CN106056060A - Method and system for masked veil detection in video image - Google Patents

Abstract

The invention relates to a method and system for masked veil detection in a video image. According to the method, after a video stream is read in, a background model is built through adoption of mixed gauss background modeling, and a pixel point of a mobile object is got by utilizing a background subtraction method and then is taken as a foreground point; morphological operation is performed on the foreground point, a noise is removed, a hole is filled at the same time, a connected region is obtained and marked, and a continuously marked region is extracted and is taken as a final foreground region; extraction of characteristics of the mobile object is performed on the final foreground region; a classifier is adopted to classify the extracted characteristics, and whether a pedestrian exists at the foreground region is detected; when the pedestrian is detected, a head and shoulder position of the pedestrian is detected, and, on this basis, whether a human face of the pedestrian is complete; and, if the human face is not complete, the pedestrian is marked, a movement direction of the pedestrian is continuously tracked and recorded, and, when the movement direction of the pedestrian is determined to be a predetermined direction, an alarm is given out. So an effective detection means is provided for the effective masked veil detection in the video image.

Description

The method and system of carpe detection in a kind of video image

Technical field

The invention belongs to image identification technical field, be specifically related to the method for carpe detection in a kind of video image and be System.

Background technology

In some unsafe areas, for the needs of social safety, need some personages wearing special dress ornament are entered Row detects identification especially, accomplishes prevention and control in advance.Although corresponding region can be entered with the image in acquisition monitoring region at present Row in real time monitoring, but cannot some non-persistent personages targetedly, monitor knowledge in advance as worn the personage of carpe Not and carry out early warning.

Summary of the invention

It is an object of the invention to solve above-mentioned technical problem and the side of carpe detection in a kind of video image is provided Method and system.

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

A kind of method of carpe detection in video image, comprises the following steps:

Read in video flowing and use mixed Gaussian background modeling to set up background model；

Background subtraction method and background model is utilized to ask for the pixel of mobile object as foreground point；

Foreground point is carried out morphology operations, removes noise and hole is filled with simultaneously, obtain UNICOM region and carry out Labelling, and using continued labelling extracted region out as final foreground area；

The foreground area that this is final carries out the feature extraction of mobile object；

Using adaboost grader to classify the feature extracted, whether detection foreground area exists pedestrian；

When pedestrian having been detected, the head and shoulder model inspection preset is utilized to go out the head and shoulder position of this pedestrian, and basis at this The face of upper detection pedestrian is the most complete；

If face is imperfect, labelling also keeps track of the pedestrian of this labelling continuously, examines the direction of motion of this pedestrian Survey, judging that the direction of motion of this pedestrian is reported to the police when being predetermined direction.

The described haar feature that is characterized as, the face of described detection pedestrian the most completely uses eigenface detection method.

The described pixel utilizing background subtraction method and background model to ask for mobile object as the step of foreground point is: sentence Whether disconnected current modeling point mates with corresponding background model, if coupling, current modeling point is background dot, the most currently models Point is foreground point.

Whether described judgement currently models a little mates employing following steps with corresponding background model:

Judge the variable quantity of the background model that current modeling point is corresponding whether in the range of default matching threshold and or work as Whether the average gradient of front modeling point differs with the gradient in background model in the range of predetermined percentage threshold, if there being one Mate then not thinking.

Described matching threshold is expressed as:

Threshold_i=(Value_base+Theta)*Sensitivit y

Wherein, Value_baseBeing basic threshold, Theta is variance, and Sensitivit y is sensitivity.

The present invention also aims to provide the system of carpe detection in a kind of video image, including:

Background module sets up module, is used for reading in video flowing and using mixed Gaussian background modeling to set up background model；

Foreground point acquisition module, for utilizing background subtraction method and background model to ask for the pixel of mobile object as front Sight spot；

Foreground area determines module, for foreground point carries out morphology operations, removes noise and fills out hole simultaneously Fill, obtain UNICOM region and be marked, and using continued labelling extracted region out as final foreground area；

Characteristic extracting module, for carrying out the feature extraction of mobile object in this final foreground area；

Pedestrian detection module, for using adaboost grader to classify the feature extracted, detects foreground area Whether there is pedestrian；

Face detection module, has in when pedestrian having been detected, utilizes the head and shoulder model inspection preset to go out the head of this pedestrian Shoulder position, and the face detecting pedestrian on this basis is the most complete；

Direction of motion detection module, for detecting that face is imperfect and after labelling, keeping track of this labelling continuously The direction of motion of this pedestrian is also detected by pedestrian, and judges that the direction of motion of this pedestrian is to report during predetermined direction Alert.

The present invention, by above technical scheme, can detect the moving object in image also rapidly in video image Determine whether pedestrian, to detect face, the head and shoulder position detecting pedestrian after being judged as pedestrian the most completely judges that pedestrian is No carpe the tracking pedestrians direction of motion worn, and report to the police when traffic direction is predetermined warning direction, for regarding Frequently the masked man that in image, carpe is worn in quick detection provides one and effectively detects recognition methods.

Accompanying drawing explanation

The flow chart of the method for carpe detection in the video image that Fig. 1 provides for the embodiment of the present invention；

The schematic diagram of Haar feature shown in Fig. 2.

Detailed description of the invention

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

Shown in Figure 1, a kind of method of carpe detection in video image, including:

S101, reads in video flowing and uses mixed Gaussian background modeling to set up background model；

S102, utilizes background subtraction method and background model to ask for the pixel of mobile object as foreground point；

S103, carries out morphology operations to foreground point, removes noise and is filled with hole simultaneously, obtains UNICOM region also It is marked, and using continued labelling extracted region out as final foreground area；

S104, carries out the feature extraction of mobile object in the foreground area that this is final

Wherein, haar feature it is characterized as described in；

S105, uses adaboost grader to classify the feature extracted, and whether detection foreground area exists pedestrian；

In the present invention, described grader is mainly adaboost cascade of strong classifiers；

It is i.e. the adaboost grader feature to extracting will to be used to input the pedestrian's grader trained, detect prospect Whether region exists pedestrian, i.e. judges whether pedestrian, if do not existed, returning S101 and continues detection, if existing, carrying out next Step；

S106, when pedestrian having been detected, utilizes the head and shoulder model inspection preset to go out the head and shoulder position of this pedestrian, and at this On the basis of to detect the face of pedestrian the most complete；

The most complete at detection face, it is mainly based upon the detection of eigenface (Eigenface) method and carries out, detect whether to deposit In effective face information, its face is imperfect, and labelling carries out a step；

S107, if face is imperfect, labelling, and keep track of the pedestrian of this labelling continuously, the direction of motion to this pedestrian Carry out detecting, labelling, judging that the direction of motion of this pedestrian is reported to the police when being predetermined direction.

Concrete, namely judge the direction of motion of the incomplete pedestrian of this face for faced by photographic head time carry out warning and carry Show.

The present invention, by above technical scheme, can detect the moving object in image also rapidly in video image Determine whether pedestrian, to detect face, the head and shoulder position detecting pedestrian after being judged as pedestrian the most completely judges that pedestrian is No carpe the tracking pedestrians direction of motion worn, and report to the police when traffic direction is predetermined warning direction, for regarding Frequently the masked man that in image, carpe is worn in quick detection provides one and effectively detects recognition methods.

Below, in conjunction with concrete background modeling method, haar feature extraction, use adaboost grader to haar feature Classification illustrates that the present invention's realizes process.

Step1: read in video flowing；

Step2: use mixed Gaussian background modeling, set up background model；

Because video camera is fixed viewpoint so detected pedestrian is mobile target or mobile object, use mixed Gaussian background Modeling, can be mobile object by target lock-on after setting up background model, and provides rational target area for subsequent detection.

The flow process of background modeling mainly includes three steps: the initialization of background model, the coupling of background model update, the back of the body Scape learning success also detects prospect.Modeling point is divided into foreground point and background dot two class.Need to use difference during renewal learning Speed its Gauss model mated is updated.

Utilize present image to model point data for the first frame data to initialize the Gauss model distributed at first, depend on afterwards Model average, variance and the weight of correspondence is constantly trained according to the data of modeling point.When modeling point weight reaches being modeled as of setting Merit threshold value, illustrates that this modeling point models successfully, otherwise continues study, until weight meets threshold value.

By modeling point Gauss model parameter is constantly trained and learnt, increasing modeling point models successfully, system Counting modeling successfully modeling point quantity in whole two field picture, if reaching the 1/5 of entire image modeling point sum, then Background learning becomes Merit.Subsequently into the foreground detection stage, and the foreground point detected and background dot are updated background model with different rates respectively, To improve the adaptability of background model.

Step3: background modeling enters the foreground detection stage after completing, and i.e. utilizes background subtraction method to be extracted by mobile object It is used as prospect, mainly asks for what the pixel of mobile object or mobile target realized as foreground point.Concrete is permissible Current modeling point is utilized to carry out mating realizing with corresponding background model.The judgment threshold of coupling is and corresponding background model side Difference is relevant, and model can be according to the adaptive adjustment of the change of its value, and the Model Matching threshold value of each color component is represented by:

Threshold_i=(Value_base+Theta)*Sensitivit y

Wherein, Value_baseBeing basic threshold, Theta is variance, and Sensitivit y is sensitivity.If modeling is put and it The variable quantity of background model, in this threshold range, illustrates modeling point and this Model Matching.In addition current modeling point is average Gradient differs by more than certain value with the gradient in background model, such as 20%, is i.e. considered not mate.

If additionally in background model, background model does not model successfully, then being defaulted as this modeling point is background dot.Building Background model corresponding to mould point successfully in the case of, if this modeling point does not mates with its background model, then this modeling point is not necessarily For foreground point, also to do with the background model of the modeling point of its four neighborhood and mate, if coupling is the most unsuccessful, then this modeling point is true It is set to foreground point, otherwise, this point is still that background dot.

Step4: foreground point carries out morphology operations, removes noise and is filled with hole simultaneously, obtain UNICOM region And be marked.

Step5: using continued labelling extracted region out as final foreground area.

Step6: move the feature extraction of target or mobile object in foreground area.

Here feature extraction refers to haar feature extraction:

Haar feature is divided three classes: edge feature, linear character, central feature and diagonal feature, they are combined into spy Levy template, in feature templates, have white and two kinds of rectangles of black, and the eigenvalue defining this template is white rectangle pixel and subtracts Black removal rectangular pixels and.Haar eigenvalue reflects the grey scale change situation of image.

Some features of such as face simply can be described by rectangular characteristic, as eyes are deeper than cheek color, and the bridge of the nose Both sides are deeper than bridge of the nose color, and face is deeper etc. than ambient color.But rectangular characteristic is only to some simple graphic structures, such as limit Edge, line segment are more sensitive, so the structure at particular orientation (level, vertical, diagonal angle) can only be described, shown in Figure 2.

For this category feature of A, B and the D in Fig. 2, character numerical value computing formula is:

White-the Sum of v=Sum is black；

And for C, computing formula is as follows: the white-2*Sum of v=Sum is black；

Why by black region pixel be multiplied by 2, it is to make number of pixels in two kinds of rectangular areas consistent.By changing Become feature templates size and location, can in image subwindow exhaustive go out substantial amounts of feature.The feature templates of upper figure is referred to as " special Levy prototype "；Feature prototype extends the feature that (translation flexible) obtain in image subwindow and is referred to as " rectangular characteristic "；Rectangular characteristic Value be referred to as " eigenvalue ".

Step7: use adaboost grader to classify features described above, whether detection foreground area exists effective row People, then proceeds to next step as existed, otherwise re-reads next frame video.

Adboost grader principle: AdaBoost is a kind of iterative algorithm, adds a new weak typing in each wheel Device, until reaching certain predetermined sufficiently small error rate.Each training sample is endowed a weight, shows that it is by certain Individual grader is selected into the probability of training set.If certain sample point is classified exactly, then a training under construction Concentrating, its selected probability is just lowered；On the contrary, if certain sample point is not classified exactly, then its weight Just it is improved.

The algorithm flow of Adaboost is as follows:

First, the weights distribution of training data is initialized.Each training sample is endowed identical when starting most Weight: 1/N.

$D_1=(w_{11},w_{12}...w_{1i}...,w_{1N}),w_{1i}=\frac{1}{N},i=1,2,...,N$

If it follows that certain sample point is classified exactly, then under construction in a training set, it is selected In probability be just lowered；On the contrary, if certain sample point is not classified exactly, then its weight is just improved. Particularly, then:

2. for m=1,2 ..., M

A. use the training dataset study with weights distribution Dm, obtain basic binary classifier:

Gm (x): χ → {-1 ,+1}

B. the Gm (x) error in classification rate on training dataset is calculated

$e_m=P\left(G_m\right(x_i)\≠y_i)=\underset{i=1}{\overset{N}{\Σ}}{w}_{mi}I\left(G_m\right(x_i)\≠y_i)$

C. calculating the coefficient of Gm (x), am represents the Gm (x) significance level in final grader:

${\mathrm{\α}}_m=\frac{1}{2}log\frac{1-e_m}{e_m}$

From above-mentioned formula, during em≤1/2, am >=0, and am increases along with the reduction of em, it is meant that classification is by mistake The effect in final grader of the rate the least basic classification device is the biggest.

D. the weights distribution of training dataset is updated

D_m+1=(w_M+1,1, w_M+1,2…w_{M+1, i}..., w_{M+1, N}),

$w_{m+1,i}=\frac{w_{mi}}{Z_m}exp(-{\mathrm{\α}}_m{y}_i{G}_m(x_i\left)\right),i=1,2,...,N$

Make to be increased by the weights of basic classification device Gm (x) misclassification sample, and reduced by the weights of correct classification samples. Like this, by such mode, AdaBoost method can " focus on " on those samples of more difficult point.

Wherein, Zm is standardizing factor so that Dm+1 becomes a probability distribution:

$Z_m=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{w}_{mi}exp(-{\mathrm{\α}}_m{y}_i{G}_m(x_i\left)\right)$

3. build the linear combination of basic classification device

$f\left(x\right)=\underset{m=1}{\overset{M}{\Σ}}{\mathrm{\α}}_m{G}_m\left(x\right)$

Thus obtain final grader, as follows:

$G\left(x\right)=sign\left(f\right(x\left)\right)=sign\left(\underset{m=1}{\overset{M}{\Σ}}{\mathrm{\α}}_m{G}_m\right(x\left)\right)$

Step8: the pedestrian detected is utilized head and shoulder model, detects pedestrian's head and shoulder position.

Head and shoulder model presents Ω shape mainly by head and shoulder shape, screens.Gather the head and shoulder picture conduct of pedestrian Positive sample, uses feature learning method above to carry out features training and study, obtains pedestrian's head and shoulder position.

Step9: detect complete face on this basis, then illustrates that this person is non-wear carpe people as there is complete face Scholar, if there is not complete face then for masked, is marked masked personage.

Step10: trace labelling target continuously, travel direction detects.Masked personage is tracked labelling and passes through position Put change calculations and go out the direction of target travel.

Step11: carry out alarm decision for above testing result, if incomplete face and the direction of motion are to face shooting Machine prospect, then it is assumed that be carpe personage, report to the police, otherwise occurs without effective target, reads next frame.

Step12: output is reported to the police.

The present invention also aims to provide the system of carpe detection in a kind of video image, including:

Background module sets up module, is used for reading in video flowing and using mixed Gaussian background modeling to set up background model；

Foreground point acquisition module, for utilizing background subtraction method and background model to ask for the pixel of mobile object as front Sight spot；

Foreground area determines module, for foreground point carries out morphology operations, removes noise and fills out hole simultaneously Fill, obtain UNICOM region and be marked, and using continued labelling extracted region out as final foreground area；

Characteristic extracting module, for carrying out the feature extraction of mobile object in this final foreground area；

Pedestrian detection module, for using adaboost grader to classify the feature extracted, detects foreground area Whether there is pedestrian；

Face detection module, has in when pedestrian having been detected, utilizes the head and shoulder model inspection preset to go out the head of this pedestrian Shoulder position, and the face detecting pedestrian on this basis is the most complete；

Direction of motion detection module, for detecting that face is imperfect and after labelling, keeping track of this labelling continuously The direction of motion of this pedestrian is also detected by pedestrian, and judges that the direction of motion of this pedestrian is to report during predetermined direction Alert.

The detection method of system of carpe detection and principle in this video image, with embodiments of the invention video image The Method And Principle of middle carpe detection is identical, and at this, it will not be described.

The present invention, by above technical scheme, can detect the moving object in image also rapidly in video image Determine whether pedestrian, to detect face, the head and shoulder position detecting pedestrian after being judged as pedestrian the most completely judges that pedestrian is No carpe the tracking pedestrians direction of motion worn, and report to the police when traffic direction is predetermined warning direction, for regarding Frequently the masked man that in image, carpe is worn in quick detection provides one and effectively detects recognition methods.

The above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For Yuan, under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, these improvements and modifications also should It is considered as protection scope of the present invention.

Claims (10)

1. the method for carpe detection in a video image, it is characterised in that comprise the following steps:

Read in video flowing and use mixed Gaussian background modeling to set up background model；

Background subtraction method and background model is utilized to ask for the pixel of mobile object as foreground point；

Foreground point is carried out morphology operations, removes noise and hole is filled with simultaneously, obtain UNICOM region and be marked, And using continued labelling extracted region out as final foreground area；

The foreground area that this is final carries out the feature extraction of mobile object；

Using adaboost grader to classify the feature extracted, whether detection foreground area exists pedestrian；

When pedestrian having been detected, utilize the head and shoulder model inspection preset to go out the head and shoulder position of this pedestrian, and examine on this basis The face surveying pedestrian is the most complete；

If face is imperfect, labelling also keeps track of the pedestrian of this labelling continuously, detects the direction of motion of this pedestrian, Judging that the direction of motion of this pedestrian is reported to the police when being predetermined direction.

The method of carpe detection in video image the most according to claim 1, it is characterised in that described in be characterized as that haar is special Levying, the face of described detection pedestrian the most completely uses eigenface detection method.

The method of carpe detection in video image the most according to claim 1, it is characterised in that described utilize background subtraction Method and background model ask for the pixel of mobile object: judge current modeling point and corresponding background Whether model mates, if coupling, current modeling point is background dot, and otherwise current modeling point is foreground point.

The method of carpe detection in video image the most according to claim 2, it is characterised in that described judgement currently models Put and whether mate employing following steps with corresponding background model:

Judge the variable quantity of the background model that current modeling point is corresponding whether in the range of default matching threshold and or currently build Whether the average gradient of mould point differs in the range of predetermined percentage threshold with the gradient in background model, if there being one not exist Then not think and mate.

The method of carpe detection in video image the most according to claim 3, it is characterised in that described matching threshold represents For:

Threshold_i=(Value_base+Theta)*Sensitivity

Wherein, Value_baseBeing basic threshold, Theta is variance, and Sensitivity is sensitivity.

6. the system of carpe detection in a video image, it is characterised in that including:

Background module sets up module, is used for reading in video flowing and using mixed Gaussian background modeling to set up background model；

Foreground point acquisition module, for utilizing background subtraction method and background model to ask for the pixel of mobile object as prospect Point；

Foreground area determines module, for foreground point carries out morphology operations, removes noise and is filled with hole simultaneously, To UNICOM region and be marked, and using continued labelling extracted region out as final foreground area；

Characteristic extracting module, for carrying out the feature extraction of mobile object in this final foreground area；

Pedestrian detection module, for using adaboost grader to classify the feature extracted, whether detection foreground area There is pedestrian；

Face detection module, has in when pedestrian having been detected, utilizes the head and shoulder model inspection preset to go out the head and shoulder position of this pedestrian Put, and the face detecting pedestrian on this basis is the most complete；

Direction of motion detection module, for detecting that face is imperfect and after labelling, keeping track of the pedestrian of this labelling continuously And the direction of motion of this pedestrian is detected, and judge that the direction of motion of this pedestrian is to report to the police during predetermined direction.

The system of carpe detection in video image the most according to claim 6, it is characterised in that described in be characterized as that haar is special Levying, the face of described detection pedestrian the most completely uses eigenface detection method.

The most according to claim 6 the video image culminant star moon pattern detection system, it is characterised in that described utilize background to subtract Division and background model ask for the pixel of mobile object: judge current modeling point and the corresponding back of the body Whether scape model mates, if coupling, current modeling point is background dot, and otherwise current modeling point is foreground point.

The system of carpe detection in video image the most according to claim 6, it is characterised in that described judgement currently models Put and whether mate employing following steps with corresponding background model:

Judge the variable quantity of the background model that current modeling point is corresponding whether in the range of default matching threshold and or currently build Whether the average gradient of mould point differs in the range of predetermined percentage threshold with the gradient in background model, if there being one not exist Then not think and mate.

10. according to the system of carpe detection in video image described in any one of claim 6-9, it is characterised in that described Join threshold value table to be shown as:

Threshold_i=(Value_base+Theta)*Sensitivity

Wherein, Value_baseBeing basic threshold, Theta is variance, and Sensitivity is sensitivity.