CN103745230B - Adaptive abnormal crowd behavior analysis method - Google Patents
Adaptive abnormal crowd behavior analysis method Download PDFInfo
- Publication number
- CN103745230B CN103745230B CN201410016292.8A CN201410016292A CN103745230B CN 103745230 B CN103745230 B CN 103745230B CN 201410016292 A CN201410016292 A CN 201410016292A CN 103745230 B CN103745230 B CN 103745230B
- Authority
- CN
- China
- Prior art keywords
- density
- image
- arteries
- pixel
- particle
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses an adaptive abnormal crowd behavior analysis method, which is used for analyzing crowd behaviors in a video image. The method comprises the following steps of performing streak line calculation on the video image; calculating a streak line flow; detecting abnormal behaviors; performing foreground detection on the video image of abnormal crowd behaviors; performing adaptive crowd density estimation comprising pixel-counting-based density estimation and texture-analysis-based density estimation, and finally dividing estimated density into four density levels, i.e. a low density level, a medium density level, a high density level and an ultrahigh density level, thereby finishing grading the abnormal crowd behaviors. According to the method, the concepts of streak line and streak line flow are introduced to analyze whether a crowd in the video image is abnormal or not; the method has the advantage of detection accuracy; the densities of crowds involved in the abnormal crowd behaviors in different density scenarios are estimated in an adaptive way, and the detected abnormal crowd behaviors are graded by using density estimation results as main characteristics; the method is used for accurately grading the abnormal behaviors (such as mass brawl) in crowded public places, and giving alarms.
Description
Technical field
The present invention relates to the analysis method of the group abnormality behavior of intelligent video monitoring is based in a kind of public safety field,
More particularly to a kind of self adaptation group abnormality behavior analysis method based on arteries and veins line model, belong at machine vision and intelligent information
Reason field.
Background technology
The generation of group abnormality behavior will constitute harm to social public security, and different grades of group abnormality behavior pair
The hazardness that social public security is produced also is not quite similar, and corresponding attention rate and sensitivity are also different.When group abnormality behavior
During generation, for the different brackets of Deviant Behavior, different measures should be also taken.Such as when in the scene that anomalous event occurs
During population density relatively low (or Population is less), it is believed that the attention rate and sensitivity of the event is relatively low;But,
When population density higher (or Population is more) in the scene that anomalous event occurs, corresponding attention rate and sensitivity are just
Should be lifted rapidly, because the group abnormality event may cause very big threat to social public security.Based on such
Understanding, just seems very necessary to being analyzed, being understood and grading forewarning system is carried out to Deviant Behavior based on the group behavior of video.
And at present in video monitoring system both domestic and external, rarely have the product of such maturation.
The grading forewarning system of group abnormality behavior mainly includes the analysis of group abnormality behavioral value and colony on technology is realized
Two parts of Deviant Behavior hierarchical analysis.
First, during understanding video content, needs judge that it is belonging to normally according to video content
The video of behavior or the video of Deviant Behavior.For normal behaviour video, it is not necessary to give special attention, we mainly close
Note those videos comprising Deviant Behavior.This part is actually to carry out Deviant Behavior identification according to video;And in colony
In scene, due to mutually blocking between serious colony, the impact of the not first-class factor of size of individual in population so that colony
Deviant Behavior recognizes inherently one challenging problem.In some current group abnormality Activity recognition methods,
Mainly have based on hidden Markov model (E.Andrade, S.Blunsden, R.Fisher.Modeling Crowd Scenes
For Event Detection [C] .ICPR 2006, pp.175-178), Lagrangian coherent structure (S.Ali, M.Shah.A
Lagrangian Particle Dynamics Approach for Crowd Flow Segmentation and
Stability Analysis [C] .CVPR 2007, pp.1-6), social force model (R.Mehran, A.Oyama,
M.Shah.Abnormal Crowd Behavior Detection using Social Force Model[C].CVPR
2009, pp.935-942), Markov random field (J.Kim, K.Grauman.Observe locally, infer
globally:A space-time MRF for detecting abnormal activities with incremental
Updates [C] .CVPR 2009, pp.2921-2928), chaos invariance (Wu S, Moore B E, Shah M.Chaotic
invariants of lagrangian particle trajectories for anomaly detection in
Crowded scenes [C] .CVPR, 2010, pp.2054-2060) and dynamic characteristic (S.Ali, M.Shah.Human
action recognition in videos using kinematic features and multiple instance
learning[J].IEEE Transactions onPatternAnalysis and Machine Intelligence,
2010,32)(2):288-303.) the anomaly detection method of scheduling theory, these methods show in some indexs compared with
Good performance.However, when video is present, resolution is relatively low, shake, or in video group movement excessive velocities or excessively slowly etc. because
When plain, these methods may examine the Deviant Behavior not measured in corresponding colony's scene.2012, Hassner et al. was published in
Article " Violent flows on CVPRW (international computer vision and pattern recognition seminar):Real-time
Detection ofviolent crowd behavior " propose based on violence stream description (Violence flows
Descriptor video set well adapting to property of the Deviant Behavior recognition methodss) to These characteristics, but recognition accuracy
Have much room for improvement.
Secondly, according to the result of group abnormality behavioral value, for the group abnormality behavior for detecting is carried out immediately to it
Feature extraction, and according to the feature of the reflection group abnormality behavior scale degree for being extracted carrying out to the group abnormality behavior
It is classified and reports to the police.The feature for describing group abnormality behavior scale degree for generally adopting is mainly the group abnormality behavior field
Crowd density feature in scape.If the people for participating in the group abnormality behavior is more, so its security implication to scene around
Degree it is also bigger.Extracting the method for crowd density comparative maturity in colony's scene at present mainly has based on the side of pixels statisticses
Method and the method based on textural characteristics.In the population density statistics based on pixels statisticses, there are many scholars to propose some effects
Fruit preferably method, such as Davies et al. is published in magazine " Electronics&Communication Engineering
Article " Crowd monitoring using image processing " on Journal " proposes to utilize image procossing first
Technology carries out population surveillance, and primary analysis population's quantity carries out crowd movement's estimation.It is published in " IEEE Conference on
Article " On pixel count based crowd on Cybernetics and Intelligent Systems "
Density estimation for visual surveillance " are characterized estimation population density with foreground pixel number,
Simultaneously algorithm realizes projection modification using camera calibration.It is published in " International Conference on
Article " Aviewpoint invariant approach for crowd on Pattern Recognition "
Counting " is using feedforward neural network (Feed-forward neural network) training characteristics amount and prospect number
Relation, and detect crowd density using training pattern.In the population density statistics based on textural characteristics, also there are some scholars to carry
The pretty good method of effect is gone out, such as Marana et al. is published in " IEEE International Symposium on
Article " On the efficacy on Computer Graphics, Image Processing, and Vision "
Oftexture analysis for crowd monitoring " com-parison and analysis population density estimate in four kinds of textures point
Analysis method, then realizes the classification to different densities colony using Bayes Method.It is published in CVPR (world meters in 2008
Calculation machine vision and pattern recognition meeting) on article " Privacy preserving crowd monitoring:Counting
People without people models or tracking " utilize the textural characteristics such as the uniformity, energy, entropy, using height
This process regression analysis is trained to realize that Population is counted.It is published in 2009 ICCV (international computer vision conference)
Article " Bayesian Poisson regression for crowd counting " is right using Bayes-Poisson regression analysis
Textural characteristics are analyzed, and obtain population density estimation.Two class methods respectively have the scope of application, and it is relatively low that the former is primarily adapted for use in density
Colony's scene, the latter is primarily adapted for use in the higher colony's scene of density.But, for the video set of These characteristics, at present still
It is not better able to the method for adaptive analysis different densities colony scene.
The content of the invention
The purpose of the present invention is that and a kind of self adaptation colony based on arteries and veins line model is provided to solve the above problems
Deviant Behavior analysis method.
The present invention is achieved through the following technical solutions above-mentioned purpose:
A kind of self adaptation group abnormality behavior analysis method, for being analyzed to the group behavior in video image, bag
Include following steps:
(1) arteries and veins line computation is carried out to the video image:
Arteries and veins line is defined:Hypothesis has a particle a at p points, and according to light stream field direction, particle flux moves every time a step
Long, at the next step-length moment, p points are initialized by new particle b again, and then, two particles of a and b continue to be moved with flow direction,
Repeat this process, in time interval tsThe a number of particle position at p points is inside just obtained, it is described a number of
The line of particle position is arteries and veins line;
IfFor initial point p the frame of t i-th a particle position, i, t=0,1,2 ... ts, according to
Light stream direction, improved stream repeats initialization p points, and arteries and veins line computation formula is as follows:
Wherein, u, v are respectively optical flow velocity vector field;To all of i, t=0,1,2 ... ts, using quadravalence Long Ge-storehouse
Tower equation makees particle advection to formula (1);Each particle in convection cell, we define one comprising particle position and initial velocity
Extra particles PiCorrespond to therewith:
Pi={ xi(t),yi(t),ui,vi} (2)
Wherein,
(2) arteries and veins line stream calculation:
The definition of arteries and veins line stream is:Ωs=(us,vs)T, wherein, T represents transposition, us、vsArteries and veins line stream velocity is represented respectively
;
If U=is [ci], A=[a1,a2,a3], wherein, ciIPi, to arbitrary i, p, all pixels are calculated in the x direction
Arteries and veins line stream, can learn that particle has sub-pixel level precision according to formula (1), to each particle, need calculate its closest three
The analog value of individual pixel, in calculating process, using the linear interpolation of three vicinity points c is obtainedi, formula is defined as follows:
ci=a1us(k1)+a2us(k2)+a3us(k3) (3)
Wherein, kjFor the label of vicinity points, j=1,2,3, ajFor the known Based on Triangle Basis of j-th neighborhood pixels,
According to the particle and its adjacent three pixel, u can be asked for by trigonometric interpolation formulas(kj), to all data in U
Point, using formula (3), just defines following one group system of linear equations:
Aus=U (4)
Equation group (4) is solved using method of least square, obtains us, vsSolution procedure is similar to;Again by Ωs=(us,
vs)TObtain arteries and veins line stream Ωs;
(3) unusual checking:Classified using the support vector machine based on Radial basis kernel function, detected video figure
Normal population behavior and Anomaly groups behavior as in.
Further, in order to realize the accurate monitoring to Anomaly groups behavior and warning, the analysis method also includes abnormal
Group behavior is classified, and its step is described as follows:
1. foreground detection:The video image of the Anomaly groups behavior detected to the step (3) is processed, and obtains group
Body sport foreground;
2. self adaptation population density is estimated:After obtaining group movement prospect, whole image area is accounted for according to foreground area
Whether good more than predefined actual ratio proportion threshold value carry out density estimation being adaptive selected different methods:When this
When actual ratio is less than proportion threshold value, show that now foreground target area is less, population density is relatively low, then using following A method
To its density Further Division;Conversely, showing that population density now is higher, its density is further drawn using following B methods
Point;
A, the density estimation algorithm counted based on pixel:Step is as follows:
A, using canny edge detection operators (JohnCanny in 1986 propose Canny edge detection operators, belonging to is
The method differentiated afterwards is first smoothed, is a kind of conventional universal method) edge graph is obtained to adjacent two frames gray level image extraction edge
Picture;
B, interference noise is eliminated by morphological dilations and etching operation;
C, to two frame border image differences, and it is carried out and operation with the edge image of present frame, obtain foreground edge
Image;
D, the total pixel number that edge is calculated according to foreground edge image;
E, fitted using the method for linear fit crowd's number and crowd's foreground edge total pixel number relation, estimate
Go out general population density, and be divided into low and medium two density ratings;
B, the density estimation method based on texture analysiss:
First, gray level co-occurrence matrixes are generated:Grey level histogram is that the single pixel with certain gray value on image is entered
The result of row statistics, gray level co-occurrence matrixes are that respectively the situation with certain gray scale is entered to keeping two pixels of a certain distance on image
Row statistics is obtained, if f (x, y) is a width two-dimensional digital image, its size is M × N, and M, N distinguishes the width and height of representative image
Degree, gradation of image rank is L, then the gray level co-occurrence matrixes for meeting certain space relation are:
P (m, n)=# { (x1,y1),(x2,y2)∈M×N|f(x1,y1)=m, f (x2,y2)=n } (5)
Wherein, # represents the element number in set, and m and n represents image intensity value, as can be seen that P is L from formula (5)
The matrix of × L, if (x1,y1) and (x2,y2) distance be d, both angles with abscissa line are q, then can obtain each inter-species
Away from and angle gray level co-occurrence matrixes P (m, n | d, θ), it represents that the gray value of a pixel in image is m, one other pixel
Gray value is n, and neighbor distance is d, the number of times that direction occurs for the two such pixel of θ, θ=0 °, 45 °, 90 °, 135 °,
D takes 1;
Then, the gray level co-occurrence matrixes to generating carry out following feature extraction:
Entropy
Energy
Contrast
Uniformity
Dependency
Wherein:
It is a characteristic vector by the latent structure of taking-up mentioned above, and by this characteristic vector input linear supporting vector
Machine, estimates the crowd density in high-density scene, and is divided into two density ratings such as high and superelevation, so as to complete exception
Group behavior is classified.
Specifically, the step 1. in, process video image using three conventional frame frame difference methods.Three frame frame difference methods are also referred to as
Three-frame differencing, is a kind of improved method of adjacent two frame differences algorithm.The ultimate principle of the algorithm is first to select video figure
The difference image of adjacent two frame is calculated as continuous three two field picture in sequence and respectively, it is then that difference image is appropriate by choosing
Threshold value carries out binary conversion treatment, obtains binary image;Finally bianry image is carried out into logic and operation, obtains common ground,
So as to obtain moving target foreground information.
In the step (3), the method for the unusual checking is:
To the every frame video being input into, using below equation (11) arteries and veins line stream amplitude R of each pixel is calculatedx,y,t:
In formula,Respectively it is calculated and u in the step (2)s、vsCorresponding t i-th
It is located in frameParticle arteries and veins line flow valuve;
Then to the result of calculation of formula (11), as follows (12) carry out binaryzation:
In formula, θ is according in adjacent two frames video | Rx,y,t-Rx,y,t-1| meansigma methodss adaptive setting a threshold value;
On the basis of the calculating of formula (12), by the B for adding up all frames in each pixelx,y,t, then ask again flat
Average, sequence of frames of video will obtain an arteries and veins line stream amplitude mean change amount on corresponding pixelIt is expressed as follows:
Wherein, N represents frame of video quantity;
Then based on the result of calculation of formula (13), carried out point using the support vector machine based on Radial basis kernel function
Class, detects the normal population behavior and Anomaly groups behavior in video image.
The beneficial effects of the present invention is:
Present invention introduces whether the colony that the concept of arteries and veins line and arteries and veins line stream is come in analysis video image is abnormal, it is accurate with detection
True advantage, and adaptively realize that the crowd included to the group abnormality behavior in different densities scene enters line density and estimates
Meter, then the group abnormality behavior for detecting is classified by principal character of density estimation result, for accurate classifying alarm,
Public place Deviant Behavior (such as crowd fighting) monitoring and enforcement concentrated for crowd is reported to the police, process has important effect.
Description of the drawings
Fig. 1 is the overall flow figure of self adaptation group abnormality behavior analysis method of the present invention.
Specific embodiment
Below in conjunction with the accompanying drawings the invention will be further described:
As shown in figure 1, the flow process in figure is the main process summary of following methods, using following steps to being used to monitor group
The video image of body is analyzed:
(1) arteries and veins line computation is carried out to the video image:
Arteries and veins line is defined:Hypothesis has a particle a at p points, and according to light stream field direction, particle flux moves every time a step
Long, at the next step-length moment, p points are initialized by new particle b again, and then, two particles of a and b continue to be moved with flow direction,
Repeat this process, in time interval tsThe a number of particle position at p points is inside just obtained, it is described a number of
The line of particle position is arteries and veins line;
IfFor initial point p the frame of t i-th a particle position, i, t=0,1,2 ... ts, according to
Light stream direction, improved stream repeats initialization p points, and arteries and veins line computation formula is as follows:
Wherein, u, v are respectively optical flow velocity vector field;To all of i, t=0,1,2 ... ts, using quadravalence Long Ge-storehouse
Tower equation makees particle advection to formula (1);Each particle in convection cell, we define one comprising particle position and initial velocity
Extra particles PiCorrespond to therewith:
Pi={ xi(t),yi(t),ui,vi} (2)
Wherein,
(2) arteries and veins line stream calculation:
The definition of arteries and veins line stream is:Ωs=(us,vs)T, wherein, T represents transposition, us、vsArteries and veins line stream velocity is represented respectively
;
If U=is [ci], A=[a1,a2,a3], wherein, ciIPi, to arbitrary i, p, all pixels are calculated in the x direction
Arteries and veins line stream, can learn that particle has sub-pixel level precision according to formula (1), to each particle, need calculate its closest three
The analog value of individual pixel, in calculating process, using the linear interpolation of three vicinity points c is obtainedi, formula is defined as follows:
ci=a1us(k1)+a2us(k2)+a3us(k3) (3)
Wherein, kjFor the label of vicinity points, j=1,2,3, ajFor the known Based on Triangle Basis of j-th neighborhood pixels,
According to the particle and its adjacent three pixel, u can be asked for by trigonometric interpolation formulas(kj), to all data in U
Point, using formula (3), just defines following one group system of linear equations:
Aus=U (4)
Equation group (4) is solved using method of least square, obtains us, vsSolution procedure is similar to;Again by Ωs=(us,
vs)TObtain arteries and veins line stream Ωs。
(3) unusual checking:
To the every frame video being input into, using below equation (11) arteries and veins line stream amplitude R of each pixel is calculatedx,y,t:
In formula,Respectively it is calculated and u in the step (2)s、vsCorresponding t i-th
It is located in frameParticle arteries and veins line flow valuve;
Then to the result of calculation of formula (11), as follows (12) carry out binaryzation:
In formula, θ is according in adjacent two frames video | Rx,y,t-Rx,y,t-1| meansigma methodss adaptive setting a threshold value;
On the basis of the calculating of formula (12), by the B for adding up all frames in each pixelx,y,t, then ask again flat
Average, sequence of frames of video will obtain an arteries and veins line stream amplitude mean change amount on corresponding pixelIt is expressed as follows:
Wherein, N represents frame of video quantity;
Then based on the result of calculation of formula (13), carried out point using the support vector machine based on Radial basis kernel function
Class, detects the normal population behavior and Anomaly groups behavior in video image.
(4) foreground detection:The video of the step (3) is detected Anomaly groups behavior using conventional three frame frame difference methods
Image is processed, and obtains group movement prospect.
(5) self adaptation population density is estimated:After obtaining group movement prospect, whole image area is accounted for according to foreground area
Whether good more than predefined proportion threshold value ξ actual ratio σ carry out density estimation being adaptive selected different methods:When
σ<During ξ, show that now foreground target area is less, population density is relatively low, then its density is further drawn using following A method
Point;Conversely, as σ >=ξ, show that population density now is higher, using following B methods to its density Further Division;
A, the density estimation algorithm counted based on pixel:Step is as follows:
A, using canny edge detection operators to adjacent two frames gray level image extract edge obtain edge image;
B, interference noise is eliminated by morphological dilations and etching operation;
C, to two frame border image differences, and it is carried out and operation with the edge image of present frame, obtain foreground edge
Image;
D, the total pixel number that edge is calculated according to foreground edge image;
E, fitted using the method for linear fit crowd's number and crowd's foreground edge total pixel number relation, estimate
Go out general population density, and be divided into low and medium two density ratings;
B, the density estimation method based on texture analysiss:
First, gray level co-occurrence matrixes are generated:Grey level histogram is that the single pixel with certain gray value on image is entered
The result of row statistics, gray level co-occurrence matrixes are that respectively the situation with certain gray scale is entered to keeping two pixels of a certain distance on image
Row statistics is obtained, if f (x, y) is a width two-dimensional digital image, its size is M × N, and M, N distinguishes the width and height of representative image
Degree, gradation of image rank is L, then the gray level co-occurrence matrixes for meeting certain space relation are:
P (m, n)=# { (x1,y1),(x2,y2)∈M×N|f(x1,y1)=m, f (x2,y2)=n } (5)
Wherein, # represents the element number in set, and m and n represents image intensity value, as can be seen that P is L from formula (5)
The matrix of × L, if (x1,y1) and (x2,y2) distance be d, both angles with abscissa line are q, then can obtain each inter-species
Away from and angle gray level co-occurrence matrixes P (m, n | d, θ), it represents that the gray value of a pixel in image is m, one other pixel
Gray value is n, and neighbor distance is d, the number of times that direction occurs for the two such pixel of θ, θ=0 °, 45 °, 90 °, 135 °,
D takes 1;
Then, the gray level co-occurrence matrixes to generating carry out following feature extraction:
Entropy
Energy
Contrast
Uniformity
Dependency
Wherein:
It is a characteristic vector by the latent structure of taking-up mentioned above, and by this characteristic vector input linear supporting vector
Machine, estimates the crowd density in high-density scene, and is divided into two density ratings such as high and superelevation, so as to complete exception
Group behavior is classified.
In order to verify the accuracy and effectiveness of adaptive Deviant Behavior analysis method proposed by the present invention, below by
Experiment carries out detailed com-parison and analysis:
Experiment includes unusual checking result analysis of the accuracy and Deviant Behavior classification analysis of the accuracy two parts.Experiment
Video takes from video set http://www.openu.ac.il/home/hassner/data/violentflows, for convenience of table
Show, we are named as ViF data bases.The data base includes multiple normal and anomalous videos from YouTube.Video
Scene is extremely enriched, and video quality is uneven, and there is Population in scene have few more.Due to the present invention's it is important that how will
Deviant Behavior is divided into different grades according to population density, so selecting to meet table 1 below requirement first from ViF data bases
Typical Deviant Behavior video amount to 93 anomalous video collection as us.Wherein, low, medium, high, superelevation isodensity
Number of videos is respectively 22,21,25 and 25.During using carrying out population density estimation based on the method for texture analysiss,
We select 6 videos as our training video from the videos such as high and superelevation respectively, and remaining 19 video is used as me
Test video;Meanwhile, selecting from the data base when 25 normal videos are classified as our Deviant Behavior recognizers makes
With.
Classification accuracy is as described in Table 2.From Table 2, it can be seen that for the Deviant Behavior video set of different densities,
Preferable result can be obtained using analysis method proposed by the invention, while experimental result also demonstrates analysis side of the present invention
The effectiveness and accuracy of method.
The crowd density tier definition of table 1
The Deviant Behavior grade recognition accuracy of table 2
Above-described embodiment is presently preferred embodiments of the present invention, is not the restriction to technical solution of the present invention, such as:This
Inventing the analysis method can be only applied to group abnormality behavioral value, it is also possible to be applied to other with colony of the present invention
The group abnormality behavior analysiss of feature similarity.As long as can realize on the basis of above-described embodiment without creative work
Technical scheme, in the rights protection scope for being regarded as falling into patent of the present invention.
Claims (1)
1. a kind of self adaptation group abnormality behavior analysis method, for being analyzed to the group behavior in video image, it is special
Levy and be:Comprise the following steps:
(1) arteries and veins line computation is carried out to the video image:
Arteries and veins line is defined:Hypothesis has a particle a at p points, and according to light stream field direction, particle flux moves every time a step-length,
Next step-length moment, p points are initialized by new particle b again, and then, two particles of a and b continue to be moved with flow direction, repeat
This process, in time interval tsInside just obtain a number of particle position at p points, a number of particle
The line of position is arteries and veins line;
IfFor initial point p the frame of t i-th a particle position, i, t=0,1,2 ... ts, according to light stream
Direction, improved stream repeats initialization p points, and arteries and veins line computation formula is as follows:
Wherein, u, v are respectively optical flow velocity vector field;To all of i, t=0,1,2 ... ts, using quadravalence Runge-Kutta equation
Particle advection is made to formula (1);Each particle in convection cell, we define an extra grain comprising particle position and initial velocity
Sub- PiCorrespond to therewith:
Pi={ xi(t),yi(t),ui,vi} (2)
Wherein,
(2) arteries and veins line stream calculation:
The definition of arteries and veins line stream is:Ωs=(us,vs)T, wherein, T represents transposition, us、vsArteries and veins line stream velocity vector field is represented respectively;
If U=is [ci], A=[a1,a2,a3], wherein, ciIPi, to arbitrary i, p, the arteries and veins line of all pixels is calculated in the x direction
Stream, can learn that particle has sub-pixel level precision according to formula (1), to each particle, need to calculate its closest three picture
The analog value of vegetarian refreshments, in calculating process, using the linear interpolation of three vicinity points c is obtainedi, formula is defined as follows:
ci=a1us(k1)+a2us(k2)+a3us(k3) (3)
Wherein, kjFor the label of vicinity points, j=1,2,3, ajFor the known Based on Triangle Basis of j-th neighborhood pixels, according to
The particle and its adjacent three pixel, can ask for u by trigonometric interpolation formulas(kj), to all data points in U, adopt
With formula (3), following one group system of linear equations is just defined:
Aus=U (4)
Equation group (4) is solved using method of least square, obtains us, vsSolution procedure is similar to;Again by Ωs=(us,vs)T
To arteries and veins line stream Ωs;
(3) unusual checking:Classified using the support vector machine based on Radial basis kernel function, in detecting video image
Normal population behavior and Anomaly groups behavior;
(4) Anomaly groups behavior classification, specifically includes following steps:
1. foreground detection:The video image of the Anomaly groups behavior detected to the step (3) is carried out using three frame frame difference methods
Process, obtain group movement prospect;
2. self adaptation population density is estimated:After obtaining group movement prospect, the reality of whole image area is accounted for according to foreground area
Whether good more than predefined ratio proportion threshold value carry out density estimation being adaptive selected different methods:When the reality
When ratio is less than proportion threshold value, show that now foreground target area is less, population density is relatively low, then using following A method to it
Density Further Division;Conversely, show that population density now is higher, using following B methods to its density Further Division;
A, the density estimation algorithm counted based on pixel:Step is as follows:
A, using canny edge detection operators to adjacent two frames gray level image extract edge obtain edge image;
B, interference noise is eliminated by morphological dilations and etching operation;
C, to two frame border image differences, and it is carried out and operation with the edge image of present frame, obtain foreground edge image;
D, the total pixel number that edge is calculated according to foreground edge image;
E, fitted using the method for linear fit crowd's number and crowd's foreground edge total pixel number relation, estimate big
General population density, and it is divided into low and medium two density ratings;
B, the density estimation method based on texture analysiss:
First, gray level co-occurrence matrixes are generated:Grey level histogram is that the single pixel with certain gray value on image is united
The result of meter, gray level co-occurrence matrixes are that respectively the situation with certain gray scale is united to keeping two pixels of a certain distance on image
Meter is obtained, if f (x, y) is a width two-dimensional digital image, its size is M × N, and the width and height of M, N difference representative image are schemed
As grey level is L, then the gray level co-occurrence matrixes for meeting certain space relation are:
P (m, n)=# { (x1,y1),(x2,y2)∈M×N|f(x1,y1)=m, f (x2,y2)=n } (5)
Wherein, # represents the element number in set, and m and n represents image intensity value, as can be seen that P is L × L's from formula (5)
Matrix, if (x1,y1) and (x2,y2) distance be d, both angles with abscissa line are q, then can obtain various spacing and
The gray level co-occurrence matrixes P (m, n | d, θ) of angle, it represents that the gray value of a pixel in image is m, the gray scale of one other pixel
It is worth for n, and neighbor distance is d, the number of times that direction occurs for the two such pixel of θ, θ=0 °, and 45 °, 90 °, 135 °, d takes
1;
Then, the gray level co-occurrence matrixes to generating carry out following feature extraction:
Entropy
Energy
Contrast
Uniformity
Dependency
Wherein:
It is a characteristic vector by the latent structure of taking-up mentioned above, and by this characteristic vector input linear support vector machine, estimates
The crowd density in high-density scene is counted out, and is divided into two density ratings such as high and superelevation, so as to complete Anomaly groups
Behavior is classified;
In the step (3), the method for the unusual checking is:
To the every frame video being input into, using below equation (11) arteries and veins line stream amplitude R of each pixel is calculatedx,y,t:
In formula,Respectively it is calculated and u in the step (2)s、vsIn the frame of corresponding t i-th
It is located atParticle arteries and veins line flow valuve;
Then to the result of calculation of formula (11), as follows (12) carry out binaryzation:
In formula, θ is according in adjacent two frames videoMeansigma methodss adaptive setting a threshold value;
On the basis of the calculating of formula (12), by the B for adding up all frames in each pixelx,y,t, then average again,
Sequence of frames of video will obtain an arteries and veins line stream amplitude mean change amount on corresponding pixelIt is expressed as follows:
Wherein, N represents frame of video quantity;
Then based on the result of calculation of formula (13), classified using the support vector machine based on Radial basis kernel function,
Detect the normal population behavior and Anomaly groups behavior in video image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410016292.8A CN103745230B (en) | 2014-01-14 | 2014-01-14 | Adaptive abnormal crowd behavior analysis method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410016292.8A CN103745230B (en) | 2014-01-14 | 2014-01-14 | Adaptive abnormal crowd behavior analysis method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103745230A CN103745230A (en) | 2014-04-23 |
CN103745230B true CN103745230B (en) | 2017-05-10 |
Family
ID=50502247
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410016292.8A Active CN103745230B (en) | 2014-01-14 | 2014-01-14 | Adaptive abnormal crowd behavior analysis method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103745230B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109766834A (en) * | 2019-01-10 | 2019-05-17 | 成都师范学院 | A kind of crowd evacuation method based on particle flux |
Families Citing this family (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105809092A (en) * | 2014-12-30 | 2016-07-27 | 深圳中兴力维技术有限公司 | Population target detection method and device thereof |
CN106157326B (en) * | 2015-04-07 | 2020-10-27 | 中国科学院深圳先进技术研究院 | Group abnormal behavior detection method and system |
CN106157325B (en) * | 2015-04-07 | 2020-10-27 | 中国科学院深圳先进技术研究院 | Group abnormal behavior detection method and system |
CN104820824A (en) * | 2015-04-23 | 2015-08-05 | 南京邮电大学 | Local abnormal behavior detection method based on optical flow and space-time gradient |
CN105389567B (en) * | 2015-11-16 | 2019-01-25 | 上海交通大学 | Group abnormality detection method based on dense optical flow histogram |
CN106874885B (en) * | 2017-03-03 | 2020-10-16 | 燕山大学 | Crowd abnormity detection method based on energy level distribution change |
US10972860B2 (en) | 2017-05-25 | 2021-04-06 | International Business Machines Corporation | Responding to changes in social traffic in a geofenced area |
CN107330372B (en) * | 2017-06-05 | 2021-05-28 | 四川大学 | Analysis method of video-based crowd density and abnormal behavior detection system |
CN108563991A (en) * | 2018-03-08 | 2018-09-21 | 佛山市云米电器科技有限公司 | Kitchen fume concentration division methods and oil smoke concentration detection and interference elimination method |
CN108960290A (en) * | 2018-06-08 | 2018-12-07 | Oppo广东移动通信有限公司 | Image processing method, device, computer readable storage medium and electronic equipment |
CN108921086A (en) * | 2018-06-29 | 2018-11-30 | Oppo广东移动通信有限公司 | Image processing method and device, storage medium, electronic equipment |
CN109215269A (en) * | 2018-09-18 | 2019-01-15 | 长春众鼎科技有限公司 | Intelligent bus platform with video alarm function |
CN110009490A (en) * | 2019-01-14 | 2019-07-12 | 阿里巴巴集团控股有限公司 | Abnormal financial transaction Stock discrimination method and device |
CN110231624B (en) * | 2019-06-06 | 2021-08-10 | Oppo广东移动通信有限公司 | Object detection method and related product |
CN110427894A (en) * | 2019-08-06 | 2019-11-08 | 齐鲁工业大学 | A kind of crowd's abnormal motion monitoring method, system, equipment and medium |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102800106A (en) * | 2012-06-29 | 2012-11-28 | 刘怡光 | Self-adaptation mean-shift target tracking method based on optical flow field estimation |
-
2014
- 2014-01-14 CN CN201410016292.8A patent/CN103745230B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102800106A (en) * | 2012-06-29 | 2012-11-28 | 刘怡光 | Self-adaptation mean-shift target tracking method based on optical flow field estimation |
Non-Patent Citations (4)
Title |
---|
A Streakline Representation of Flow in Crowded Scenes;Ramin Mehran et al.;《Computer Vision - ECCV 2010》;20101231;第2页第3段,第4页倒数第2段-第6页倒数第一段,第10页倒数第二段 * |
基于帧间差的区域光流分析及其应用;李超 等;《计算机工程与应用》;20051231;195-222 * |
适应于复杂背景的前景检测算法的研究与实现;王晶;《中国优秀硕士学位论文全文数据库 信息科技辑》;20120515;I138-1458,第45页倒数第2段 * |
重点视频监控区域的人群密度估计改进算法的研究与实现;陈鸿;《中国优秀硕士学位论文全文数据库 信息科技辑》;20130615;I138-1451 第32页第1段,第34-38页第4.2.2、4.2.3、4.2.4节,第50页第2段 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109766834A (en) * | 2019-01-10 | 2019-05-17 | 成都师范学院 | A kind of crowd evacuation method based on particle flux |
CN109766834B (en) * | 2019-01-10 | 2021-07-27 | 成都师范学院 | Crowd evacuation method based on particle flow |
Also Published As
Publication number | Publication date |
---|---|
CN103745230A (en) | 2014-04-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103745230B (en) | Adaptive abnormal crowd behavior analysis method | |
CN107330372B (en) | Analysis method of video-based crowd density and abnormal behavior detection system | |
CN108230364B (en) | Foreground object motion state analysis method based on neural network | |
Andrade et al. | Hidden markov models for optical flow analysis in crowds | |
Swathi et al. | Crowd behavior analysis: A survey | |
CN105023019B (en) | A kind of character description method detecting crowd's abnormal behaviour automatically for video monitoring | |
Wang et al. | Crowd anomaly detection for automated video surveillance | |
de Almeida et al. | Change detection in human crowds | |
Kaltsa et al. | Swarm-based motion features for anomaly detection in crowds | |
Yang et al. | Abnormal crowd behavior detection based on local pressure model | |
Hu et al. | Anomaly detection based on local nearest neighbor distance descriptor in crowded scenes | |
Yen et al. | Abnormal event detection using HOSF | |
Xia et al. | Vision-based traffic accident detection using matrix approximation | |
Nemade et al. | Comparative performance analysis of optical flow algorithms for anomaly detection | |
Singh et al. | Crowd escape event detection via pooling features of optical flow for intelligent video surveillance systems | |
CN106951827A (en) | A kind of global abnormal behavioral value method based on target kinetic characteristic | |
Leyva et al. | Video anomaly detection based on wake motion descriptors and perspective grids | |
George et al. | Crowd panic detection using autoencoder with non-uniform feature extraction | |
Nallaivarothayan et al. | Anomalous event detection using a semi-two dimensional hidden markov model | |
Wang et al. | A classification method based on streak flow for abnormal crowd behaviors | |
Lee et al. | Hostile intent and behaviour detection in elevators | |
Esan et al. | Anomalous detection in noisy image frames using cooperative median filtering and KNN | |
Xu et al. | Unusual scene detection using distributed behaviour model and sparse representation | |
Lamba et al. | A large scale crowd density classification using spatio-temporal local binary pattern | |
CN106874885B (en) | Crowd abnormity detection method based on energy level distribution change |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |