CN106022279A - Method and system for detecting people wearing a hijab in video images - Google Patents

Abstract

The invention relates to a method and system for detecting people wearing a hijab in video images. The method includes that acquiring a frame of video image and setting the interest region to be detected in the frame video image; performing foreground detection on the interest region according to a preset foreground detection method, and then acquiring the size and position of the foreground region to be detected; performing pedestrian detection in the foreground region to determine whether a pedestrian exists in the foreground region; if yes, acquiring the pedestrian contour and detecting the pedestrian by a classifier with hijab characteristics to determine whether the pedestrian has the hijab characteristics; if the pedestrian has the hijab characteristics, marking the pedestrian and determining whether the number of image frames having the hijab characteristics detected within a predetermined statistical region in a subsequent predetermined number of image frames is greater than the preset misdetection frame number, if yes, outputting an alarm. Therefore, an effective detection means for effectively performing hijab detection in video images is provided.

Description

A kind of video image invading the interior cuts gram method and system for detection

Technical field

The invention belongs to image identification technical field, be specifically related to a kind of video image invading the interior and cut gram detection Method and system.

Background technology

In some unsafe areas, for the needs of social safety, some are worn special dress ornament Personage needs to detect identification the most especially, overcomes the personage of decorations to carry out detection identification as cut in dress, To accomplish to prevent in advance.The personage that image detection can realize in image is identified, but conventional images inspection Surveying knowledge technology cannot overcome the personage of decorations to prevent detection to identify in advance to cutting in dress effectively, accomplishes to carry Front prevention.

Summary of the invention

It is an object of the invention to solve above-mentioned technical problem and provide a kind of video image invading the interior to cut gram The method and system of detection.

For achieving the above object, the present invention adopts the following technical scheme that

A kind of video image invading the interior cuts a gram method for detection, comprises the following steps:

Obtain a frame video image and area-of-interest to be detected in this frame video image is set；

Described area-of-interest is carried out foreground detection by default foreground detection method, then obtains to be detected The size of foreground area and position；

Carry out pedestrian detection in described foreground area, determine in described foreground area whether there is pedestrian；

If there is pedestrian, then obtain pedestrian contour and with the grader cutting korte's sign in comprising, pedestrian carried out Detection, it is judged that this pedestrian cuts korte's sign in whether having；

If this pedestrian cuts korte's sign in having, then this pedestrian of labelling judge that this pedestrian is at follow-up predetermined number In picture frame in default statistical regions in detect have in cut the number of image frames of korte's sign the biggest In default flase drop frame number, if then exporting alarm.

Described area-of-interest is the humanoid district that in this frame video image, definition meets default definition values Territory.

Described foreground detection method is Vibe foreground detection method.

The size of the foreground area that described acquisition is to be detected and position employing following steps:

The binary image exported by foreground detection is repeatedly corroded and connects into a region, so Size and the position of this foreground area is obtained afterwards by edge detection method.

Cutting korte's sign in described is Haar feature, and described grader is Adaboost cascade classifier.

The present invention also aims to provide a kind of video image invading the interior to cut a gram system for detection, including:

Area-of-interest acquisition module, treats for obtaining a frame video image arranging in this frame video image The area-of-interest of detection；

Foreground area determines module, before carrying out described area-of-interest by default foreground detection method Scape detects, and then obtains size and the position of foreground area to be detected；

Pedestrian detection module, for carrying out pedestrian detection in described foreground area, determines described foreground area In whether there is pedestrian；

Li Qieke judge module, in the case of being used for detecting and there is pedestrian, obtains pedestrian contour and with wrapping Containing the inner grader cutting korte's sign, pedestrian is detected, it is judged that this pedestrian cuts korte's sign in whether having；

Detection confirms module, and in time cutting korte's sign in this pedestrian has, this pedestrian of labelling also judges to be somebody's turn to do Pedestrian in follow-up predetermined number picture frame in default statistical regions in detect have in cut cut Whether the number of image frames levied is more than presetting flase drop frame number, and the output alarm when judging to be.

Whether the present invention, by above technical scheme, can detect in image rapidly in video image There is pedestrian, after being judged as pedestrian, be quickly detected from pedestrian has cutting korte's sign, and after validation Output is reported to the police, and provides one effectively for cutting the pedestrian of korte's sign in quickly detection has in video image Detection recognition methods.

Accompanying drawing explanation

Fig. 1 cuts a gram flow chart for the method for detection for the video image invading the interior that the embodiment of the present invention provides；

The expression figure of Haar-like feature shown in Fig. 2；

Fig. 3 show integral image；

Fig. 4 show the schematic diagram of cascade of strong classifiers.

Detailed description of the invention

Below, in conjunction with example, substantive distinguishing features and the advantage of the present invention are further described, but this Bright it is not limited to listed embodiment.

Shown in Figure 1, a kind of video image invading the interior cuts a gram method for detection, including:

S101, obtains a frame video image and arranges area-of-interest to be detected in this frame video image；

S102, carries out foreground detection to described area-of-interest by default foreground detection method, then obtains The size of foreground area to be detected and position；

S103, carries out pedestrian detection in described foreground area, determines in described foreground area whether there is row People；If there is not pedestrian, return step S101,

S104, if there is pedestrian, then obtains pedestrian contour and cuts the grader of korte's sign to row in comprising People detects, it is judged that this pedestrian cuts korte's sign in whether having；It cuts korte's sign in not existing, then return Return step S101,

S105, if this pedestrian cuts korte's sign in having, then this pedestrian of labelling judge that this pedestrian is follow-up pre- In fixed number mesh picture frame in default statistical regions in detect have in cut the number of image frames of korte's sign Whether more than presetting flase drop frame number, if then exporting alarm, otherwise return step S101.

Whether the present invention, by above technical scheme, can detect in image rapidly in video image There is pedestrian, after being judged as pedestrian, be quickly detected from pedestrian has cutting korte's sign, and after validation Output is reported to the police, and provides one effectively for cutting the pedestrian of korte's sign in quickly detection has in video image Detection recognition methods.

Implementing, in the present invention, described area-of-interest is that in this frame video image, definition meets The humanoid region of testing requirement, i.e. definition meet the humanoid region of default definition values, are specifically realizing Time, can be after the width obtaining sequence of video images and high data, according to the wide and high number of video image According to calculating the position in video image of area-of-interest and width height to determine this region interested, And follow-up detection will be carried out in this area-of-interest, so ratio detects institute's used time in entire image Between to lack a lot, thus be effectively improved detection speed.

Implementing, in the present invention, described foreground detection method is Vibe (Visual Background Extractor) foreground detection method.

Described Vibe foreground detection method is mainly detected by following principle: it is by static or non- The most slowly moving object regards background object as, and the object of corresponding movement regards foreground object as, thus handle Object detection regards a classification problem as, namely determines whether an image slices vegetarian refreshments belongs to background dot, Background dot is distinguished with foreground point, thus realizes foreground detection, it is thus achieved that corresponding foreground area.

In ViBe detection model, its background model is that each background dot stores a sample set, then The pixel value new by each and sample set are compared to judge whether to belong to background dot.If one new Observed value belong to background dot, then it should with the sampled value in sample set relatively.

Specifically, note v (x) is the pixel value at x point；M (x)={ V₁,V₂,...V_NIt it is the background sample at x This collection (sample set size is N)；SR (v (x)) is the region as radius of the R centered by x, if M (x) [{SR(v(x))∩{v₁,v₂,...,v_N] more than given threshold value min, it is judged that x point belongs to Background dot.

ViBe model initialization is exactly the process of the sample set of filler pixels, but due in a two field picture not The spatial and temporal distributions information of pixel may be comprised, utilize close pixel to have close spatial and temporal distributions special Property, it is exactly specifically that, for a pixel, the pixel value of the random neighbours' point selecting it is as it Model sample value.M₀(x)={ v₀(y | y ∈ NG (x)) }, t=0 initial time, NG (x) is neighbours Point.This initial method advantage is that the reaction for noise is sensitiveer, and the little speed of amount of calculation is fast, Can the detection carrying out moving object quickly, shortcoming is to be readily incorporated Ghost region.

The more New Policy used that updates of model is conservative more New Policy+foreground point method of counting.Foreground point Counting is i.e. to add up pixel, if certain pixel N continuous time is detected as prospect, then will It is updated to background dot.Random sub sampling: go to update in background model in each new video frame Each pixel sample value it is not necessary that, when a pixel is classified as background dot, It hasProbability go update background model.

Concrete update method, each background dot hasProbability go to update the model sample value of oneself, Also have simultaneouslyProbability go to update its model sample value of neighbours' point.The sample value profit of more new neighbor By the spatial transmission characteristic of pixel value, background model is gradually to external diffusion, and this is also beneficial to Ghost The faster identification in region.It is changed into background when foreground point counting reaches marginal value simultaneously, and has Probability go to update the model sample value of oneself.When sample value in selecting sample set to be replaced, We are to randomly select a sample value to be updated, and so can ensure that the smooth Life Cycle of sample value Phase, the probability that such a sample value is not updated at moment t was (N-1)/N owing to being random renewal, The time of hypothesis is continuous print, then after past time of dt, the probability that sample value still retains is

$P(t,t+dt)={\left(\frac{N-1}{N}\right)}^{(t+dt)-t}$

Can also write

$P(t,t+dt)=e^{-\ln \left(\frac{N}{N-1}\right)dt}.$

It is unrelated with time t that this indicates that the most whether a sample value is replaced, and randomized policy is Suitably.

In the present invention, being carried out after foreground detection completes by above method, described obtaining is to be detected The size of foreground area and position can use following steps:

The binary image exported by foreground detection is repeatedly corroded and connects into a district of UNICOM Territory, then obtains size and the position of this foreground area by edge detection method.It is so follow-up spy Levy and detect to determine whether the basis of detection established for pedestrian.

In the present invention, export after described foreground detection is the foreground image of a binaryzation, this foreground picture As pixel value is 1, background image pixel values is 0, and such foreground image is not a connected region, In order to subsequent detection needs, doing this two-value foreground image and repeatedly corrode, connection becomes a continuous print district Territory, then the size and location of foreground area place rectangle can be obtained by rim detection.In the prospect of acquisition After the size and location of place, region rectangle, it is possible to described foreground area carries out pedestrian detection, determine institute State in foreground area and whether there is pedestrian, thus substantially increase pedestrian with or without the speed detected and effect Rate.

In the present invention, described in cut korte's sign be Haar feature, described grader is Adaboost cascade Grader, is specially Adaboost cascade classifier based on Haar feature.By using Haar special Levy Adaboost cascade classifier and carry out pedestrian detection in foreground area, reduce Li Qieke detection further Region, improve efficiency and the accuracy rate of detection.

The principle of Adaboost cascade classifier based on Haar feature is described below:

About Haar feature and integrogram

1) Haar-like feature can be represented by Fig. 2, and each feature is made up of 2-3 rectangle, at this In a little small echo schematic diagrams, light areas represents " cumulative data ", and darker regions represents and " deducts this region Data ".Detect respectively boundary characteristic (Edge features), line feature (Line features), Center ring characteristics (Center-surround features), these features are represented by:

${\mathrm{feature}}_j=\underset{i\∈(1,...,N)}{\mathrm{\Σ}}{\mathrm{\ω}}_i*\mathrm{Re}ctSum\left(r_i\right)$

Wherein, w_iFor the power of rectangle, RectSum (r_i) it is rectangle r_iThe gray integration of enclosed image, N It is composition feature_jRectangle number.

As in the picture of a 24*24, there are 115984 features, are far longer than its number of pixels. If calculate each feature pixel and, amount of calculation can be very big, and computing many times is to repeat.

To this end, Paul Viola proposes a kind of method utilizing integral image method quickly to calculate Haar feature, The method is in brief, it is simply that first (Integral image is also Summed to structure one " integrogram " Area Table), any one Haar rectangular characteristic can be by the method tabled look-up and limited number of time afterwards Simple operation obtains, and greatly reduces operation times.

Rectangle is expressed as:

R=(x, y, w, h, α) 0≤x, x+w≤W, 0≤y, y+h≤H, x, y >=0, w, h ＞ 0, α ∈ { 0 °, 45 ° }

Wherein, x, y represent that starting point coordinate, w, h represent wide, and high, a represents angle.

1. in rectangle, pixel value sum is expressed as: RecSum (r)

2. integral image (Summed Area Table) is constructed, as shown in Figure 3.

In integral image, the storage of each point is its upper left side all pixels sum:

$SAT(x,y)=\underset{x^{\′}\≤x,y^{\′}\≤y}{\mathrm{\Σ}}I(x^{\′},y^{\′})$

Wherein, (x y) represents image (x, y) pixel value of position to I.Integral image can use increment Mode calculates:

SAT (x, y)=SAT (x, y-1)+SAT (x-1, y)+I (x, y)-SAT (x-1, y-1)

Initial boundary: SAT (-1, y)=SAT (x ,-1)=SAT (-1 ,-1)=0

So, it is only necessary to the most just can be in the hope of the integral image of this figure to whole image traversal.

About adaboost grader.

Adaboost grader is to utilize Weak Classifier that classification capacity is general by certain method superposition (boost) get up, the strong classifier that composition and classification is very capable.Adaboost training strong classifier Arthmetic statement is as follows:

Given a series of training sample (x₁,y₁),(x₂,y₂),...(x_n,y_n), wherein x_iRepresent the I sample, y_iIt is positive sample (face) when=1, y_i=0 represents negative sample (non-face).To often Individual feature feature_j, train a Weak Classifier h_j(x), weak point afterwards each feature generated Class device calculating weighted error:

${\mathrm{\ϵ}}_j=\underset{i}{\mathrm{\Σ}}{\mathrm{\ω}}_i|h_j\left(x_i\right)-y_i|$

To have minimum error e_jGrader be added in strong classifier, and update the general of training sample Rate is distributed:

$D_j\left(i\right)=D_j\left(i\right){{\mathrm{\β}}_j}^{1-e_i}$

Wherein β_j=ε_j/1-ε_j,e_i=0 represents sample x_iCorrectly classified, otherwise e_i=1 represents not Correctly classified, finally constituted strong classifier:

$H\left(x\right)=sign\[\underset{j=1}{\mathrm{\Σ}}{h}_j\left(x\right)-b\]$

Wherein b is the threshold value arranged, and is defaulted as 0.

The schematic diagram of cascade of strong classifiers as shown in Figure 4, by multiple Haar-like features F₁F₂…F_nConnection Collectively form decision tree, by data and the plurality of Haar-like feature F will be inputted₁F₂…F_nEnter Row compare, it is judged that meet and be output as face (face), the most incongruent be output as non- Face (not face).

Viola-Jones detector utilizes waterfall (Cascade) algorithm classification device to be organized as screening type Cascade classifier, each node of cascade is that AdaBoost trains the strong classifier obtained.In cascade Each node arranges threshold value b so that nearly all face sample can be transferred through, and the overwhelming majority is non-face Sample can not pass through.Node arranges from simple to complex, and position node the most rearward is the most complicated, i.e. comprises The most Weak Classifiers.So energy minimization refusal image but amount of calculation during region, notice ensures classification The high detection rate of device and low reject rate.Such as it is 99.9% at discrimination, when reject rate is 50%, (99.9% Face and 50% non-face can pass through), total discrimination of 20 nodes is: 98%, and mistake Receptance is only: 0.0001%.

It should be noted that in the present invention, there is pedestrian in judgement, and obtain pedestrian contour and with comprising In cut the grader of korte's sign pedestrian detected, it is judged that when this pedestrian cuts korte's sign in whether having, Its detection method is identical with described foreground detection method principle, and the training sample simply entered is for cutting in comprising The image of korte's sign.I.e. can detect that pedestrian cuts korte's sign in whether having by the detection of this step, and Obtain Li Qieke target position in this two field picture and size labelling.

It should be noted that in the present invention, in order to prevent flase drop, when adding up, be with multiframe figure As being the unit number of times that carries out cutting in statistic mixed-state goes out korte's sign, when in default statistical regions, Carry out alarm when cutting the number of times of korte's sign in detecting more than the threshold value set, do not report to the police, continue Continue and detect as stated above.

Concrete can be with 10 two field pictures is that a statistic unit is added up, in its first two field picture Korte's sign is cut in being tested with, and in detecting in 9 follow-up two field pictures in default statistical regions The number of times cutting korte's sign is more than 5, then export warning, otherwise it is assumed that flase drop, continues detection.

In the present invention, described statistical regions is that the first frame being determined by a statistical unit detects In cut korte's sign position in the picture and size after, centered by cutting korte's sign place rectangle in this, to Outer expansion, as expanded 1/3 one enlarged area of formation of this rectangle length of side, using this enlarged area as system Counting region and formed, after this frame, continuous print multi frame detection is all to enter in the region preset of this expansion OK, so desirably prevent missing inspection, improve detection precision.

The present invention also aims to provide a kind of video image invading the interior to cut a gram system for detection, including:

Area-of-interest acquisition module, treats for obtaining a frame video image arranging in this frame video image The area-of-interest of detection；

Foreground area determines module, before carrying out described area-of-interest by default foreground detection method Scape detects, and then obtains size and the position of foreground area to be detected；

Pedestrian detection module, for carrying out pedestrian detection in described foreground area, determines described foreground area In whether there is pedestrian；

Li Qieke judge module, in the case of being used for detecting and there is pedestrian, obtains pedestrian contour and with wrapping Containing the inner grader cutting korte's sign, pedestrian is detected, it is judged that this pedestrian cuts korte's sign in whether having；

Detection confirms module, and in time cutting korte's sign in this pedestrian has, this pedestrian of labelling also judges to be somebody's turn to do Pedestrian in follow-up predetermined number picture frame in default statistical regions in detect have in cut cut Whether the number of image frames levied is more than presetting flase drop frame number, and the output alarm when judging to be.

Described video image invading the interior is cut in gram implementation method of the system of detection and above-mentioned video image The method of Li Qieke detection is identical, is no longer described in detail at this.

Whether the present invention, by above technical scheme, can detect in image rapidly in video image There is pedestrian, after being judged as pedestrian, be quickly detected from pedestrian has cutting korte's sign, and after validation Output is reported to the police, and provides one effectively for cutting the pedestrian of korte's sign in quickly detection has in video image Detection recognition methods.

The above is only the preferred embodiment of the present invention, it is noted that general for the art For logical technical staff, under the premise without departing from the principles of the invention, it is also possible to make some improvement and profit Decorations, these improvements and modifications also should be regarded as protection scope of the present invention.

Claims (10)

1. a video image invading the interior cuts a gram method for detection, it is characterised in that comprise the following steps:

Obtain a frame video image and area-of-interest to be detected in this frame video image is set；

Described area-of-interest is carried out foreground detection by default foreground detection method, then obtains to be detected The size of foreground area and position；

Carry out pedestrian detection in described foreground area, determine in described foreground area whether there is pedestrian；

If there is pedestrian, then obtain pedestrian contour and with the grader cutting korte's sign in comprising, pedestrian carried out Detection, it is judged that this pedestrian cuts korte's sign in whether having；

If this pedestrian cuts korte's sign in having, then this pedestrian of labelling judge that this pedestrian is at follow-up predetermined number In picture frame in default statistical regions in detect have in cut the number of image frames of korte's sign the biggest In default flase drop frame number, if then exporting alarm.

The most according to claim 1, video image invading the interior cuts a gram method for detection, it is characterised in that Described area-of-interest is the humanoid region that in this frame video image, definition meets default definition values.

The most according to claim 2, video image invading the interior cuts a gram method for detection, it is characterised in that Described foreground detection method is Vibe foreground detection method.

The most according to claim 2, video image invading the interior cuts a gram method for detection, it is characterised in that The size of the foreground area that described acquisition is to be detected and position employing following steps:

The binary image exported by foreground detection is repeatedly corroded and connects into a region, so Size and the position of this foreground area is obtained afterwards by edge detection method.

The most according to claim 3, video image invading the interior cuts a gram method for detection, it is characterised in that Cutting korte's sign in described is Haar feature, and described grader is Adaboost cascade classifier.

6. a video image invading the interior cuts a gram system for detection, it is characterised in that including:

Area-of-interest acquisition module, treats for obtaining a frame video image arranging in this frame video image The area-of-interest of detection；

Foreground area determines module, before carrying out described area-of-interest by default foreground detection method Scape detects, and then obtains size and the position of foreground area to be detected；

Pedestrian detection module, for carrying out pedestrian detection in described foreground area, determines described foreground area In whether there is pedestrian；

Li Qieke judge module, in the case of being used for detecting and there is pedestrian, obtains pedestrian contour and with wrapping Containing the inner grader cutting korte's sign, pedestrian is detected, it is judged that this pedestrian cuts korte's sign in whether having；

Detection confirms module, and in time cutting korte's sign in this pedestrian has, this pedestrian of labelling also judges to be somebody's turn to do Pedestrian in follow-up predetermined number picture frame in default statistical regions in detect have in cut cut Whether the number of image frames levied is more than presetting flase drop frame number, and the output alarm when judging to be.

The most according to claim 6, video image invading the interior cuts a gram system for detection, it is characterised in that Described area-of-interest is the humanoid region that in this frame video image, definition meets default definition values.

The most according to claim 6 the video image culminant star moon pattern detection system, it is characterised in that Described foreground detection method is Vibe foreground detection method.

The most according to claim 6, video image invading the interior cuts a gram system for detection, it is characterised in that The size of the foreground area that described acquisition is to be detected and position employing following steps:

The binary image exported by foreground detection is repeatedly corroded and connects into a region, so Size and the position of this foreground area is obtained afterwards by edge detection method.

10. cutting a gram system for detection according to video image invading the interior described in any one of claim 6-9, it is special Levy and be, described in cut korte's sign be Haar feature, described grader is Adaboost cascade classifier.