CN101719216B - Movement human abnormal behavior identification method based on template matching - Google Patents

Abstract

The invention is a method for identifying abnormal behavior of a moving human body based on template matching. It mainly includes video image acquisition, behavior feature extraction, sample-based statistical learning and pattern recognition technology. Use computer vision technology to analyze and understand human motion, conduct behavior recognition and record and alarm directly based on the geometric calculation of the motion area, use the combination of Gaussian filter denoising and neighborhood denoising to achieve denoising, and improve the performance of the intelligent monitoring system With independent analysis performance and intelligent monitoring capabilities, it has high recognition accuracy for abnormal behaviors, can effectively remove the complex background and noise of visually collected images, and improves the efficiency and robustness of detection algorithms. The invention has simple modeling, concise algorithm and accurate detection, and can be widely used in banks, museums and other places. It also helps to improve the level of security monitoring in public places.

Description

Movement human abnormal behaviour recognition methods based on template matches

Technical field

The invention belongs to computer vision, Intelligent Information Processing field.Relate to computer monitoring technology, based on the mode identification technology of statistical learning based on moving image.Relating generally to a kind of video monitoring content intelligent analysis method, specifically is a kind of movement human abnormal behaviour recognition methods based on template matches.

Background technology

Along with the attention of society to the public safety problem, monitoring has in real time obtained application more and more widely.The problem of existing supervisory system mainly is that a large amount of monitor messages is difficult to handled timely and effectively, through the identification of computer assisted to human behavior and incident, has become a hot issue of computer vision field.

The Intellectual Analysis Technology of vision monitoring is the focus and the difficult point problem of computer vision field, relates to problems such as Flame Image Process, machine learning.Academicly carry out more correlative study in recent years, comprised the research of intelligent monitoring project in national high-tech research development plan and the state key fundamental research development plan.

Chinese scholars has been done a lot of work on the abnormality detection based on video sequence, be broadly divided into two types: based on the method for model with based on the method for similarity measure.First kind method preestablishes criterion; From image sequence, extract the information such as profile, motion of moving target then; According to the characteristic information that these obtained; Artificial or use the model of semi-supervised method definition normal behaviour, to select usually to use HMM or graph model to carry out to carrying out modeling by the represented state of sequential image feature, the observation of those normal behaviour models that do not match all is considered to unusual.All normal events by the situation of fine modeling under, detect functional based on the method for model.But, when modeling encounters difficulties fully, detect effect and just can descend when normal behavior quantity is very big.In second class methods, utilize the difficulty definition of abnormal behaviour, detectable characteristics to make to need not people in advance that explicit definition goal behavior model just can detect it.Its ultimate principle is the normal pattern of automatically from video sequence data, learning, and infers suspicious abnormal behaviour then.

The abnormal behaviour detection is meant at first to be analyzed respectively and modeling self-defining regular event behavior and other behavioral datas, judges according to behavioral test and both similar programs whether behavior has unusually then.Present research mainly concentrates moving region, speed through human body whether to satisfy restrictions such as predetermined condition, detects elementary events such as the personage occurs, crosses the border.Company such as U.S. Object Video has developed corresponding product on this basis, and has obtained certain application.

Yet also there are certain problem in these researchs and application: present study portion vision understanding method adopts the manikin method of machinery; Through the match video image; Recover the state of human synovial motion; And then understand the behavior of human body, handling has like this increased intermediate link, makes model complicated more and be difficult to accurate realization.

The surveillance of existing widespread use also all is to be main with shooting and document image, or is manual monitoring, or is afterwards playback and analysis; Basically there is not Intellectualized monitoring; Even some monitoring to environment and still life has warning system, also can only judge for the normal and unusual of dynamic or moving object with moving and motionless as the whether normal criterion of difference; Be in theoretical research and discussion stage mostly, the abnormal behaviour recognition methods that is not of practical significance is employed.

Project team of the present invention does not find report or the document closely related and the same with the present invention as yet to domestic and international patent documentation and the journal article retrieval of publishing.

Summary of the invention

The objective of the invention is to overcome the shortcoming of above-mentioned technology or method existence; A kind of motion that utilizes the computer vision technique analysis and understand the people is provided; Directly carry out behavior and discern go forward side by side line item and warning based on the geometrical calculation of moving region; The autonomous analytical performance and the intelligent monitoring ability of system have been improved; Abnormal behaviour there is higher recognition accuracy, can effectively removes the complex background and the noise of video acquisition image, improved the movement human abnormal behaviour recognition methods based on template matches of the efficient and the robustness of detection algorithm.The application of this method helps to improve the security monitoring level of public place.

Be elaborated in the face of the present invention down

The present invention is a kind of movement human abnormal behaviour recognition methods based on template matches, it is characterized in that: said method comprising the steps of:

One. sample video data acquiring: through the camera collection normal behaviour video sequence in the define behavior in advance, as training sample;

Two. graphical analysis and behavioural characteristic are extracted: utilize the background subtraction point-score to extract the moving region of gathering in the video sequence image;

Three. measure the length and the width of the moving region that extracts, calculate long and wide ratio, classification and storage is on corresponding view data;

Four. behavior modeling: utilize length with the wide ratio different learning model buildings of being correlated with of a large amount of motion samples of above-mentioned collection, set up corresponding criterion behavior model by every kind of behavior;

Five. set up model bank: the criterion behavior category of model that obtains is saved in the model database, makes up the criterion behavior storehouse;

Six. real-time data acquisition: gather the behavior video sequence in real time through camera, and the employing method identical with step 2 carried out feature extraction;

Seven. interpretation of result: utilize similarity comparison algorithm based on Weighted H u invariant moments; Predefined normal behaviour relatively according to maximum-likelihood criterion, calculates the behavior to be identified and the similarity of define behavior in advance in tracking target that current frame image is promptly extracted and the criterion behavior database; Compare with setting threshold; Greater than similarity threshold be normal behaviour, less than similarity threshold be abnormal behaviour, determine unusual;

Eight. to the compact tracking of abnormal behaviour, mark and warning.

The present invention automatically learns normal behavior pattern from video sequence data, infer suspicious abnormal behaviour then.Be through analysis, extract the moving region and calculate length breadth ratio,, set up the human motion characteristic model, realize the real-time detection of human body abnormal behaviour through a large amount of normal and abnormal motion data are carried out machine learning to the moving region.

The present invention relates generally to computer vision and area of pattern recognition; Obtain the human motion zone through the background subtraction point-score; Real-time detection through moving region length breadth ratio realization abnormal behaviour makes monitor message be easy to get, and has also reduced information processing capacity; Can handle the moving region timely and effectively, algorithm is simple.

That is to say that the present invention does not need the accurate tracking movement human; Do not need accurately to extract human body contour outline yet, only need to extract the moving region, make that overall plan is able to simplify; Model Matching detection method operand is little, computational accuracy is high than preestablishing, thereby has improved recognition accuracy.

Realization of the present invention also is: when graphical analysis in carrying out step 2 and behavioural characteristic are extracted, comprise the steps:

Step S1: the method that adopts gaussian filtering method denoising and neighborhood denoising to combine is removed noise.

Step S2: utilize mixed Gauss model to carry out background modeling to image sequence after the denoising;

Step S3: region of variation is extracted from background model, obtain moving target;

Step S4: utilize the HSV model that shade separately and with mixed Gauss model is removed in moving target and motion shadow region.

Denoising is the important step that Flame Image Process and image extract, and the method denoising that the present invention adopts gaussian filtering denoising and neighborhood denoising to combine has obtained good effect, for next step processing lays the foundation.And then with prospect from image with background separation; Obtain moving target, again this moving target is removed shade, each step here all relates to the accuracy rate that abnormal behaviour is judged; The present invention adopts the HSV model to obtain doubtful shade as much as possible; Adopt mixed Gauss model to remove doubtful shade, got accurate effect, further improved recognition effect.

Realization of the present invention also is: in carrying out step 2, utilize mixed Gauss model to carry out background modeling to comprise the steps:

Step S21: video sequence image is divided into W * W little subregion:

Step S22: each little subregion is carried out modeling according to mixed Gauss model.

The present invention is divided into W * W little subregion with video sequence image, and W is carried out different values, with region segmentation, when next update, can only upgrade the zone that those change noticeably, and changes inapparent zone and then needn't upgrade.The present invention carries out subregion to image, has reduced the operand of mixed Gauss model, has improved efficient.

Realization of the present invention also is: the HSV model that utilizes in carrying out step 2 separately and with mixed Gauss model removal shade comprises the steps: moving target and motion shadow region

Step S41: utilize the HSV model to judge that whether current pixel is shade, detects all doubtful shades;

Step S42: the pixel value of doubtful shade is added the parameter learning of mixed Gaussian shadow model, thereby judge whether this doubtful shade is real shade, eliminate these flase drops then.

At first utilize shadow model, judge whether current pixel value possibly be the shadows pixels value based on the hsv color space.Current pixel value possibly be a shade, just can participate in the parameter learning of mixed Gaussian shadow model, thereby judges whether this doubtful shade is real shade.Therefore; The parameter of doubtful shadow model is selected should be suitably looser; Be all real shadow Detection doubtful shade as far as possible; And allow also to be judged to be doubtful shade to some moving target pixel values, to eliminate these flase drops with the mixed Gaussian shadow model then, the prospect that obtains is moving target.The present invention utilizes the shadow model based on the hsv color space, is all real shadow Detection doubtful shade as far as possible, and allows also to be judged to be doubtful shade to some moving target pixel values, eliminates these flase drops with the mixed Gaussian shadow model then.Make that the motion detection ratio of precision is higher.

Because the present invention through the analysis to the moving region, extracts the length breadth ratio of moving region, through a large amount of normal and abnormal motion data are carried out machine learning; Set up the human motion characteristic model, realize the real-time detection of human body abnormal behaviour, in detection; The present invention has organically used the background subtraction point-score to extract the moving region of gathering in the video sequence image; And used gaussian filtering denoising and the denoising of neighborhood denoising realization combining therein, and when adopting mixed Gauss model to carry out background modeling, the background of video sequence image is segmented, reduced the calculated amount of mixed Gauss model; Shorten computation process, improved the efficient and the robustness of detection algorithm.Whether the present invention also utilizes the parameter learning of HSV model with the pixel value adding mixed Gaussian shadow model of doubtful shade, be real shade with this doubtful shade of accurate judgement, and reduced flase drop, has also improved the identification accuracy.It is simple that the improvement of several method and fusion, the present invention have solved modeling effectively, and terse algorithm detects technical matters accurately, realized the movement human abnormal behaviour recognition methods based on template matches that a kind of verification and measurement ratio is higher.The present invention is mainly used in places such as bank, museum, parking lot, airport.

Description of drawings:

Fig. 1 is an operational scheme synoptic diagram of the present invention;

Fig. 2 moving target extraction algorithm of the present invention process flow diagram;

Fig. 3 is the original graph A that comprises moving target;

Fig. 4 is the original graph B that comprises moving target;

Fig. 5 is the foreground extraction figure that comprises the original graph A of moving target;

Fig. 6 is the foreground extraction figure that comprises the original graph B of moving target;

Fig. 7 is criterion behavior collection figure;

Fig. 8 is that the present invention is to abnormal behaviour recognition result figure A;

Fig. 9 is the recognition result figure B of the present invention to abnormal behaviour;

Figure 10 is the recognition result figure C of the present invention to abnormal behaviour;

Figure 11 is the recognition result figure D of the present invention to abnormal behaviour.

Embodiment:

Below in conjunction with accompanying drawing the present invention is elaborated:

Embodiment 1:

Referring to Fig. 1, the present invention is a kind of movement human abnormal behaviour recognition methods based on template matches, is applied to bank, museum etc. more and relates to the public safety problem needs place of monitoring in real time.Utilize the computer vision technique analysis and understand people's motion, and carry out real time image collection, record, discriminating abnormal behaviour and warning, see Fig. 1.The needed hardware minimalist configuration of the inventive method is: P4 3.0G CPU, the computing machine of 512M internal memory; Lowest resolution is 320 * 240 healthy camera or DV video camera; Frame per second is the video frequency collection card and the MD decoding card of 25 frame per seconds.Fig. 1 is an operational scheme synoptic diagram of the present invention, and according to Fig. 1 flow process, detection method of the present invention may further comprise the steps: one. sample video data acquiring: through the camera collection normal behaviour video sequence in the define behavior in advance, as training sample; In this example in advance define behavior be the walking, normal behaviour is referring to Fig. 7.

Two. graphical analysis and behavioural characteristic are extracted: utilize the background subtraction point-score to extract the moving region of gathering in the video sequence image;

Three. measure the length and the width of the moving region that extracts, calculate long and wide ratio, classification and storage is on corresponding view data; Referring to Fig. 8-11.

Four. behavior modeling: utilize length with the wide ratio different learning model buildings of being correlated with of a large amount of motion samples of above-mentioned collection, set up corresponding criterion behavior model by every kind of behavior; What is called is promptly enough judged a certain amount of sample of abnormal behaviour to every kind of behavior collection in a large number.

Five. set up model bank: the criterion behavior category of model that obtains is saved in the model database, makes up the criterion behavior storehouse;

Six. real-time data acquisition: gather the behavior video sequence in real time through camera, and the employing method identical with step 2 carried out feature extraction; Promptly utilize the background subtraction point-score to extract the moving region of gathering in the video image.

Seven. interpretation of result: utilize similarity comparison algorithm based on Weighted H u invariant moments; Predefined normal behaviour relatively according to maximum-likelihood criterion, calculates the behavior to be identified and the similarity of define behavior in advance in tracking target that current frame image is promptly extracted and the criterion behavior database; Compare with setting threshold; Greater than similarity threshold be normal behaviour, less than similarity threshold be abnormal behaviour, determine unusual; Calculate the length breadth ratio of moving region for each criterion behavior, obtain through the mass data statistics.

Eight. follow the tracks of and warning abnormal behaviour is compact.Such as to the walking target following, after keeping watch on out abnormal behaviour, report to the police.

The present invention utilizes the model bank of training; The matching degree of its template satisfies maximum-likelihood criterion, calculates behavior to be identified and the define behavior and the matching degree of define behavior in advance in advance, in this example; Define behavior is to walk in advance; Calculate the length breadth ratio of moving region, obtaining threshold value through the mass data statistics is 0.7, judges abnormal behaviour less than the behavior of threshold value.

The present invention extracts the pairing original image of moving target zone earlier; Calculate Weighted H u invariant moments under its HSV space again as the feature descriptor of human body behavior; And utilize this characteristic to each two field picture in the video sequence and the regular behavior image calculation Euclidean distance that is stored in advance in the database, whether the behavior of differentiating this target is unusual.Therefore, before system detects unusual human body behavior, must a clearly regulation be arranged to normal behaviour.The present invention sets up database to the normal behaviour of moving target, adopts the method for frame template matches to realize that abnormal behaviour detects.Stored the picture of normal behaviour in the database.

In advance with the picture-storage of a series of normal behaviours in database because symmetric relation, in the storehouse image according to direction of travel can be divided into ' left side is to right ' with ' right to a left side big type of ' two, be called " criterion behavior storehouse " or title " regular motion ATL ".At real-time monitor stages; Read video sequence by frame, every two field picture is carried out processing such as gray scale conversion and mathematical morphology, utilize the moving target detecting method in the system to extract target information; Calculate its Weighted H u invariant moment features value then; And mate with normal behaviour picture in the template database, maximum similarity and the preset threshold value of calculating coupling compare, if in threshold value in specialized range then be regarded as " normal behaviour "; Otherwise, then be " abnormal behaviour ".

In above-mentioned human body behavior identifying; A crucial step is exactly current frame image and normal behaviour picture coupling; Be template matches, farthest calculate the similarity of two width of cloth images, through repetition test and contrast; Adopted the similarity comparison algorithm based on Weighted H u invariant moments, experimental result shows that this algorithm has adaptability and precision preferably.

M.K.Hu proposed the definition of continuous function square and about the fundamental property of square in 1962, and specifically provided have translation, the expression formula of 7 invariant moments of rotation, constant rate property, it has good discrimination when the object of identification similarity.

The definition of Hu square is following:

If f (x, y) be certain two dimensional image h (x, y), s (x, y) or v (then its (p+q) rank moment of the orign is defined as for x, y) function:

$m_{\mathrm{Pq}}=\underset{x,y\∈\mathrm{\Ω}}{\∫\∫}{x}^p{y}^{q}f(x,y)\mathrm{Dxdy}$ Formula (8)

In the formula, Ω is x, the interval of y, m _PqConcrete implication be f (x, y) projection on monomial.Obvious m _Pq(x, y) unique definite, vice versa by f.Because m _PqDo not have translation invariance, therefore definition (p+q) rank central moment is:

${\mathrm{\μ}}_{\mathrm{Pq}}=\underset{x,y\∈\mathrm{\Ω}}{\∫\∫}{(x-\stackrel{\‾}{x})}^p{(y-\stackrel{\‾}{y})}^{q}f(x,y)\mathrm{Dxdy}$ Formula (9)

(p+q) rank normalization central moment note is made η _Pq, be defined as:

${\mathrm{\η}}_{\mathrm{Pq}}=\frac{{\mathrm{\μ}}_{\mathrm{Pq}}}{{\mathrm{\μ}}_{00}^{\mathrm{\γ}}}$ $\mathrm{\γ}=\frac{p+q+1}{2}$ (p+q=2,3 ...) formula (10)

Utilize second order and three rank normalization central moment can derive following 7 invariant moments groups:

M ₁＝η ₂₀+η ₀₂

$M_2={({\mathrm{\η}}_{20}-{\mathrm{\η}}_{02})}^2+4{\mathrm{\η}}_{11}^2$

M ₃＝(η ₃₀-3η ₁₂) ²+(3η ₃₁-η ₀₃) ²

M ₄＝(η ₃₀+η ₁₂) ²+(η ₂₁+η ₀₃) ²

M ₅＝(η ₃₀-3η ₁₂)(η ₃₀-η ₁₂)[(η ₃₀+η ₁₂) ²-3(η ₂₁+η ₀₃) ²]+(3η ₂₁-η ₀₃)(η ₂₁+η ₀₃)[3(η ₃₀+η ₁₂) ²-(η ₂₁+η ₀₃) ²]

M ₆＝(η ₂₀-η ₀₂)[(η ₃₀+η ₁₂) ²-(η ₂₁+η ₀₃) ²]+4η ₁₁(η ₃₀+η ₁₂)(η ₂₁+η ₀₃)

M ₇＝(3η ₁₂-η ₃₀)(η ₃₀+η ₁₂)[(η ₃₀+η ₁₂) ²-3(η ₂₁+η ₀₃) ²]

Formula (11)

(3η ₂₁-η ₀₃)(η ₂₁+η ₀₃)[3(η ₀₃+η ₁₂) ²-(η ₁₂+η ₃₀) ²]

The quick invariant moments algorithm that the present invention adopts Hatamian to propose carries out the Moment Feature Extraction of image.The two dimensional image square is decomposed into two one dimension squares, and calculates with recursion method, greatly reduce the multiplication number of times, the addition number of times also reduces than direct calculating to some extent.

Experiment shows when the brightness in two zones differs greatly, and can be easy to it is separated; And when the identical only tone of brightness in two zones not simultaneously, can not cause shape vision clearly.If let the brightness in two zones move closer to, it is fuzzy gradually that their outline line will become, and two region shapes become uncertain gradually.That is to say that brightness, tone, color saturation can both form shape under certain condition, but they are different to the contribution of target shape separately, brightness is maximum, and tone takes second place, and color saturation is taken second place again.Therefore this present invention proposes to utilize the Hu invariant moments of weighting to come the behavioural characteristic of human body is described.The computing formula of its invariant moments is following:

M _i=w _vM _Vi+ w _kM _Ki+ w _sM _Si(i=1,2 ..., 7) and formula (12)

In the formula, M _ViBe brightness Hu invariant moments, M _HiBe tone Hu invariant moments, M _SiBe colour saturation Hu invariant moments, M _iBe human body behavior Weighted H u invariant moments, w _v, w _h, w _sBe the weights of each component, wherein w _v+ w _s+ w _h=1, and w _v＞w _h＞w _s

In this example; The behavior of definition " normally " human body is to walk in advance; Referring to Fig. 7, this walking picture is stored in the criterion behavior storehouse, calculates the Euclidean distance of image in every two field picture of video sequence and the normal behaviour criterion behavior storehouse then; The minimum Eustachian distance of all templates and this image in the outbound, differentiate.If M represents the image of current video sequence, calculate the Euclidean distance between each template s among itself and the normal behaviour storehouse S, its computing formula is following:

$\begin{array}{cc}{d}_{\mathrm{Ms}}=\underset{i=1}{\overset{7}{\mathrm{\Σ}}}{(M_{\mathrm{Mi}}-M_{\mathrm{Si}})}^2& s\∈S\end{array}$ Formula (13)

Obtain minor increment d _Mj=min{d _Ms, s, j ∈ S, when minor increment less than set threshold value, think that then the human body behavior in the present image is a normal behaviour; Otherwise, judge that the human body behavior in this image is an abnormal behaviour.The present invention is to the recognition result of abnormal behaviour such as Fig. 8-shown in Figure 11.

Embodiment 2:

With embodiment 1, referring to Fig. 2, graphical analysis among the present invention in the step 2 and behavioural characteristic are extracted and are comprised the steps: based on the movement human abnormal behaviour recognition methods of template matches

Step S1: the method that adopts gaussian filtering method denoising and neighborhood denoising to combine is removed noise.

Step S2: utilize mixed Gauss model to carry out background modeling to image sequence after the denoising;

Step S3: region of variation is extracted from background model, obtain moving target;

Step S4: utilize the HSV model that shade separately and with mixed Gauss model is removed in moving target and motion shadow region.

The purpose of moving object detection is from sequence image, region of variation segmented extraction from background image to be come out; Because understanding, the post-processed process of image such as target classification, tracking and behavior only consider in the image to be important basic work so moving target is effectively cut apart corresponding to the pixel of motion target area.Yet when detecting moving target, the motion shade of moving target projection also can be detected as the part of sport foreground, causes merging, the geometry deformation of moving target, even makes track rejection.How to obtain better moving target foreground segmentation effect, in detected sport foreground, separate moving target and the motion shade of its projection better, directly have influence on the effect of Target Recognition.

At first utilize shadow model, judge whether current pixel value possibly be the shadows pixels value based on the hsv color space.Current pixel value possibly be a shade, just can participate in the parameter learning of mixed Gaussian shadow model, thereby judges whether this doubtful shade is real shade.Therefore; The parameter of doubtful shadow model is selected should be suitably looser; Be all real shadow Detection doubtful shade as far as possible; And allow also to be judged to be doubtful shade to some moving target pixel values, to eliminate these flase drops with the mixed Gaussian shadow model then, the prospect that obtains is moving target.

In addition, video sequence is obtaining and is transmitting often that regular meeting receives various noise, and for example white Gaussian noise, impulsive noise and multiplicative noise etc. also have in extraneous actual environment simultaneously because leaf a large amount of noise that is brought such as rock.And these noises can be given follow-up processing, and such as rim detection, profile extracts, and target detection is brought very big inconvenience, also influences the visual perception to a certain extent.The present invention has adopted the gaussian filtering method and has faced the method removal noise effect that the territory denoising combines.The original graph A that comprises moving target sees Fig. 3 and shown in Figure 4 with the original graph B that comprises moving target; Adopted the gaussian filtering method and faced method that the territory denoising combines and remove noise effect and target detection is extracted moving target its effect such as Fig. 5 and shown in Figure 6 from background through the present invention.

Embodiment 3:

, utilize mixed Gauss model to carry out background modeling to comprise the steps: among the present invention in the step 2 with embodiment 1-2 based on the movement human abnormal behaviour recognition methods of template matches

Step S21: video sequence image is divided into W * W little subregion:

Step S22: each little subregion is carried out modeling according to mixed Gauss model.

Video sequence image is divided into 3 * 3 little subregions, when next update, can only upgrades the zone that those change noticeably, change inapparent zone and then needn't upgrade.The present invention carries out subregion to image, has reduced the operand of mixed Gauss model, has improved efficient.

The present invention proposes and has adopted improved mixed Gaussian method to detect moving target, adopts the mixed Gaussian method to carry out modeling to each little subregion.The basic thought of mixed Gauss model is: to each pixel, define K state, each state is represented with a Gaussian function, the pixel value of these state part expression backgrounds, and remainder is then represented the pixel value of prospect.If each pixel color value is used variable X _tExpression, its probability density function can use following K three-dimensional Gaussian function to describe:

$f(X_t=x)=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{it}}\×\mathrm{\η}(x,{\mathrm{\μ}}_{\mathrm{it}},{\mathrm{\Σ}}_{\mathrm{it}})---\left(1\right)$

η (x, μ wherein _It, ∑ _It) be t i Gaussian distribution constantly, its average is μ _ItCovariance is a ∑ _It, ω _ItBe i Gaussian distribution in t weight constantly,

I=1,2 ..., k

Wherein, n representes X _tDimension.In order to reduce calculated amount, suppose that generally R, G, B three Color Channels of each pixel are separate, and have identical variance that this just is equivalent to respectively set up an one dimension mixed Gauss model for each Color Channel.To each input pixel value I _tIf, | I _t-μ _{I, t-1}|≤D * σ _{I, t-1}, wherein, μ _{I, t-1}Be average, D is a parameter, σ _{I, t-1}Be standard deviation, then I _tWith this Gaussian function coupling, its parameter is upgraded by following formula

$\left\{\begin{array}{c}{\mathrm{\ω}}_{i,t}=(1-\mathrm{\α}){\mathrm{\ω}}_{i,t-1}+\mathrm{\α}\\ {\mathrm{\μ}}_{i,t}=(1-\mathrm{\ρ}){\mathrm{\μ}}_{i,t-1}+\mathrm{\ρ}{I}_t\\ {\mathrm{\σ}}_{i,t}^2=(1-\mathrm{\ρ}){{\mathrm{\σ}}^2}_{i,t-1}+\mathrm{\ρ}{(I_t-{\mathrm{\μ}}_{i,t})}^2\end{array}\right.---\left(2\right)$

Wherein α is user-defined learning rate; And 0≤α≤1; ρ is the parameter learning rate, and

If do not have Gaussian distribution and I _tCoupling, then the minimum Gaussian distribution of weights will be upgraded by new Gaussian distribution, and the new average that distributes is I _t, bigger standard deviation sigma of initialization and less weights ω.Remaining Gaussian distribution keeps identical average and variance, but their weights can decay, that is:

ω _i，t＝(1-α)ω _i，t-1 (3)

At last, all weights normalization, and press ω to each Gaussian distribution _{I, t}/ σ _{I, t}Arrange from big to small.ω _{I, t}/ σ _{I, t}Big person representes to have less variance and bigger probability of occurrence, and this has just embodied the characteristic of scene background pixel value, because the probability of pixel display background state is more much bigger than the probability that shows arbitrary prospect state usually.If i ₁, i ₂...., i _KBe that each Gaussian distribution is by ω _{I, t}/ σ _{I, t}Descending ordering, if following formula is satisfied in preceding M distribution, then this M distribution is considered to background distributions, that is:

$\underset{K=i}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\ω}}_{k,t}\≥\mathrm{\τ}---\left(4\right)$

Wherein τ is a weight threshold.I _tWith the absolute value of the difference of each background distributions average all doubly greater than the D of this distribution standard deviation, I then _tBe considered to sport foreground, otherwise I _tBe judged to background pixel.

Fig. 3 is moving target A, and Fig. 5 is foreground extraction result of the present invention.Fig. 4 is moving target B, and Fig. 6 is foreground extraction result of the present invention.

Embodiment 4:

With embodiment 1-3, in carrying out step 2, utilize the HSV model separately and remove shade with mixed Gauss model and comprise the steps: based on the abnormal behaviour recognition methods of template matches with moving target and motion shadow region

Step S41: utilize the HSV model to judge that whether current pixel is shade, detects all doubtful shades;

Step S42: the pixel value of doubtful shade is added the parameter learning of mixed Gaussian shadow model, thereby judge whether this doubtful shade is real shade, eliminate these flase drops then.

The purpose of moving object detection is from sequence image, region of variation segmented extraction from background image to be come out; Because understanding, the post-processed process of image such as target classification, tracking and behavior only consider in the image to be important basic work so moving target is effectively cut apart corresponding to the pixel of motion target area.Yet when detecting moving target, the motion shade of moving target projection also can be detected as the part of sport foreground, causes merging, the geometry deformation of moving target, even makes track rejection.How to obtain better moving target foreground segmentation effect, in detected sport foreground, separate moving target and the motion shade of its projection better, directly have influence on the effect of Target Recognition.

Embodiment 5:

With embodiment 3, the application scenario is that bank, parking lot etc. relate to the public safety problem needs place of monitoring in real time based on the abnormal behaviour recognition methods of template matches.Utilize the computer vision technique analysis and understand people's motion, and carry out real time image collection, record, discriminating abnormal behaviour and warning.

The needed hardware minimalist configuration of the inventive method is: P4 3.0G CPU, the computing machine of 512M internal memory; Lowest resolution is 320 * 240 healthy camera or DV video camera; Frame per second is the video frequency collection card and the MD decoding card of 25 frame per seconds.Referring to Fig. 1 is operational flow diagram of the present invention, and the inventive method comprises following concrete steps according to Fig. 1 flow process:

One. sample video data acquiring: adopt the machine learning principle to carry out the modeling and the identification of human body behavior, therefore work such as study of being correlated with of a considerable amount of behavior samples and verification.In this example in advance define behavior be to jog, normal behaviour has been gathered the plurality of sections video sequence referring to 3, and gets wherein a part and learn as training set, a part is carried out model checking as test set, has made up the video data training sample.

Two. graphical analysis and behavioural characteristic are extracted: institute is merged in steps, and concrete leaching process flow process as shown in Figure 2 is carried out, and introduces as follows:

(1) obtains video data: comprise the video of gathering in real time in training video and the detection.

(2) method that adopts gaussian filtering method denoising and neighborhood denoising to combine to the video data that obtains is removed noise.

(3) frame of video after the denoising is divided into 4 * 4 little subregion.

(4) each little subregion is carried out modeling according to mixed Gauss model.

(5) carry out differential ratio to operation and carry out binary conversion treatment by current input image and background model.

(6) utilize the HSV model to detect doubtful shade, utilize mixed Gauss model to remove real shade and obtain the moving region.

Three. measure the length and the width of the moving region that extracts, calculate long and wide ratio, classification and recording storage are on corresponding view data;

Four. behavior modeling: the different of long and wide ratio are pressed in every kind of behavior, set up corresponding criterion behavior model, promptly a certain amount of sample of abnormal behaviour is enough judged in every kind of behavior collection, extract, acquisition criterion behavior model through graphical analysis and behavioural characteristic.

Five. set up model bank: it is in the training pattern storehouse that the criterion behavior category of model that obtains is saved in model database;

Six. real-time data acquisition: gather the behavior video sequence in real time through camera, and the employing method identical with step 2 carried out feature extraction; Promptly utilize the background subtraction point-score to extract the moving region of gathering in the video image;

Seven. interpretation of result: utilize similarity comparison algorithm based on Weighted H u invariant moments; Predefined normal behaviour relatively according to maximum-likelihood criterion, calculates the behavior to be identified and the similarity of define behavior in advance in tracking target that current frame image is promptly extracted and the criterion behavior database; Compare with setting threshold; Greater than similarity threshold be normal behaviour, less than similarity threshold be abnormal behaviour, determine unusual; Calculate the length breadth ratio of moving region for each criterion behavior, obtain through the mass data statistics.

The present invention utilizes the model bank of training, and the matching degree of its template satisfies maximum-likelihood criterion, calculates behavior to be identified and the define behavior and the matching degree of define behavior in advance in advance.

In this example in advance define behavior be to jog as normal behaviour; Hurry up, jump, escape away, sidle, fight all as abnormal behaviour; Calculate the length breadth ratio of moving region, obtaining threshold value through the mass data statistics is 0.75, judges abnormal behaviour less than the behavior of 0.75 threshold value;

Eight. follow the tracks of and warning abnormal behaviour is compact.Such as to the target following of jogging, after keeping watch on out abnormal behaviour, 0.3-0.5 reported to the police after time second.

The present invention is to the recognition result of abnormal behaviour such as Fig. 8-shown in Figure 11.

Embodiment 6:

Based on the abnormal behaviour recognition methods of template matches with embodiment 3

The application scenario is the parking lot, relates to the public safety problem needs place of monitoring in real time.Utilize the computer vision technique analysis and understand people's motion, and carry out real time image collection, record, discriminating abnormal behaviour and warning.

Embodiment 7:

Abnormal behaviour recognition methods based on template matches is banks with embodiment 3 application scenarios, relates to the public safety problem needs place of monitoring in real time.Utilize the computer vision technique analysis and understand people's motion, and carry out real time image collection, record, discriminating abnormal behaviour and warning.Difference is, among the step S21: video sequence image is divided into 10 * 10 little subregions, and 3 * 3 accuracy of detection are better than being divided into, and the sub-district is many more along with video sequence image is cut apart, and corresponding precision increases.

The invention solves prospect background and separate, remove technical barriers such as motion shade, simplification technology.Realized having higher intelligent monitoring ability and detected than the abnormal behaviour of high detection accuracy rate.The efficient and the robustness of detection algorithm have also been improved.

The present invention has high recognition, and complex scene is also had very high recognition accuracy.This technological achievement can be applied to the video monitoring system of all trades and professions; Aspects such as the monitoring of the for example security monitoring of the security monitoring of financial instrument insurance, government bodies, examination hall discipline, border defendance, prison security, community's security protection, even very wide application prospect is all arranged in the national defence field.

Claims (2)

1. A method for identifying abnormal behavior based on template matching, characterized in that: the method comprises the following steps: 1. Sample video data collection: collect video sequences of normal behaviors in predefined behaviors through cameras as training samples; 2. Image analysis and behavioral feature extraction: Utilize the background difference method to extract the motion region in the collected video sequence images; wherein the image analysis and behavioral feature extraction include the following steps: Step S1: The Gaussian filtering method combined with the neighborhood denoising method is used to remove noise. Step S2: performing background modeling on the collected image sequence using a Gaussian mixture model; Said step S2 uses Gaussian mixture model to perform background modeling on the collected image sequence, including the following steps: Step S21: Divide the video sequence image into W×W small partitions: Step S22: modeling each small subregion according to the mixed Gaussian model. Step S3: extracting the changed area from the background model from the image sequence to obtain the moving target; Step S4: Use the HSV model to distinguish the moving object from the moving shadow, and use the Gaussian mixture model to remove the shadow. 3. Measure the length and width of the extracted motion area, calculate the ratio of length and width, and classify and store in the corresponding image data; 4. Behavior modeling: use the large number of motion samples collected above to carry out relevant learning modeling according to the ratio of length and width of each behavior, and establish a corresponding standard behavior model; 5. Establish a model library: classify and save the obtained standard behavior models into the model database to build a standard behavior library; 6. Real-time data collection: collect behavioral video sequences in real time through the camera, and use the same method as step 2 for feature extraction; Seven. Result analysis: Using the similarity comparison algorithm based on weighted Hu invariant moments, compare the tracking target extracted from the current frame image with the pre-defined normal behavior in the standard behavior database, and calculate the target to be identified according to the maximum likelihood criterion. The similarity between the behavior and the predefined behavior is compared with the set threshold. If it is greater than the similarity threshold, it is normal behavior, and if it is less than the similarity threshold, it is abnormal behavior, and the abnormality is judged; 8. Tightly track, mark and call the police for abnormal behavior. 2. The abnormal behavior recognition method based on template matching according to claim 1, characterized in that: said step S4 utilizes the HSV model to distinguish moving objects and moving shadows and removes shadows with a mixed Gaussian model comprising the following steps: Step S41: use the HSV model to determine whether the current pixel is a shadow, and detect all suspected shadows; Step S42: adding the pixel value of the suspected shadow into the parameter learning of the mixed Gaussian shadow model, so as to determine whether the suspected shadow is a real shadow, and then eliminate the real shadow.

Images