CN107491749A - Global and local anomaly detection method in a kind of crowd's scene - Google Patents
Global and local anomaly detection method in a kind of crowd's scene Download PDFInfo
- Publication number
- CN107491749A CN107491749A CN201710684991.3A CN201710684991A CN107491749A CN 107491749 A CN107491749 A CN 107491749A CN 201710684991 A CN201710684991 A CN 201710684991A CN 107491749 A CN107491749 A CN 107491749A
- Authority
- CN
- China
- Prior art keywords
- crowd
- global
- local anomaly
- src
- light stream
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses global and local anomaly detection method in a kind of crowd's scene.First, the present invention proposes a kind of feature of new entitled mixing light stream histogram, then, it is proposed that a kind of method of double rarefaction representations solves this problem for the global abnormal behavior present invention, finally for local anomaly behavior, by first detecting the prospect of present frame area-of-interest, the detection of local anomaly behavior is then carried out using the method for online weighted cluster.The advantages of experimental verification in UMN data sets and UCSD data sets the inventive method.Test result indicates that our method has higher precision compared with the motor behavior of the crowd in video is analyzed of method before.
Description
Technical field
The invention belongs to global and local abnormal behaviour in technical field of image processing, more particularly to a kind of crowd's scene to examine
Survey method.
Background technology
Video monitoring equipment is widely used to the public places such as traffic intersection, shop, bank, subway station and school,
To ensure social safety.Situation about can be easily recorded in by video monitoring in the video of camera surveillance.It is however, right
In most of existing monitoring systems, it is impossible to check anomalous event automatically, and a people can not possibly monitor at any time
Monitor.
Because monitoring device is commonly installed in public places, therefore crowd behaviour analysis is always computer vision research neck
One new focus in domain.Due to eclipse phenomena, the classical way of individual behavior analysis cannot be used for crowded scene.Congested area
Video monitoring more and more important effect is played to public safety.The present invention proposes a kind of gathering around based on crowd movement's feature
The abnormality detection new method squeezed in scene.
In daily life, most of behaviors that we run into, as pedestrian walk on pavement or station subway station people,
All it is normal.However, some behaviors, as people to all the winds scatter suddenly, an automobile driven over the speed limit appears in people
On trade, or people fight, the difference being frequently seen with us, therefore they are considered as abnormal behavior.Generally speaking, it is different
Chang Hangwei seldom occurs in different forms.Therefore, unusual checking is exactly that these exceptions are identified from normal event
Event, this is two classification problems.This problem is divided into global abnormal behavioral value (GAA) and local unusual checking
(LAA).For global abnormal behavioral value, abnormal behaviour or normal behaviour are only existed in the synchronization of Same Scene.For
Local anomaly behavioral value, normal behaviour and abnormal behaviour occur in one scenario simultaneously.
The content of the invention
The technical problems to be solved by the invention are overcome the deficiencies in the prior art, propose global in a kind of crowd's scene
With local anomaly detection method, solve to ignore the information gap between regional in existing detection method, for dividing
The problem of flase drop be present in scattered abnormal behaviour.
It is of the invention specifically to solve above-mentioned technical problem using following technical scheme：
Global and local anomaly detection method in a kind of crowd's scene, is specifically comprised the steps of：
Step 1, crowd's scene image to be detected is inputted；
Step 2, the mixing light stream histogram of crowd's scene image to be detected is extracted；
Step 3, global abnormal behavioral value is carried out to the mixing light stream histogram of step 2 feature extraction, it is specific as follows：
Step 3.1, it is whether abnormal that the image after judging characteristic extraction is distinguished by two rarefaction representation processes；
Step 3.2, handled, and then obtained a result with judgement of the fuzzy integral to two rarefaction representation processes；
Step 4, local anomaly behavioral value is carried out to the mixing light stream histogram of step 2 feature extraction, it is specific as follows：
Step 4.1, the prospect of area-of-interest in input picture is extracted；
Step 4.2, local anomaly detection is carried out to area-of-interest by online weighted cluster；
Step 4.3, noise is filtered out using the method for multiple target tracking；
Step 5, global and local unusual checking in more crowd's scenes is completed by step 3 and step 4.
As the further preferred scheme of global and local anomaly detection method in a kind of crowd's scene of the present invention,
In step 2, light stream histogram H is mixedGIt is specific to represent as follows：
HG=(μ, δ)
Wherein,μ represents the expectation of light stream, and δ is represented in the side of equidirectional
Difference, there are r to continue frame, H in empty container when r is representediRepresent i-th of light stream histogram.
As the further preferred scheme of global and local anomaly detection method in a kind of crowd's scene of the present invention,
In step 3.1, described two rarefaction representation processes include the process of a dynamic lexicon renewal, wherein, dynamic lexicon renewal
The expression formula of process is as follows：
Wherein, the SRC of all elements is expressed as SRC in dictionary Φdic=[SRC1 dic,SRC2 dic,...,SRCN dic]T,
The SRC of next sample is represented by SRCy, m* is the sample index of dictionary, and it reaches next sample y distance minimum, ω
M* be it is new will the m* sample in substitution dictionary Φ, bm* is the m* sample in old dictionary；
As the further preferred scheme of global and local anomaly detection method in a kind of crowd's scene of the present invention, two
The formula of the rarefaction representation of individual rarefaction representation process is as follows：
Wherein, SRC is sparse reconstruct cost, and x represents the coefficient of feature weight, and y is characterized the table after being multiplied by weight coefficient
Up to formula, λ is parameter, and Φ is dictionary, | | | |2For Euclidean distance, | | x | |1For L1Norm.
As the further preferred scheme of global and local anomaly detection method in a kind of crowd's scene of the present invention, institute
Step 4.1 is stated specifically to comprise the following steps：
Step 4.11, background modeling is carried out to input picture by gauss hybrid models method, prospect interested is carried
Take out；
Step 4.12, the prospect and area-of-interest r of sliding window Scanning Detction present frame are passed throughi, then extract in region
Mixing light stream histogram feature, and ri={ di,li, wherein, di and liMixing light stream histogram feature and space are corresponded to respectively
Position.
As the further preferred scheme of global and local anomaly detection method in a kind of crowd's scene of the present invention, institute
Step 4.2 is stated specifically to comprise the following steps：
Step 4.21, the locus of a two field picture is divided into 2L-1×2L-1Individual grid, wherein, L is size of space level；
Step 4.22, we are modeled by OWC algorithms to the motion feature in certain time in each grid：
Wherein, the cluster renewal in grid is specific as follows：
ωk←ωk+α(ok-ωk)
Ck←Ck+ok(α/ωk)(di-Ck)
Wherein,For weight, CkIt is the barycenter that k is clustered in a grid, α is learning rate, and 0<α<1, if from diTo Ck
Euclidean distance be less than threshold θ, then ok1 is set to, otherwise okIt is set to 0；
Step 4.23, the HOHFs that normally clusters will be belonged to and thinks that there is larger weight, will belong to what is clustered extremely
HOHFs thinks there is less weight, will cluster by ωkDescending sort, then define first cluster B be it is normal, i.e.,
Wherein, δ is to determine the minimum threshold value that can position normal HOHFs in cluster；If diFirst cluster B can be matched
In one, then it be detected as normally, be otherwise exception.
As the further preferred scheme of global and local anomaly detection method in a kind of crowd's scene of the present invention, institute
It is as follows to state step 4.3 multiple target tracking：
Step 4.31, in moment t, one group of track is obtained：T={ t1,...,tm, wherein ti={ xi,t-1,Pi,t-1,A,
H,Q,R}；M is the total number of tracks obtained, xi,t-1And Pi,t-1Estimated state and estimated state covariance, A for the t-1 moment are
State-transition matrix, H are calculation matrix, and Q is process noise covariance, and R is measurement noise covariance；
Step 4.32, extracted by connected component labeling algorithm from the abnormal ROIs detected foreground pixel a series of
The barycenter of new block, wherein block is defined as Z={ z1,t,...,zn,t}；
Step 4.33, to an already present track ti, use the estimated state x at t-1 momenti,t-1With estimated state association side
Poor Pi,t-1To predict its new state ^xi,tWith state covariance ^Pi,t；Specifically it is calculated as follows：
Wherein, A condition transfer matrix, Q are process noise covariances；
Step 4.34, existing track t will be associated by gating calculatingiAnd zj,tTo play the effect of multiple target tracking.
The present invention uses above-mentioned technical proposal, can produce following technique effect：
1st, in feature level, the present invention proposes a kind of descriptor of new entitled mixing light stream histogram, for global different
Normal behavioral value is made up of two sparse processes, and each process has the process of a dynamic lexicon renewal, two Thinning Process
Two probability are provided, last result is judged by fuzzy integral；
2nd, for the blockage of the detection of local anomaly behavior, first extraction area-of-interest, then with online weighted cluster
Algorithm judged, finally noise is filtered using the method for multiple target tracking；
3rd, method proposed by the present invention is avoided that the interference of the shade of illumination variation, can accurately detect global abnormal and not
The local anomaly behavior of same type, there is certain robustness, validity.
Brief description of the drawings
Fig. 1 is the inventive method flow chart；
Fig. 2 is the experimental result schematic diagram under three kinds of different scenes in UMN data sets；
Fig. 3 is the ROC curve of the global abnormal behavioral value in UMN data sets；
Fig. 4 (a) is the ped1 experimental result schematic diagrames in UCSD data sets；
Fig. 4 (b) is the ped2 experimental result schematic diagrames in UCSD data sets；
Fig. 5 is the ROC curve of local anomaly behavioral value in ped1 data sets；
Fig. 6 is the ROC curve of local anomaly behavioral value in ped2 data sets.
Embodiment
Technical scheme is described in detail below in conjunction with the accompanying drawings：
It is an object of the invention to provide a kind of new global and local anomaly detection method, its realization approach is：Carry
A kind of feature of the entitled mixing light stream histogram of new descriptor is gone out.There are two sparse mistakes for the detection of global abnormal behavior
Journey, one judges whether region is normal, and another judges whether region is abnormal, and each process has what a dynamic lexicon updated
Process, two Thinning Process provide two probability, finally the two are judged to handle with fuzzy integral, obtain the region
Final result.For local anomaly behavioral value, the prospect of area-of-interest is first extracted, is then carried out by online weighted cluster
Detection, noise is filtered out with the method for multiple target tracking.The inventive method flow chart is as shown in Figure 1.
Wherein, the method for feature extraction：
Frame of video is divided into some size identical blockages first, and along the same position in the successive frame of time shaft
Several such blockages form a container.For when empty container in each piece, calculate the distributed intelligence of light stream, its optics
Vector can be expressed as F (x, y)=(vx(x,y),vy(x, y)), v herex(x,y),vy(x, y) represents light stream along x-axis, y-axis,
Therefore, size of the light stream on (x, y) can be expressed as：
Its direction can be expressed as：
The light stream direction of all pixels can [0,360 °) in the range of be divided into the equal regions of b, pixel is to scope
The contribution of interior light stream is：
Here bkIt is the direction of k-th of scope, 1≤k≤b.For when empty container in each piece, calculate along b directions
Light stream distributed intelligence, and histogram H=(G can be obtained1,G2,...,Gb)∈R1*b, then, for it is same when empty container
With a series of blockages in different frame, desired value and variance are calculated.Assuming that when empty container V in there are r to continue frames, calculate this
The light stream histogram of a little frames, their expectation and variance are as follows：
Here μk(1≤k≤b) represents light stream in the expectation in k-th of direction, δk(1≤k≤b) is represented in the side of equidirectional
Difference, wherein μ=(μ1,μ2,...,μb)∈R1*bAnd δ=(δ1,δ2,...,δb)∈R1*b.Larger variance yields causes to become apparent from
Optical flow velocity change.In addition, the change for also remembering speed is acceleration, therefore, the value of k directional accelerations can be expressed as α
∝δk(1≤k≤b) we combine expectation in histogram and variance describes the motion feature in local space time's container, Nogata
Figure can be expressed as HG=(μ, δ).Descriptor is called mixing light stream histogram.Compared with general feature is extracted, our method
It is the change for calculating light stream on direction and size, so as to form acceleration information, so as to improve the expression of motion state.
The method of sparse reconstruct：
Frame of video generally represents b=[H with a series of feature1,H2,...,HN], N is the total amount of key character here.
In order to represent the video sample with high dimensional feature, rarefaction representation provides a kind of rational method.These features are multiplied by one
Group represents the coefficient of feature weight, and we can be obtained by：
Y=[H1X1,H2X2,…,HNXN]T(6) X here1,X2,...,XNIt is coefficient.To obtain optimal coefficient sets, it should
MakeIt is sufficiently small, it means that we need to establish an optimization problem, are expressed as：
Here L2Norm | | | |2Represent Euclidean distance.With matrix Φ expression characteristics, formula 7 is：
It is determined that during sample, in order that with less feature, there must be some 0 elements in coefficient vector x.L0Norm | | x |
|0It is added in formula 8：
L0Norm calculation nonzero element number, therefore, the optimal coefficient sets of minimum feature calculation of formula 9.Use L1Norm
To replace L0Norm has identical effect, herein：
Formula 10 is the final form of rarefaction representation.Wherein x*It is reconstructed coefficients.We are revised as formula 11, and will
Sparse reconstructed cost is expressed as formula 12.
Normal frame reconstruct cost is smaller, and abnormal frame generally produces larger reconstruction cost.It is therefore possible to use SRC makees
For the exception measurement of oneclass classification problem.
The renewal of dynamic lexicon：
The purpose of dictionary renewal is dictionary is more accurately described normal sample.Initial dictionary is only including training data just
Normal sample, it is intended to the normal sample of some test samples is added in dictionary, so that dictionary can preferably handle test number
According to.We solve the problems, such as how to select extensive sample to update dictionary.Dictionary Φ is established have selected enough samples
=[b1,b2,...bN]TAfterwards, judge whether upcoming sample is normal.In the case where the dimension of dictionary is constant, should increase
Various normal samples.Dictionary is scanned, a sample minimum with this normal sample distance is found, its frame is merged into word
In allusion quotation.Here distance was weighed originally with sparse be reconstructed into, and the SRC of all elements is expressed as SRC in dictionary Φdic=
[SRC1 dic,SRC2 dic,...,SRCN dic]T, the SRC of next sample is represented by SRCy, dynamic lexicon, which updates, to be expressed as：
Here m*It is the sample index of dictionary, it reaches next sample y distance minimum, ωm*It is that new will substitute word
M in allusion quotation Φ*Individual sample.bm*It is m in old dictionary*Individual sample.
The process of double rarefaction representations：
Double rarefaction representations in the present invention are a systems for including two rarefaction representation processes.Only comprising normal in dictionary
Base, the foundation of another dictionary include the exceptional sample obtained by first rarefaction representation process.Abnormal dictionary can be by very
The base composition of low probability.The equation of second rarefaction representation is：
Here ΦabIt is abnormal dictionary, formula 14 and 15 has no difference with formula 11 and 12, but new dictionary ΦabWith it is normal
The similar processes for needing a dynamic lexicon renewal of dictionary Φ, various exceptional samples are added in dictionary.Test video first
In several test samples it is whether normal.If sample is normal, Φ can then update.When sample if it is determined that abnormal be then added
To Φab.In ΦabAfter foundation, the sample that arrives afterwards has two Thinning Process and judges that the two Thinning Process are simultaneously respectively
Capable.Finally judgement is provided with fuzzy integral.As final result, renewal process must continue to down.If final result
It is normal, then test result is added in Φ, is otherwise added to ΦabIn.
The extraction of the area-of-interest prospect：
After carrying out background modeling by gauss hybrid models (GMM) method, sliding window is used for Scanning Detction to currently
The prospect of frame may be relevant with local abnormal behaviour with area-of-interest (ROI).Specifically, for all in sliding window
Pixel, if the ratio of foreground pixel is more than threshold value, window is highlighted as ROI.As can be seen that the filtering of this process is big
Partial background, reduce the robustness for calculating cost and improving abnormality detection.
Because an exception not only occurs in a frame, so we employ the sliding window centered on a time
t:T-n, and t-1, t, t+1, t+n } go to extract HOHF descriptors.Except movable information, we also retain often
Individual ROI locus (i.e. x and y coordinates).From now on, a ROI is expressed as ri={ di,li, di and l hereiRespectively
Corresponding HOHF and locus.Here HOHF features are by catching direction and exercise intensity data separation motor pattern.
The online weighted cluster algorithm：
The locus of one two field picture is divided into 2L-1×2L-1Individual grid, L is size of space level here, and ROIs is positioned at identical
Grid is clustered by on-line study.The step for help to overcome perspective distortion, this may cause extract feature yardstick
Change.
We are modeled by OWC algorithms to the feature in the nearest time in each grid.Specifically, allow
ω1,...,ωKAs weight, and C1,...,CKIt is the barycenter that K clusters (K≤T) in a grid, ∑ K hereK=1ωk=
1, K and T is the quantity of existing cluster and the maximum cluster numbers specified respectively.For a new ri, we are first its diDistribution
It is as follows with li and cluster renewal within a grid into a grid：
ωk←ωk+α(ok-ωk) (16)
Ck←Ck+ok(α/ωk)(di-Ck) (17)
Here α (0<α<1) it is learning rate, if from diTo CkEuclidean distance be less than threshold θ ok1 is set to, is otherwise set to 0.
If without existing cluster match di, then a new cluster ω is producedK+1=α and CK+1=di.If K<T, new
Cluster is added in existing cluster.If K=T, minimal weight ωkCluster substituted by new cluster, weight is at this
Reformed after step.
Generally fall into the HOHFs normally clustered and be considered to have larger weight, however represent abnormal cluster have compared with
Small weight.Therefore, if cluster is by ωkDescending sort, it is normal that we, which define first B cluster,：
Here δ is to determine the minimum threshold value that can position normal HOHFs in cluster.In other words, if diFirst can be matched
One in individual cluster B, then it be detected as normally, be otherwise exception.
One important advantage of the algorithm is that it is adaptive.From formula 16, we draw a conclusion, represent normal
If one of HOHFs clusters it and can not match a new diUpdated with log δ/log (1- α) for principle, then it will be considered as
It is the cluster of an exception.Likewise, if we always detect some similar abnormal HOHFs, these HOHFs clusters are represented
Weight will gradually increase and ultimately become normal cluster.
The multiple target tracking algorithm：
Although we detect ROIs abnormal in every frame video, because moving object is seriously blocked.Still there are some wrong
The detection for missing and losing.Assuming that an anomalous event appears in a continuous space-time position, we employ simplified MTT and calculated
Method improves detection performance.
In moment t, we can obtain one group of track T={ t1,...,tm, wherein the state of each track is by karr
Graceful modeling filter.It is defined as ti={ xi,t-1,Pi,t-1,A,H,Q,R}.At the same time, we pass through connected component labeling algorithm
A series of new blocks are extracted from the abnormal ROIs detected foreground pixel.The barycenter of these blocks is defined as Z={ z1,t,...,
zn,t}.To an already present track ti, we are the estimated state x first by the t-1 momenti,t-1With estimated state covariance
Pi,t-1To predict its new state ^xi,tWith state covariance ^Pi,t：
Here A condition transfer matrix, Q are process noise covariances.Next, existing track will be associated by gating calculating
tiAnd zj,t.Specifically, Distance matrix D is constructed, here DijIt is tiAnd zj,tBetween mahalanobis distance square measurement：
Here zj,tIt is barycenter of the block j in moment t, H is calculation matrix, and R is measurement noise covariance.And if only if Dij≤G
When, we are the z candidatej,tWith track tiAssociation.Here G is gate size.When there is multiple pieces in door, we are using global
Nearest block is distributed to track by arest neighbors (GNN) scheme.Because sparse is distributed block and the anomaly detector robust of proposition, it
Effect is fine in our method.
One block can distribute to existing track, can also start a new track.An if block zj,tIt is allocated
To existing track ti, track is updated as follows：
Here K is Kalman's income, and I is unit matrix.If no block is assigned to existing track ti, track is updated
For：
If distributing block without for it in a number of frame, existing track is deleted, according to path length, the rail
Mark may disappear or noise.
In order to test it is proposed that method effect, the algorithm is used in several announced data sets.This hair
It is bright that algorithm performance in our global abnormal behavioral values is tested using UMN data sets, and UCSN data sets are used to handle
Algorithm performance is tested during local anomaly behavior.
Experimental result in UMN data sets：
UMN data sets are made up of two Outdoor Scenes and an indoor scene, quick comprising several crowds under each scene
Human behavior reaction.Totalframes is 7,739, and resolution ratio is 320 × 240.For each scene, we are normal using 500 to 1600
Frame is trained, and remaining frame is used to test.Parameter r during feature extraction is set to 4, b in experiment and is set to 8.
Comparatively, UMN data sets are fairly simple compared with other data sets, so current Outlier Detection Algorithm
This data set is performed well.What Fig. 2 was provided is normal and abnormal experiment of the inventive algorithm under three different scenes
Result figure.Image result is labeled in upper left side.Three kinds of distinct methods on UMN data sets are compared by we, including society
Can power model Social Force Model, optical flow method Optical Flow and streak line representation Streakline
Representation Method.ROC curve is as shown in Figure 3.Real class rate (the True Positive that ROC curve is shown
Rate, TPR) and false positive class rate (False Positive Rate, FTR).TPR is the ratio of correct mark frame, and FPR is mistake
Mark the ratio of frame.Table 1 is AUC and the comparison of EER values and remaining method of method proposed by the invention.
Table 1
| Method | AUC (%) | EER (%) |
| Optical Flow | 83 | 24 |
| Social Force | 93.8 | 12.6 |
| SPM | 96.2 | 13.5 |
| Our Method | 97.5 | 13.2 |
It can be seen from the graph that behavior normal and abnormal under three kinds of scenes, and crowd is spread to surrounding suddenly
Abnormal behaviour is labeled as in the upper right corner, the more other three kinds of sides of algorithm of the present invention are can be seen that from ROC curve, AUC and EER value
Method has certain superiority, robustness, validity.
Experimental result in UCSD data sets：
We have carried out the experiment of local anomaly behavioral value to UCSD data sets, and the data set includes two different pedestrians
The subset of passage.First group, be expressed as " Ped1 ", there are 34 training editings and 36 test clips, each with 158 ×
238 resolution ratio, fixed length are 200 frames.And second is represented as " Ped2 ", there are 16 training editings and 12 tests
Editing, the resolution ratio of each test clips is 240 × 360, length frame from 120 to 180.In experiment, we have in application
1/2 20 × 10 overlapping sliding window extract ROI, and both time spans are arranged to 6.Parameter θ is set to 0.15, N=8, L=
4, T=15, and α=0.001.
Same Fig. 4 (a) is ped1 experimental result, and Fig. 4 (b) is ped2 experimental result, and Fig. 5 and Fig. 6 are the present invention
Used method is the same as other three kinds of methods：Method (Social Force) based on social force model, grown based on incremental encoding
Spend (Incremental Coding Length, ICL), the abnormality detection of the stratified condition random field based on mixing dynamic texture
(Aconditional random field with a hierarchical mixture of dynamic texture, H-
MDT-CRF) the comparison of the ROC curve on ped1 and ped2.Table 2 is AUC and the comparison of EER values of these four methods.
Table 2
It can be seen from the graph that in ped1 and ped2 data sets, sliding slide plate, is cycled in crowd, and car, pedestrian steps on
Enter lawn and be correctly labeled as red, the ROC curve of test result and the extensive anomaly detection method of currently used comparison,
AUC and EER values are contrasted, and the superiority of proposition method of the present invention is fully proved in aspect of performance.
Claims (7)
1. global and local anomaly detection method in a kind of crowd's scene, it is characterised in that specifically comprise the steps of：
Step 1, crowd's scene image to be detected is inputted；
Step 2, the mixing light stream histogram of crowd's scene image to be detected is extracted；
Step 3, global abnormal behavioral value is carried out to the mixing light stream histogram of step 2 feature extraction, it is specific as follows：
Step 3.1, it is whether abnormal that the image after judging characteristic extraction is distinguished by two rarefaction representation processes；
Step 3.2, handled, and then obtained a result with judgement of the fuzzy integral to two rarefaction representation processes；
Step 4, local anomaly behavioral value is carried out to the mixing light stream histogram of step 2 feature extraction, it is specific as follows：
Step 4.1, the prospect of area-of-interest in input picture is extracted；
Step 4.2, local anomaly detection is carried out to area-of-interest by online weighted cluster；
Step 4.3, noise is filtered out using the method for multiple target tracking；
Step 5, global and local unusual checking in more crowd's scenes is completed by step 3 and step 4.
2. global and local anomaly detection method in a kind of crowd's scene according to claim 1, it is characterised in that
In step 2, light stream histogram H is mixedGIt is specific to represent as follows：
HG=(μ, δ)
Wherein,μ represents the expectation of light stream, and δ represents the variance in equidirectional, r tables
There are r to continue frame, H when showing in empty containeriRepresent i-th of light stream histogram.
3. global and local anomaly detection method in a kind of crowd's scene according to claim 1, it is characterised in that
In step 3.1, described two rarefaction representation processes include the process of a dynamic lexicon renewal, wherein, dynamic lexicon is more
The expression formula of new process is as follows：
Wherein, the SRC of all elements is expressed as SRC in dictionary Φdic=[SRC1 dic,SRC2 dic,...,SRCN dic]T, it is next
The SRC of sample is represented by SRCy, m* is the sample index of dictionary, and it is reached, and next sample y distance is minimum, and ω m* are new
Will the m* sample in substitution dictionary Φ, bm* is the m* sample in old dictionary.
4. global and local anomaly detection method in a kind of crowd's scene according to claim 1, it is characterised in that
The formula of the rarefaction representation of two rarefaction representation processes is as follows：
Wherein, SRC is sparse reconstruct cost, and x represents the coefficient of feature weight, and y is characterized the expression formula after being multiplied by weight coefficient,
λ is parameter, and Φ is dictionary , ║ ║2For Euclidean distance , ║ x ║1For L1Norm.
5. global and local anomaly detection method in crowd's scene according to claim 1, it is characterised in that the step
Rapid 4.1 specifically comprise the following steps：
Step 4.11, background modeling is carried out to input picture by gauss hybrid models method, foreground extraction interested is gone out
Come；
Step 4.12, the prospect and area-of-interest r of sliding window Scanning Detction present frame are passed throughi, then extract the mixing in region
Light stream histogram feature, and ri={ di,li, wherein, di and liMixing light stream histogram feature and locus are corresponded to respectively.
6. global and local anomaly detection method in crowd's scene according to claim 1, it is characterised in that the step
Rapid 4.2 specifically comprise the following steps：
Step 4.21, the locus of a two field picture is divided into 2L-1×2L-1Individual grid, wherein, L is size of space level；
Step 4.22, we are modeled by OWC algorithms to the motion feature in certain time in each grid：
Wherein, the cluster renewal in grid is specific as follows：
ωk←ωk+α(ok-ωk)
Ck←Ck+ok(α/ωk)(di-Ck)
Wherein,For weight, CkIt is the barycenter that k is clustered in a grid, α is learning rate, and 0<α<1, if from diTo CkIt is European
Distance is less than threshold θ, then ok1 is set to, otherwise okIt is set to 0；
Step 4.23, the HOHFs that normally clusters will be belonged to and thinks that there is larger weight, the HOHFs that clusters extremely will be belonged to and recognized
With less weight, will to cluster by ωkDescending sort, then define first cluster B be it is normal, i.e.,
Wherein, δ is to determine the minimum threshold value that can position normal HOHFs in cluster；If diIt can match in first cluster B
One, then it be detected as normally, be otherwise exception.
7. global and local anomaly detection method in crowd's scene according to claim 1, it is characterised in that the step
Rapid 4.3 multiple target tracking is as follows：
Step 4.31, in moment t, one group of track is obtained：T={ t1,...,tm, wherein ti={ xi,t-1,Pi,t-1,A,H,Q,
R}；M is the total number of tracks obtained, xi,t-1And Pi,t-1Estimated state and estimated state covariance for the t-1 moment, A is state
Transfer matrix, H are calculation matrix, and Q is process noise covariance, and R is measurement noise covariance；
Step 4.32, extracted by connected component labeling algorithm from the abnormal ROIs detected foreground pixel a series of new
The barycenter of block, wherein block is defined as Z={ z1,t,...,zn,t}；
Step 4.33, to an already present track ti, use the estimated state x at t-1 momenti,t-1With estimated state covariance
Pi,t-1To predict its new state ^xi,tWith state covariance ^Pi,t；Specifically it is calculated as follows：
Wherein, A condition transfer matrix, Q are process noise covariances；
Step 4.34, existing track t will be associated by gating calculatingiAnd zj,tTo play the effect of multiple target tracking.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710684991.3A CN107491749B (en) | 2017-08-11 | 2017-08-11 | Method for detecting global and local abnormal behaviors in crowd scene |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710684991.3A CN107491749B (en) | 2017-08-11 | 2017-08-11 | Method for detecting global and local abnormal behaviors in crowd scene |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107491749A true CN107491749A (en) | 2017-12-19 |
| CN107491749B CN107491749B (en) | 2020-11-17 |
Family
ID=60645076
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710684991.3A Active CN107491749B (en) | 2017-08-11 | 2017-08-11 | Method for detecting global and local abnormal behaviors in crowd scene |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107491749B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108563217A (en) * | 2018-05-29 | 2018-09-21 | 济南浪潮高新科技投资发展有限公司 | The robust method for detecting abnormality analyzed based on part and global statistics |
| CN108596045A (en) * | 2018-04-02 | 2018-09-28 | 四川大学 | A kind of group abnormality behavioral value method based on aerial monitor supervision platform |
| CN108629316A (en) * | 2018-05-08 | 2018-10-09 | 东北师范大学人文学院 | A kind of video accident detection method of various visual angles |
| CN108764209A (en) * | 2018-06-11 | 2018-11-06 | 南京邮电大学 | A kind of personage's abnormal behaviour automatic testing method based on monitor video |
| CN109117774A (en) * | 2018-08-01 | 2019-01-01 | 广东工业大学 | A kind of multi-angle video method for detecting abnormality based on sparse coding |
| CN109300144A (en) * | 2018-09-21 | 2019-02-01 | 中国矿业大学 | A kind of pedestrian track prediction technique of mosaic society's power model and Kalman filtering |
| CN109614893A (en) * | 2018-11-28 | 2019-04-12 | 中国电子科技集团公司电子科学研究院 | Abnormal behaviour whereabouts intelligent identification Method and device based on situation reasoning |
| CN110443287A (en) * | 2019-07-19 | 2019-11-12 | 北京航空航天大学 | A kind of mobile stream method for drafting of the crowd based on sparse track data |
| CN112016701A (en) * | 2020-09-09 | 2020-12-01 | 四川大学 | Abnormal change detection method and system integrating time sequence and attribute behaviors |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105138982A (en) * | 2015-08-21 | 2015-12-09 | 中南大学 | Crowd abnormity detection and evaluation method based on multi-characteristic cluster and classification |
| CN106033548A (en) * | 2015-03-13 | 2016-10-19 | 中国科学院西安光学精密机械研究所 | Crowd abnormity detection method based on improved dictionary learning |
| CN106228149A (en) * | 2016-08-04 | 2016-12-14 | 杭州电子科技大学 | A kind of video anomaly detection method |
-
2017
- 2017-08-11 CN CN201710684991.3A patent/CN107491749B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106033548A (en) * | 2015-03-13 | 2016-10-19 | 中国科学院西安光学精密机械研究所 | Crowd abnormity detection method based on improved dictionary learning |
| CN105138982A (en) * | 2015-08-21 | 2015-12-09 | 中南大学 | Crowd abnormity detection and evaluation method based on multi-characteristic cluster and classification |
| CN106228149A (en) * | 2016-08-04 | 2016-12-14 | 杭州电子科技大学 | A kind of video anomaly detection method |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108596045A (en) * | 2018-04-02 | 2018-09-28 | 四川大学 | A kind of group abnormality behavioral value method based on aerial monitor supervision platform |
| CN108629316A (en) * | 2018-05-08 | 2018-10-09 | 东北师范大学人文学院 | A kind of video accident detection method of various visual angles |
| CN108563217A (en) * | 2018-05-29 | 2018-09-21 | 济南浪潮高新科技投资发展有限公司 | The robust method for detecting abnormality analyzed based on part and global statistics |
| CN108764209A (en) * | 2018-06-11 | 2018-11-06 | 南京邮电大学 | A kind of personage's abnormal behaviour automatic testing method based on monitor video |
| CN109117774B (en) * | 2018-08-01 | 2021-09-28 | 广东工业大学 | Multi-view video anomaly detection method based on sparse coding |
| CN109117774A (en) * | 2018-08-01 | 2019-01-01 | 广东工业大学 | A kind of multi-angle video method for detecting abnormality based on sparse coding |
| CN109300144A (en) * | 2018-09-21 | 2019-02-01 | 中国矿业大学 | A kind of pedestrian track prediction technique of mosaic society's power model and Kalman filtering |
| CN109300144B (en) * | 2018-09-21 | 2021-12-24 | 中国矿业大学 | Pedestrian trajectory prediction method integrating social force model and Kalman filtering |
| CN109614893A (en) * | 2018-11-28 | 2019-04-12 | 中国电子科技集团公司电子科学研究院 | Abnormal behaviour whereabouts intelligent identification Method and device based on situation reasoning |
| CN109614893B (en) * | 2018-11-28 | 2023-10-24 | 中国电子科技集团公司电子科学研究院 | Intelligent abnormal behavior track identification method and device based on situation reasoning |
| CN110443287A (en) * | 2019-07-19 | 2019-11-12 | 北京航空航天大学 | A kind of mobile stream method for drafting of the crowd based on sparse track data |
| CN110443287B (en) * | 2019-07-19 | 2022-01-14 | 北京航空航天大学 | Crowd moving stream drawing method based on sparse trajectory data |
| CN112016701A (en) * | 2020-09-09 | 2020-12-01 | 四川大学 | Abnormal change detection method and system integrating time sequence and attribute behaviors |
| CN112016701B (en) * | 2020-09-09 | 2023-09-15 | 四川大学 | Abnormal change detection method and system integrating time sequence and attribute behaviors |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107491749B (en) | 2020-11-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107491749A (en) | Global and local anomaly detection method in a kind of crowd's scene | |
| Junior et al. | Crowd analysis using computer vision techniques | |
| US8213679B2 (en) | Method for moving targets tracking and number counting | |
| CN103985114B (en) | A kind of monitor video personage's foreground segmentation and the method for classification | |
| CN103020986B (en) | A kind of motion target tracking method | |
| CN110084228A (en) | A kind of hazardous act automatic identifying method based on double-current convolutional neural networks | |
| CN109919122A (en) | A kind of timing behavioral value method based on 3D human body key point | |
| CN105718857A (en) | Human body abnormal behavior detection method and system | |
| CN109344736A (en) | A kind of still image people counting method based on combination learning | |
| CN107239762A (en) | Patronage statistical method in a kind of bus of view-based access control model | |
| CN106204640A (en) | A kind of moving object detection system and method | |
| CN105844234A (en) | People counting method and device based on head shoulder detection | |
| CN106529499A (en) | Fourier descriptor and gait energy image fusion feature-based gait identification method | |
| CN106295532B (en) | A kind of human motion recognition method in video image | |
| CN102254183B (en) | Face detection method based on AdaBoost algorithm | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN104850865A (en) | Real-time compression tracking method of multi-characteristic transfer learning | |
| CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
| CN109948690A (en) | A kind of high-speed rail scene perception method based on deep learning and structural information | |
| CN103136534A (en) | Method and device of self-adapting regional pedestrian counting | |
| CN112836640A (en) | Single-camera multi-target pedestrian tracking method | |
| CN106570490A (en) | Pedestrian real-time tracking method based on fast clustering | |
| CN106127812A (en) | A kind of passenger flow statistical method of non-gate area, passenger station based on video monitoring | |
| CN110909672A (en) | Smoking action recognition method based on double-current convolutional neural network and SVM | |
| CN107590427A (en) | Monitor video accident detection method based on space-time interest points noise reduction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information |
Address after: Room 201, building 2, phase II, No.1 Kechuang Road, Yaohua street, Qixia District, Nanjing City, Jiangsu Province Applicant after: NANJING University OF POSTS AND TELECOMMUNICATIONS Address before: 210023 Guang Yue Road, Qixia street, Nanjing, Jiangsu Province, No. 9-1 Applicant before: NANJING University OF POSTS AND TELECOMMUNICATIONS |
|
| CB02 | Change of applicant information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |