CN101908236A - Public traffice passenger flow statistical method - Google Patents

Abstract

The invention discloses a public traffice passenger flow statistical method based on a multi-threshold region growing and optical flow method. Aiming at the special environment of a bus, a head part is used as a counting target, two thresholds are dynamically adopted, the head part is separated from the background through two region growing algorithms and characteristic extraction, a growing threshold is determined through a variance of an image, and a piecewise function relation is formed between the growing threshold and the variance of the image. For other special conditions such as hat-wearing and the like, a moving human body is used as a counting target by using the optical flow method. In the image, a getting-on counting line and a getting-off counting line are manually designated, when the tracked target passes through the counting lines, the distance moved in the directions of the counting lines from coordinates detected for the first time and coordinates at the moment is calculated, if the distance is more than the pre-determined threshold, the counting is generated; or else the counting is not carried out, the influence of a disturbing object to the counting is effectively reduced.

Description

Public traffice passenger flow statistical method

Technical field

The present invention relates to the moving object detection and tracking technology, particularly a kind of public traffice passenger flow statistical method based on many threshold region growths and optical flow method.

Background technology

For moving object detection, classic method mainly contains background subtraction method, frame difference method and optical flow method.It is little or change environment comparatively slowly that the background method is mainly used in change of background, before target enters the video camera scope, extracts Background, after target enters, with present image subtracting background figure, obtains moving target.Context update is the emphasis of background method, mainly contains single gaussian sum mixed Gaussian background model method at present.Because each platform environment is different, and it is very big influenced by light and blocking on the bus, so be unsuitable for adopting the background subtraction method.Though frame difference method is influenced not quite by light, in crowded, a plurality of passengers are sticked together easily, and how independent passenger being split is a difficult problem.Even consider that passenger's head also seldom is close together when crowding, so be easy to realize as the counting target with head.Have many scholars to adopt edge extracting and Hough to change at present and seek the number of people, this method calculated amount is big, and it is also bigger to take storage space.

Optical flow method has its advantage in detecting moving target: can not know to detect moving target under the situation of any scene information.Optical flow computation method based on gradient has obtained using widely at present.The problem that adopts optical flow method to carry out the moving object detection is that mainly optical flow method calculating is consuming time, real-time and practicality are all relatively poor, pure to adopt optical flow method to detect moving target not-so-practical, detects moving target but optical flow computation method and additive method can be combined.

Summary of the invention

For solving the problem that prior art exists, the present invention proposes a kind of public traffice passenger flow statistical method based on many threshold region growths and optical flow method, catch the passenger getting on/off image in real time by the camera that is installed on the bus, adopt this method to calculate the number of getting on or off the bus.

For solving the problems of the technologies described above, technical scheme of the present invention is:

A kind of public traffice passenger flow statistical method may further comprise the steps:

At first obtain two threshold values by twice big Tianjin threshold method, filtering can not be the zone of head to big threshold value as the value of roughly selecting, and less threshold value is as head average gray threshold value;

The variance of computed image is determined two growths of size threshold value G then _mAnd G _n

Adopt for the first time G _mImage is carried out region growing, do not comprise the zone of head by area and average gray filtering;

To the zone that keeps, judge at first whether it comprises detected any one head center-of-mass coordinate in the previous frame, if comprise, be 1 then with step-length, the growth threshold value is from G _nTo G _mEach value region growing is carried out in this zone, and seek the zone the most similar to this head by normalization coupling;

To not comprising the zone of any one head barycenter in the previous frame, whether at first judge its circularity greater than the circularity threshold value, if greater than would directly think head zone and preserve;

Otherwise, is that the circle of R is made opening operation to it to this zone with radius, and R represents the average head size, to its average gray threshold value of region decision of obtaining whether less than head average gray threshold value, and circularity is greater than the circularity threshold value, if meet then think head and preserve.Otherwise get G _nCarry out region growing once more, find the head zone of separating by feature extraction;

Preserve the initial horizontal coordinate of each head barycenter and the coordinate in present frame with an array;

Calculate the optical flow field of this image, obtain the moving region, each moving region and the detected head of region-growing method are done and computing, non-NULL is then deleted as a result, otherwise keeps;

Adopt the detected head of its algorithm to carry out the most contiguous coupling respectively with in the previous frame to detected head of region-growing method and the detected head of optical flow method, upgrade the position of each head;

Judge whether each head crosses counting line, if, calculate the distance that it is passed by on the counting line direction, if greater than threshold value then count, otherwise do not process.

After first time region growing, judge that whether every zone comprises the barycenter of certain head that the previous frame image detection arrives, if comprise, by a series of growth threshold value and normalization form fit, extracts the head zone the most similar with the previous frame number of people.

For special case such as being branded as, adopt optical flow method to detect movement human; Optical flow method and region growing algorithm synthesis use, respectively counting.

Artificial two counting lines of getting on or off the bus of delimiting when tracking target is crossed counting line, judge that distance that target passes by whether greater than pre-determined threshold value, counts greater than then producing, otherwise do not count on the counting line direction.

Compared with prior art, beneficial effect of the present invention can be:

Bus passenger flow statistics is significant, and the crowded degree by writing down each website and the ridership of each bar circuit can be the arrangement of bus station, and the scheduling of bus and the adjustment of public bus network provide abundant foundation and information.

The bus passenger flow statistic algorithm that the present invention is based on many threshold region growths and optical flow method is based on the region growing algorithm, and optical flow method is auxilliary, by the normalization form fit, improves and extracts precision.Whether to cross counting line and whether two condition judgment of displacement produce counting, effectively reduce the influence of disturbing counting.

Embodiment

The inventive method adopts the region growing algorithm that head is split from background, by feature extraction head detection is come out as the counting target again.The region growing algorithm adopts two threshold values, threshold value choose component of variance section funtcional relationship with image.This piecewise function is determined after the statistical study great amount of images.

After detecting a head, calculate the coordinate of its barycenter.This algorithm supposes that the center-of-mass coordinate of each its former frame of head is contained in its in the zone of next frame (frame per second of camera is set up substantially) more than 25 frames.So after finishing the region growing algorithm first time, judge at first whether every zone comprises the barycenter of detected certain head in the former frame, if comprise, then choose many these zones of a plurality of growth threshold values and carry out region growing, extract the head zone the most similar to the previous frame head by normalized form fit method.

Wait other special circumstances for being branded as, this algorithm adopts optical flow method as auxiliary counting, and after calculating the light stream image, passing threshold is chosen the bigger zone of light stream.Undertaken and computing by light stream zone and the detected head zone of region growing algorithm that each is extracted, avoid repeat count.

Its information is preserved with a three-dimensional array in the detected moving region of each head and optical flow method (following be called for short with moving target), suppose that counting line is a benchmark with the x coordinate, then first element of this array is exactly the barycenter horizontal ordinate of this moving target when detecting for the first time, and second and third element is the abscissa value and the ordinate value of the barycenter of moving target in present image.When target is crossed counting line, judge the distance that target is passed by on the counting line direction, if greater than threshold value then produce counting, otherwise do not count.

The flow process of this method comprises following step:

1. camera is caught the first frame gray level image, as the initial pictures of optical flow method.

2. calculate big Tianjin threshold value T1 of this image, to ask for its big Tianjin threshold value T2 once more less than the pixel of T1, as the average gray threshold value of head.

3. ask for pixel gray variance, obtain corresponding two growth threshold value G according to the piecewise function of variance and growth threshold value less than T1 _mAnd G _n

4. with G _mFor the growth threshold value image is carried out region growing, it is suitable to choose area, average gray less than the zone of T1 as candidate's head zone.

5. following operation is done in each candidate region:

Whether at first judge its circularity greater than the circularity threshold value, if greater than the circularity threshold value then think that this zone is a head, should preserve in the zone.Otherwise with a radius is that the circle of R (being as the criterion with the head mean size that occurs in the image) carries out opening operation to it, to the zone that obtains, selects average gray less than T2, and circularity is preserved as head zone greater than the zone of threshold value.If still can't obtain a head zone, then the region growing second time is carried out in this zone, the growth threshold value is selected G _n, be that the circle of R carries out opening operation to it with radius to the zone that obtains, select average gray less than T2 then, circularity is preserved as head zone greater than the zone of threshold value.

6. preserve the initial horizontal ordinate (with the x direction as counting direction) of each head zone and the center-of-mass coordinate in present frame with an array, promptly each head zone is preserved its information with three data.

7. camera is caught second two field picture, 2,3,4 it is handled set by step, to each zone that obtains, judges whether it comprises the barycenter of certain head in the previous frame, if comprise then use G _nTo G _m, step-length is that a series of growth threshold values of 1 are carried out region growing to this zone, adopts the normalization matching algorithm, seeks the zone the most similar to this head as position and the shape of this head in present frame.

8. to not comprising the zone of any head barycenter of previous frame, 5 operate set by step.

9. ask the difference image of second two field picture and first two field picture, obtain the moving region,, calculate the minimum boundary rectangle REC that comprises the moving region by opening operation filtering small movements zone.

With first two field picture as the optical flow field in the reference image calculation REC zone, select light stream greater than the pixel of threshold value as the motor point, and the calculating connected domain, select the circle of correct radial that the connected domain that obtains is carried out opening operation, purpose is the small connection that disconnects between two moving objects.

11. each optical flow field connected domain is carried out and computing with the head zone that adopts the region growing algorithm to obtain respectively.If be not empty, then should the zone have preserved and be used as head, otherwise deleted.

12. preserve the detected head position information of optical flow method with an array equally.Comprise the initial position horizontal ordinate and in the position of present frame.Replace first two field picture conduct reference picture of optical flow method next time with second two field picture.

13. head and the detected head of optical flow method that region-growing method is extracted carry out the most contiguous coupling with the head that each algorithm of previous frame obtains respectively.Upgrade the position (initial horizontal coordinate figure constant) of each head in present frame.

Whether cross counting line 14. judge each head, if, calculate the distance that it is passed by on the counting line direction, if greater than threshold value, then count, and with this head deletion.Otherwise do not handle.

15. camera continues photographic images, repeating step 2-13.

Below by concrete algorithm and formula the present invention is described in detail.

1. before the region growing algorithm image is analyzed carrying out, calculated the gray threshold of head, adopt big Tianjin threshold method.

The first step calculates the minimum and maximum gray-scale value of image, calculating mean value:

$\stackrel{\‾}{g}=\frac{g_{\max }+g_{\min }}{2}$

Second the step with

As threshold value, image is divided into two parts, calculate the average gray value of two parts image respectively

With

Obtain

${\stackrel{\‾}{g}}^{\′}=\frac{{\stackrel{\‾}{g}}_1+{\stackrel{\‾}{g}}_2}{2}---\left(1\right)$

If

$|{\stackrel{\&OverBar;}{g}}^{\&prime;}-\stackrel{\&OverBar;}{g}|<0.05---\left(2\right)$

Then

Be the big Tianjin threshold value that obtains, otherwise order

$\stackrel{\&OverBar;}{g}={\stackrel{\&OverBar;}{g}}^{\&prime;}$

Continuation with

As threshold value, repeated for second step, set up up to (2) formula, obtain big Tianjin threshold value.

2. the gray variance S of computed image determines two region growing threshold value G by variance _mAnd G _n, only calculate the variance of gray scale here, because the head gray-scale value is generally less, so the information that the information pixel more higher than gray scale that gray scale provides than big Tianjin little pixel of threshold value provides is more accurate less than big Tianjin threshold pixels point.The funtcional relationship of growth threshold value and variance is as follows:

G _m＝7，G _n＝4，S＞60

G _m＝5，G _n＝3，40≤S≤60

G _m＝4，G _n＝2，S＜40

3. the realization of region growing algorithm.The region growing algorithm that this patent adopts is as follows: from the upper left corner of image, cover with the template of a 3*3 9 points the image upper left corner, the average gray g1 of zoning R1,3 pixels again move right template, the average gray g2 of calculation template overlay area R2, if

| g1-g2|＜T (T is a threshold value)

Then R1 and R2 territory are merged, be designated as region R 1, calculate its average gray value g1 ', order

g1＝g1′

Otherwise nonjoinder continues the rightmost of movable platen to image, adopts identical method to handle.Then template is moved back into the Far Left of image, moves down 3 pixels simultaneously, the average gray g3 of zoning R3, if

|g1-g3|＜T

Then R1 and R3 are merged, otherwise nonjoinder.So circulation covers complete image up to template.

4. this method adopts normalized form fit to calculate the similarity in two zones, its computation process is as follows: the Hamming distance of at first calculating two zones, promptly calculate the number of two regional different pixels, suppose two region R egion1 and Region2, its Hamming distance is expressed as follows:

Distance＝|Norm(Regions1)&～Regions2|+

|Regions2?&～Norm(Regions1)|

Wherein Norm (Region1) expression moves to Region1 and makes itself and Region2 have the position of identical barycenter.The expression calculating pixel number that takes absolute value ， ﹠amp; Expression is carried out and operation ,～expression negate.Obtain after the Hamming distance in two zones, calculate the similarity in two zones, computing formula is:

Similarity＝1-Distance/(|Norm(Regions1)|+|Regions2|)

5. this patent adopts the classical optical flow algorithm of improved Horn-Schunck, and this method is based on image gradient and the constant thought of grey scale pixel value summation.

For image f (x, y) (i, gray-scale value j) they are f (i in the t moment point, j t), supposes that light stream is u (i at the axial component of x, j), the axial component of y be v (i, j), after then passing through the Δ t time, (i j) moves to (i+u* Δ t to point, j+v* Δ t) locate, this gray-scale value is f (i+u* Δ t, j+v* Δ t, t+ Δ t), levels off in 0, can think at Δ t

f(i，j，t)＝f(i+u*Δt，j+v*Δt，t+Δt)

Taylor expansion is carried out in this examination, and the higher order term of dividing out can obtain:

$\frac{\&PartialD;f}{\&PartialD;x}\frac{\mathrm{dx}}{\mathrm{dt}}+\frac{\&PartialD;f}{\&PartialD;y}\frac{\mathrm{dy}}{\mathrm{dt}}+\frac{\&PartialD;f}{\&PartialD;t}=0---\left(3\right)$

Order

$\frac{\&PartialD;f}{\&PartialD;x}=f_x,$ $\frac{\&PartialD;f}{\&PartialD;y}=f_y,$ $\frac{\&PartialD;f}{\&PartialD;t}=f_t$

Order

$\frac{\mathrm{dx}}{\mathrm{dt}}=u,$ $\frac{\mathrm{dy}}{\mathrm{dt}}=v$

Then (3) formula can be rewritten as:

f _xu+f _yv+f _t＝0 (4)

Wherein

$f_x=\frac{1}{4}\{f(i+1,j,t)-f(i,j,t)+f(i+1,j+1,t)-f(i,j,t)$

$+f(i+1,j,t+1)-f(i,j,t+1)+f(i+1,j+1,t+1)-f(i,j+1,t+1)$

$f_y=\frac{1}{4}\{f(i,j+1,t)-f(i,j,t)+f(i+1,j+1,t)-f(i+1,j,t)$

$+f(i,j+1,t+1)-f(i,j,t+1)+f(i+1,j+1,t+1)-f(i+1,j,t+1)$

$f_t=\frac{1}{4}\{f(i,j,t+1)-f(i,j,t)+f(i+1,j+1,t+1)-f(i+1,j+1,t)$

$+f(i,j+1,t+1)-f(i,j+1,t)+f(i+1,j,t+1)-f(i+1,j,t)$

U, v are light stream values to be asked, and an equation can't be obtained, and still the optical flow field that produces according to an object is uniformly level and smooth in its zone, and Horn-Schunck has proposed smooth restrictive condition:

$E={\&Integral;\&Integral;(f_{x}u+f_{y}v+f_t)}^2\mathrm{dxdy}+\mathrm{\&alpha;}\&Integral;\&Integral;\{{\left(\frac{\&PartialD;u}{\&PartialD;x}\right)}^2+{\left(\frac{\&PartialD;u}{\&PartialD;y}\right)}^2+{\left(\frac{\&PartialD;v}{\&PartialD;x}\right)}^2+{\left(\frac{\&PartialD;v}{\&PartialD;y}\right)}^2\}\mathrm{dxdy}---\left(4\right)$

Wherein α is the canonical coefficient, has represented the ratio of overall level and smooth property, and the size of α generally depends on the precision that image gradient calculates.U that calculates and v make the E minimum.So E is asked for the partial derivative of u and v, and making it respectively is 0, the iterative formula that obtains u and v is:

${u_{\mathrm{ij}}}^{n+1}={{\stackrel{\&OverBar;}{u}}_{\mathrm{ij}}}^n-\frac{f_x{{\stackrel{\&OverBar;}{u}}_{\mathrm{ij}}}^n+f_y{{\stackrel{\&OverBar;}{v}}_{\mathrm{ij}}}^n+f_t}{\frac{1}{\mathrm{\&alpha;}}+{f_x}^2+{f_y}^2}{f}_x$

${v_{\mathrm{ij}}}^{n+1}={{\stackrel{\&OverBar;}{v}}_{\mathrm{ij}}}^n-\frac{f_x{{\stackrel{\&OverBar;}{u}}_{\mathrm{ij}}}^n+f_y{{\stackrel{\&OverBar;}{v}}_{\mathrm{ij}}}^n+f_t}{\frac{1}{\mathrm{\&alpha;}}+{f_x}^2+{f_y}^2}{f}_y$

Wherein

With

Represent point (i, j) the average level light stream value of adjacent four pixels and vertical light flow valuve size respectively.Before iteration begins, promptly during n=0, make the light stream initial value be [0,0], continuous iteration then, when

$|{v_{\mathrm{ij}}}^{n+1}-{{\stackrel{\&OverBar;}{v}}_{\mathrm{ij}}}^n|<T$

And

$|{u_{\mathrm{ij}}}^{n+1}-{{\stackrel{\&OverBar;}{u}}_{\mathrm{ij}}}^n|<T$

T is prior preset threshold, generally gets between the 0.01-0.1, if can not allow iteration stopping, again by judging that number of iterations allows its termination, stops after falling generation 100 times.

The precision that optical flow field calculates depends on the precision that image gradient calculates to a great extent, so before carrying out optical flow computation, generally to just filtering and noise reduction processing of image.Such as passing through gaussian filtering.The limitation of optical flow method is that the computing time of its requirement is longer, so in handling in real time, how to improve its computing velocity, the not too big again optical flow computation precision that influences is the primary problem that solves.This method is accelerated the computing velocity of optical flow method by two improved methods.

(1) image is carried out the metric space conversion, reduce calculative pixel.The picture that shooting obtains supposes that size for M*N, becomes M/2*N/2 by change of scale with graphical rule, according to the optical flow method calculating principle, in an object area, its light stream size basically identical is not so this method can produce considerable influence to the optical flow computation precision.The optical flow field that obtains after the conversion returns M*N by neighbor interpolation method change of scale.

(2) select to carry out the zone of optical flow computation by frame difference method.What optical flow field calculated is the speed in the zone of motion, so static object and background are not needed to calculate.Frame difference method can detect the profile of moving object, thereby detects the moving region by filling scheduling algorithm.For because illumination or moving region that small items produced, by opening operation with its filtering.

The above is preferred embodiment of the present invention only, is not to be used for limiting practical range of the present invention.Have in the technical field under any and know the knowledgeable usually, without departing from the spirit and scope of the present invention, when can being used for a variety of modifications and variations, so protection scope of the present invention should be looked claims institute confining spectrum and is as the criterion.

Claims (4)

1. public traffice passenger flow statistical method is characterized in that may further comprise the steps:

At first obtain two threshold values by twice big Tianjin threshold method, filtering can not be the zone of head to big threshold value as the value of roughly selecting, and less threshold value is as head average gray threshold value;

The variance of computed image is determined two growths of size threshold value G then _mAnd G _n

Adopt for the first time G _mImage is carried out region growing, do not comprise the zone of head by area and average gray filtering;

To the zone that keeps, judge at first whether it comprises detected any one head center-of-mass coordinate in the previous frame, if comprise, be 1 then with step-length, the growth threshold value is from G _nTo G _mEach value region growing is carried out in this zone, and seek the zone the most similar to this head by normalization coupling;

To not comprising the zone of any one head barycenter in the previous frame, whether at first judge its circularity greater than the circularity threshold value, if greater than would directly think head zone and preserve;

Otherwise, is that the circle of R is made opening operation to it to this zone with radius, and R represents the average head size, to its average gray threshold value of region decision of obtaining whether less than head average gray threshold value, and circularity is greater than the circularity threshold value, if meet then think head and preserve.Otherwise get G _nCarry out region growing once more, find the head zone of separating by feature extraction;

Preserve the initial horizontal coordinate of each head barycenter and the coordinate in present frame with an array;

Calculate the optical flow field of this image, obtain the moving region, each moving region and the detected head of region-growing method are done and computing, non-NULL is then deleted as a result, otherwise keeps;

Adopt the detected head of its algorithm to carry out the most contiguous coupling respectively with in the previous frame to detected head of region-growing method and the detected head of optical flow method, upgrade the position of each head;

Judge whether each head crosses counting line, if, calculate the distance that it is passed by on the counting line direction, if greater than threshold value then count, otherwise do not process.

2. public traffice passenger flow statistical method according to claim 1, it is characterized in that: after first time region growing, judge whether every zone comprises the barycenter of certain head that the previous frame image detection arrives, if comprise, by a series of growth threshold value and normalization form fit, extract the head zone the most similar with the previous frame number of people.

3. public traffice passenger flow statistical method according to claim 1 is characterized in that: for special case such as being branded as, adopt optical flow method to detect movement human; Optical flow method and region growing algorithm synthesis use, respectively counting.

4. public traffice passenger flow statistical method according to claim 1, it is characterized in that: artificial two counting lines of getting on or off the bus of delimiting, when tracking target is crossed counting line, judge that whether distance that target passes by is greater than pre-determined threshold value on the counting line direction, count greater than then producing, otherwise do not count.