mask M of moving region ^tinside, the sampling mask M of this moving region ^tperipheral profile successfully detected as motion pedestrian region, interior pixels point, without judgement, is directly all judged as motion pedestrian region, obtains Fig. 7, has reduced a pixel quantity that needs detection, reaches the object of acceleration.

In sum, to motion pedestrian region, all extract complete.

Finally, execution step S4: the number of utilizing linear regression analysis method statistic motion pedestrian region.

In the present embodiment, motion pedestrian region can not be all single pedestrian, is likely two even a plurality of people of people yet, need to calculate the number in motion for line people region by analysis meter is carried out in motion pedestrian region.

In the present embodiment, because motion pedestrian number is directly proportional with pixel and the edge number in motion pedestrian region, therefore can be according to the number of pixel in this pedestrian region of moving and two features of edge number as the move foundation of pedestrian region one skilled in the art's number of judgement.

The relation that is directly proportional to its feature of number based on motion pedestrian, can utilize following linear function to come Describing Motion pedestrian's number and the relation between its feature, concrete as shown in formula (1):

AX+BY+C＝Z (1)

Wherein, the X pixel number in pedestrian region that represents to move, Y represent the to move edge number in pedestrian region, Z represent the to move number in pedestrian region, A, B, C are respectively the coefficients of linear function.By gathering a large amount of motion pedestrians region as sample, use the method for linear regression to carry out matching to sample, just can calculate coefficient A, B and the C of linear regression function.

Suppose and have n sample { (X ₁, Y ₁, Z ₁), (X ₂, Y ₂, Z ₂) ..., (X _n, Y _n, Z _n), wherein, Z _irepresent number measured value in motion pedestrian region, Z _jexpression is according to the linear function of formula (1) the number calculated value in pedestrian region that obtains moving, concrete as shown in formula (2):

Z _j＝AX _i+BY _i+C (2)

Utilize least square method to carry out sample data to carry out matching, order represent number measured value Z in motion pedestrian region _iwith number calculated value Z in motion pedestrian region _jbetween error, error can represent by both quadratic sums of difference, concrete as shown in formula (3):

Formula (3) substitution formula (2) can be obtained to formula (4):

In order to make error minimum, available error coefficient A, B, C are asked respectively to local derviation, and make these three partial derivatives all equal 0, concrete as shown in formula (5):

By solution formula (5), can obtain coefficient A, B, the C of linear regression function.

In the present embodiment, computer system is utilized the method for linear regression analysis, and the data of matching sample set calculate the linear equation that represents motion pedestrian region number and motion pedestrian's area pixel point number and edge number relation.Fit procedure can adopt online fitting mode or offline simulation mode.

More approaching and the reality scene of sample set that online fitting mode obtains, can accurately on-the-spot influence factor be reflected by sample set in time, be conducive to reduce the error of regretional analysis, but online fitting mode can be brought certain trouble in practice, after each deployment, all to judge great amount of samples collection by manual type.

Offline simulation mode has been simplified the difficulty of disposing link, by the great amount of samples under different scenes, carries out matching, and regretional analysis goes out one group of coefficient that is applicable to most of scene, can avoid over-fitting, and range of application is wider, and actual operation speed is faster.Therefore in the present embodiment, be preferably by offline simulation mode and obtain matching sample.

In addition, be to be understood that, although this instructions is described according to embodiment, but not each embodiment only comprises an independently technical scheme, this narrating mode of instructions is only for clarity sake, those skilled in the art should make instructions as a whole, and the technical scheme in each embodiment also can, through appropriately combined, form other embodiments that it will be appreciated by those skilled in the art that.

Claims (5)

1. the people flow rate statistical method based on ViBe, is characterized in that, the method comprises the following steps:

S1, obtain guarded region video streaming image as input picture;

S2, input picture is carried out to moving region sampling, obtain the sampling mask M of moving region ^t;

S3, by ViBe algorithm, described sampling mask is carried out to profile and preferentially detect, pedestrian region obtains moving;

S4, utilize linear regression analysis method statistic by the flow of the people in described guarded region.

2. people flow rate statistical method according to claim 1, is characterized in that, described step S1 is specially: the video streaming image that obtains guarded region by video camera is as input picture, described guarded region be positioned at video camera under.

3. people flow rate statistical method according to claim 1, is characterized in that, described step S2 is specially:

According to former frame input picture being detected to temporal information, spatial information and the frequency information comprising in the result of resulting motion pedestrian region, calculate moving region probability graph, and by least one sampling policy to obtain the mask M that samples ^t.

4. people flow rate statistical method according to claim 3, is characterized in that, described sampling policy comprises: Random Discrete sampling policy, spatial spread sampling policy and sudden change pixel sampling strategy.

5. people flow rate statistical method according to claim 1, is characterized in that, " profile preferentially detects " in described step S3 is specially:

Calculate the order of contact of each pixel in described sampling mask, and according to the detection order of the large pixel of little pixel the second priority processing order of contact of the first priority processing order of contact, little by little from the peripheral profile of sampling mask, to its inside, carry out sequence detection.