CN100543765C - Method for monitoring instruction based on computer vision - Google Patents

Abstract

A kind of method for monitoring instruction based on computer vision is used for the intelligent video monitoring field.This method is gathered the video image of some frames in advance by rig camera, set up model of place based on nonparametric statistics with this.Video image for new detects the foreground point with model of place, and obtains moving target by further processing of morphologic filtering.Then moving target is extracted the weighted histogram feature, and adopt histogram matching pursuit movement target, obtain the movement locus of target.Judge whether target trajectory enters the warning region of setting, thereby whether decision gives the alarm.This invasion inspection method for supervising adopts computer vision technique, the video information of utilizing the Computer Processing rig camera to obtain, and realization detects automatically and reports to the police, and equipment is simple, easy realizes having vast market prospect and using value.

Description

Method for monitoring instruction based on computer vision

Technical field

The present invention relates to a kind of method for monitoring instruction, be specifically related to a kind of method for monitoring instruction based on computer vision technique.

Background technology

The monitoring instruction system refers to utilizes camera to judge the system that concerns between target and the fixed area, as judging swarming into of warning region, and being strayed into etc. of dangerous area.At present, this type systematic is mainly finished by 24 hours visual inspection monitoring screens of security personnel.This is a very hard work, and owing to reasons such as fatigues, people's notice is easy to disperse.So handle the video information that camera returns by computing machine, monitor highly significant to replace manpower.

Summary of the invention

The object of the present invention is to provide a kind of method for monitoring instruction, to improve the deficiency that adopts manpower to monitor in the existing method based on computer vision.

Method for monitoring instruction based on computer vision of the present invention may further comprise the steps: obtain the successive frame video image of some, the scene information when these images are mainly described no moving target and existed; According to the pixel value of the multiple image that obtains, set up background model with nonparametric statistical method; For the image that newly obtains, utilize the background model that obtains, the thresholding comparison that each pixel is under the jurisdiction of the probability of background and sets in advance in the estimated image judges with this whether this point is the foreground point; The morphologic filtering processing is carried out in the foreground point obtain the prospect block, whether the area size decision according to the prospect block is moving target afterwards; Adopt the weighted histogram feature to modeling target, and adopt histogram matching to follow the tracks of above-mentioned moving target; Utilize pursuit path, judge whether moving target enters the sensitizing range of prior setting, report to the police if enter then

Below each step of this invention is specified:

(1) background modeling

At first collect the background image of some.Then each pixel in the background image is carried out statistical modeling.The background model that obtains is distributed by the background color of each location of pixels in the image and describes, and the nonparametric statistical method of these distributions is as follows:

$p\left(x_i\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{d}{\mathrm{\Π}}}\frac{1}{\sqrt{{2\mathrm{\π\σ}}_j^2}}\exp (-\frac{{(x_{t_j}-x_{t_j})}^2}{{2\mathrm{\σ}}_j^2})$

Wherein, x _tBe the d dimension color characteristic observed reading of certain location of pixels of moment t, { x _i} _{I=1 ..., N}This pixel position that expression is collected in advance one group of d dimension background color observed reading, exp (x) represents exponential function, σ _jBe each dimensional feature corresponding variance, then p (x _t) be t this pixel value x constantly _tThe probable value that belongs to background.

(2) moving object detection

Obtain after the background statistical model, carve t at a time, the detection of foreground target can be by comparing probable value p (x in the image _t) and the size of certain threshold values decide.Pixel value x for certain position in the image _tIf, probable value p (x _t) less than the threshold values of certain setting, then current corresponding pixel is the foreground point, otherwise corresponding pixel is a background dot.In order to increase robustness, need further carry out mathematical morphology filter for the detected foreground point of background model and handle.The step that these mathematical morphology filters are handled comprises medium filtering, morphological erosion and morphology expansive working.For the foreground area that obtains after handling, choose area wherein greater than given threshold value as moving target.

(3) modeling target

After obtaining motion target area, this method adopts weighting color histogram model to come modeling target.For certain height h, width w, center position is y _cThe foreground moving target area, this regional weighting color histogram model is

Wherein u is that color space quantizes back corresponding quantitative sequence number, and scope is 1 ..., m.Each

Computing method as follows:

${\hat{q}}_u=C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y_i-y_c}{\sqrt{h^2+w^2}}\left|\right|}^2\right)\mathrm{\δ}[b\left(y_i\right)-u]$

Wherein, y ₁Represent certain pixel coordinate of this motion target area, the number of pixels of whole target area is n, b (y _i) function obtains the quantification sequence number of this location of pixels pixel value, δ (x) function shows the Kronecker function, is defined as:

C is a normalization coefficient, makes

Satisfy ${\mathrm{\Σ}}_{u=1}^m{\hat{q}}_u=1$ Requirement.Function k (x) is the Epanechnikov function, and form is

D wherein, c _dBe constant

(4) motion tracking and invasion are judged

Here adopt the method for histogram feature coupling to come the pursuit movement target.At t constantly, the weighted histogram model of certain moving target A is

At t+1 constantly, in order to determine the reposition of target A.At first in the image of t+1, find, calculate their weighting color model, then calculate apart from the nearer several moving targets of target A original position Similarity with these targets Then the moving target of maximum similarity correspondence and target A are complementary, thereby obtain the reposition of target A.

${\hat{q}}_A^{t+1}=\arg \underset{{\hat{q}}_i^{t+1}}{\max }s({\hat{q}}_A^t,{\hat{q}}_i^{t+1})=\arg \underset{{\hat{q}}_t^{t+1}}{\max }\left({\mathrm{\Σ}}_{n=1}^m\sqrt{{\hat{q}}_A^u,{\hat{q}}_i^u}\right)$

Just can realize the tracking of moving target by the target of continuous coupling consecutive frame.When this target trajectory dropped in advance warning region by user's appointment, then system judged that invasion takes place and reports to the police.

Method for monitoring instruction based on computer vision of the present invention, adopt computer vision technique, utilize rig camera and computing machine to realize the monitoring instruction of sensitizing range, equipment needed thereby realize simply, easily and also cost lower, avoided manual supervisory hard work amount and fatigue problem.

Description of drawings

In order to understand the present invention better, the present invention will be described in detail below in conjunction with accompanying drawing, wherein:

Fig. 1 is the hardware system block diagram of method for monitoring instruction of the present invention;

Fig. 2 is the scene synoptic diagram of method for monitoring instruction of the present invention.

Fig. 3 is the realization flow figure of method for monitoring instruction of the present invention.

Embodiment

As shown in Figure 1, the hardware system of method for monitoring instruction of the present invention comprises: video camera is used to gather video image; Video Codec, the image that is used for video camera is gathered carries out encoding and decoding processing and Network Transmission; Computing machine is used to receive the video image from Video Codec, utilizes method for monitoring instruction of the present invention that the video image that is received is handled then.

Below in conjunction with Fig. 3 the method for monitoring instruction based on computer vision of the present invention is described.

At first set up initial back-ground model, promptly utilize the camera acquisition video image, collect the background image of N frame, and obtain each pixel in the image (i, j) d of position dimension pixel value.

Then, utilize N pixel value of each pixel position to set up the nonparametric statistics model, this model is expressed with a probability function:

$p\left(x_i\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{d}{\mathrm{\Π}}}\frac{1}{\sqrt{{2\mathrm{\π\σ}}_j^2}}\exp (-\frac{{(x_{t_j}-x_{t_j})}^2}{{2\mathrm{\σ}}_j^2})$ (formula 1)

Wherein, x _iBe the d dimension color characteristic observed reading of certain location of pixels of moment t, { x _i} _{T=1 ..., N}This pixel position that expression is collected in advance one group of d dimension background color observed reading, exp (x) represents exponential function, σ _jBe each dimensional feature corresponding variance, then p (x _t) be moment t pixel value x _tThe probable value that belongs to background.

(i after the background model of j) locating, utilizes given threshold values TH obtaining pixel ₁Detect the foreground point: at moment t, for pixel (i, pixel value x j) _tIf, p (x _t)≤TH ₁, then (i j) is the foreground point to pixel, denys the person, and (i j) is background dot to pixel.Because the major part point of generalized case hypograph all is a background dot, the sum term in the formula (1) had generally just surpassed threshold values before also not calculated the N item, and therefore most of point just can dispose in very short time.

After having detected the foreground point, need carry out mathematical morphology to foreground image and handle with filtering noise, filling cavity, specifically described processing comprises medium filtering, corrosion operation and expansive working.Carry out according to the following steps: foreground image is carried out 3 * 3 medium filtering, to remove isolated noise spot; The image that obtains after the filtering carried out 5 * 5 morphology expansive working; The image that obtains after expanding is carried out border tracking or marginal point connection, obtain the border of each connected region in the image, thereby the relevant information that obtains each connected region is gone out area less than certain threshold values TH then as size, area etc. ₂Or connected region in irregular shape; The pixel of inside, border is set to the foreground point, to fill the cavity that wherein may exist.Through after the aftertreatment, area is greater than given threshold values TH ₃The zone be chosen to be moving target.

Then adopt the weighting color histogram to come moving target is carried out modeling, specific as follows:

If { y _i} _{I=1,2 ..., n}Be certain height h, width w, center position is y _cThe foreground moving target area in the set of n pixel.B (y _i) represent the quantification sequence number of this pixel in the color quantizing space to quantize serial number range u=0,1 ..., m.Weighting color histogram model that then should the zone is

Each

Computing method as follows:

${\hat{q}}_u=C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y_i-y_c}{\sqrt{h^2+w^2}}\left|\right|}^2\right)\mathrm{\δ}[b\left(y_i\right)-u]$ (formula 2)

Wherein, y _iRepresent certain pixel coordinate of this motion target area, C is a normalization coefficient, makes

Satisfy ${\mathrm{\Σ}}_{u=1}^m{\hat{q}}_u=1$ Requirement.δ (x) function shows the Kronecker function, is defined as:

Function k (x) is the Epanechnikov function, and form is

D wherein, c _dBe constant.

Obtain after the color histogram model of moving target, adopt the method for histogram feature coupling to come the pursuit movement target.At t constantly, the weighted histogram model of certain moving target A is

At t+1 constantly, in order to determine the reposition of target A, at first in the image of t+1, find, and calculate their weighting color model apart from the nearer several moving targets of target A original position.Then respectively and and

Relatively, try to achieve similarity

Here adopt the Bhattacharyya coefficient to calculate similarity:

$s({\hat{q}}_A^t,{\hat{q}}_i^{t+1})={\mathrm{\Σ}}_{u=1}^m\sqrt{{\hat{q}}_A^u,{\hat{q}}_i^u}$ (formula 5)

Satisfying $s({\hat{q}}_A^t,{\hat{q}}_i^{t+1})\≥{\mathrm{TH}}_4$ All targets in, find the moving target with target A similarity maximum, this target location is exactly the reposition of target A.

${\hat{q}}_A^{t+1}=\arg \underset{{\hat{q}}_i^{t+1}}{\max }s({\hat{q}}_A^t,{\hat{q}}_i^{t+1})=\arg \underset{{\hat{q}}_t^{t+1}}{\max }\left({\mathrm{\Σ}}_{n=1}^m\sqrt{{\hat{q}}_A^u,{\hat{q}}_i^u}\right)$ (formula 6)

If target A can not find certain target and makes $s({\hat{q}}_A^t,{\hat{q}}_i^{t+1})\≥{\mathrm{TH}}_4,$ Then may target A left the target in the visual field and by other target occlusion, for this situation, A is kept with target, and in some two field pictures of back, continue to mate, if never mate, just think that this target left the target in the visual field,, then rebulid the pursuit path of A target in follow-up image if the coupling target is arranged.

Just can obtain the pursuit path of moving target by the target of continuous coupling consecutive frame.When this target trajectory dropped in advance warning region by user's appointment, then system judged that invasion takes place and reports to the police.

Claims (6)

1, a kind of method for monitoring instruction based on computer vision is characterized in that, may further comprise the steps: obtain the successive frame video image of some, the scene information when these images are mainly described no moving target and existed; According to the pixel value of the multiple image that obtains, set up background model with nonparametric statistical method; For the image that newly obtains, utilize the background model that obtains, the thresholding comparison that each pixel is under the jurisdiction of the probability of background and sets in advance in the estimated image judges with this whether this point is the foreground point; The morphologic filtering processing is carried out in the foreground point obtain the prospect block, whether the area size decision according to the prospect block is moving target afterwards; Adopt the weighted histogram feature to modeling target, and adopt histogram matching to follow the tracks of above-mentioned moving target; Utilize pursuit path, judge whether moving target enters the sensitizing range of prior setting, report to the police if enter then.

2, method for monitoring instruction according to claim 1 is characterized in that, background model is distributed by the background color of each pixel position in the image and describes, and these distributions obtain by the nonparametric statistics of multiframe background image:

$p\left(x_t\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{d}{\mathrm{\Π}}}\frac{1}{\sqrt{{2\mathrm{\π\σ}}_j^2}}\exp (-\frac{{(x_{t_j}-x_{t_j})}^2}{{2\mathrm{\σ}}_j^2})$

Wherein, x _tBe the d dimension color characteristic observed reading of certain location of pixels of moment t, { x _l} _{L=1 ..., N}The observed one group of d dimension in this pixel position color characteristic observed reading during the expression modeling, exp (x) represents exponential function, σ _jBe every dimensional feature corresponding variance, then p (x _t) be t pixel value x constantly _tThe probable value that belongs to background.

3, method for monitoring instruction according to claim 1 is characterized in that, this method is by comparing probable value p (x _t) and certain size of setting threshold values decide t this pixel value x constantly _tWhether belong to prospect: if probable value p is (x _t) less than the threshold values of certain setting, then current corresponding pixel is the foreground point, otherwise corresponding pixel is a background dot.

4, method for monitoring instruction according to claim 1, it is characterized in that, mathematical morphology filter is carried out in the detected foreground point of background model to be handled, the step that these mathematical morphology filters are handled comprises medium filtering, morphological erosion and morphology expansive working, for the foreground area after handling, the chosen area area greater than given threshold value as moving target.

5, method for monitoring instruction according to claim 1 is characterized in that, adopts weighting color histogram model to come modeling target; For height h, width w, center position is y _cThe foreground moving target area, this regional weighting color histogram model is

Wherein u is that color space quantizes back corresponding quantitative sequence number, and scope is 1 ..., m; Then each

Computing method as follows:

${\hat{q}}_u=C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y_i-y_c}{\sqrt{h^2+w^2}}\left|\right|}^2\right)\mathrm{\δ}[b\left(y_i\right)-u]$

Wherein, y _iRepresent certain pixel coordinate of this motion target area, total number of pixels of target area is n, b (y _i) function obtains the quantification sequence number of this location of pixels pixel value under quantizing in the space, C is a normalization coefficient, makes

Satisfy ${\mathrm{\Σ}}_{n=1}^m{\hat{q}}_u=1$ Requirement; δ (x) function is the Kronecker function, is defined as:

Function k (x) is the Epanechnikov function, and form is

D wherein, c _dBe constant.

6, method for monitoring instruction according to claim 1 is characterized in that, adopts based on the method for histogram feature coupling and comes the pursuit movement target; At t constantly, the weighted histogram model of moving target A is

At t+1 constantly, calculate near the weighting color model of the several moving targets of target A original position

And and

Relatively try to achieve similarity

Then the moving target of maximum similarity correspondence and target A are complementary:

${\hat{q}}_A^{t+1}=\arg \underset{{\hat{q}}_t^{t+1}}{\max }s({\hat{q}}_A^t,{\hat{q}}_i^{t+1})=\arg \underset{{\hat{q}}_t^{t+1}}{\max }\left({\mathrm{\Σ}}_{u=1}^m\sqrt{{\hat{q}}_A^u,{\hat{q}}_i^u}\right)$

Thereby realized the tracking of moving target by the target of continuous coupling consecutive frame; For the moving target on following the tracks of, judge whether its movement locus drops in the warning region of prior appointment, judge with this whether invasion takes place and report to the police.

Images