CN104010168B - A kind of non-overlapping visual field multiple-camera monitors network topology adaptation learning method - Google Patents
A kind of non-overlapping visual field multiple-camera monitors network topology adaptation learning method Download PDFInfo
- Publication number
- CN104010168B CN104010168B CN201410266226.6A CN201410266226A CN104010168B CN 104010168 B CN104010168 B CN 104010168B CN 201410266226 A CN201410266226 A CN 201410266226A CN 104010168 B CN104010168 B CN 104010168B
- Authority
- CN
- China
- Prior art keywords
- node
- pair
- target
- connection
- observation
- 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.)
- Expired - Fee Related
Links
Abstract
The present invention proposes a kind of non-overlapping visual field multiple-camera monitoring network topology adaptation learning method, is related to computer vision field.The present invention uses weighted digraph G=<V,E,W>Represent the topology of monitoring network.The present invention is using the leaving of target under the single camera ken, in-position is modeled as topological node V using mixed Gauss model.The present invention proposes a kind of Intensity correlation function computational methods based on joint appearance similarity, and judges certain connectedness to node by Intensity correlation function, so as to obtain side collection E.For the node pair for connecting, transfer time distribution is calculated by standardizing Intensity correlation function.The present invention represents this pair of transition probability of node using the interactive information of node pair, so as to obtain weight set W.The present invention proposes that a kind of " falseness connection " Solve Problem is excluded in topology possible " falseness connection ", and proposes that topology adaptation more new strategy ensures that topological structure has higher robustness to environmental change.
Description
Technical field
The invention belongs to computer vision field, and in particular to field of intelligent monitoring, more particularly to a kind of non-overlapping visual field
Multiple-camera monitors network topology adaptation learning method.
Background technology
With the development of camera supervised technology, extensive area is monitored becomes guarantee people's lives and properties peace
A full important means.However, for the larger monitoring occasion in region, all of monitor area is covered using video camera
It is very unrealistic.Therefore, the multiple-camera comprising non-overlapping visual field is generally built using the method for key area covering and monitors system
System.Compared with traditional single camera monitoring system or overlap ken multiple-camera monitoring system, being imaged non-overlapping visual field more
Machine monitoring system is all discrete in the time, spatially due to its observed object, so continuous tracking is carried out to target being more stranded
It is difficult.Common processing method is the topological structure for learning camera network, and the space-time restriction provided using topological structure is believed
Cease to help match the target under the different kens, realize the continuous tracking to target.
The topology structure learning method of non-overlapping visual field multiple-camera network is broadly divided into two kinds of supervised and non-supervisory formula.
Supervised learning method is tracked by the target to artificial mark, so that learn the topological structure of camera network, most often
See as O.Javed utilization Parzen window learning network topological structures space-time restriction information, such as document " Javed O.,
Rasheed Z.,Shafique K.,Shah M.Tracking across multiple cameras with disjoint
views[C].IEEE International Conference on Computer Vision,2003:952-957 " and text
Offer " Javed O, Shafique K, Shah M.Appearance modeling for tracking in multiple
non-overlapping cameras[C].2005IEEE Computer Society Conference on Computer
Vision and Pattern Recognition,2005:26-33”.But, supervised learning method generally needs mark big
The data of amount, and change adaptability in actual use for monitors environment is not strong, therefore it is difficult to apply to actual prison
Control system.
Unsupervised learning methods do not need artificial flag data, and system can be according to each child node inspection in monitoring network
The topological structure of the data adaptive ground Learning-memory behavior system for measuring.Unsupervised learning methods generally using Intensity correlation function come
Infer and whether connect between certain two node, such as document " Makris D., Ellis T., Black J..Bridging the
gaps between cameras[C].IEEE Computer Society Conference on Computer Vision
and Pattern Recognition,2004:205-210 ", but Intensity correlation function is only comprising two nodes in certain time
The information that interior target disappears and occurs, the similarity information not comprising target, therefore error is larger.There is scholar to propose two afterwards
Plant improved method：Colouring information is fused in Intensity correlation function, such as document " Niu C., Grimson E..Recovering
non-overlapping network topology using far-field vehicle tracking data[C]
.International Conference on Pattern Recognition,2006:944-949 ", by adding target
Colouring information effectively increases Intensity correlation function and judges connective ability, meanwhile, can be inferred that prison using connective matrix
Control the topological structure of network.On the basis of color characteristic, the identity information obtained using face detection is fused to mutually
In correlation function, such as document " Zou X.T., Bhanu B., Roy-Chowdhury A.Continuous learning of a
multilayered network topology in a video camera network.Research
Article460689,Center for Research in Intelligent Systems,University of
California, USA, 2009 ", by the identity information for adding, substantially increases only with the Intensity correlation function of colouring information
Judge connective ability.On the premise of accurate acquisition identity information, the Intensity correlation function that the method is calculated can be obtained
It is most accurately connective to judge.But, face detection all there are certain requirements for imaging precision and angle, be not suitable for pushing away
Extensively.
Therefore, the topological structure of network is monitored using unsupervised learning method adaptive learning multiple-camera, for taking the photograph more
Can camera monitoring system be applied to actual most important.In unsupervised topology learning method, mainly there are following three points to need
Note：(1) using the Intensity correlation function of node pair judge this pair of node it is connective when, how to be added in Intensity correlation function
Target signature that is representative and being easy to extraction, to improve using the connective degree of accuracy of Intensity correlation function decision node；(2)
In the topological structure for tentatively obtaining, how to exclude " falseness connection "；(3) in actual use, how adaptive updates
Topological structure, to adapt to the change of monitors environment.
The content of the invention
It is an object of the invention to propose a kind of non-overlapping visual field multiple-camera monitoring network topology adaptation learning method.
The method has very strong robustness to indoor change, complicated monitors environment.
The technical scheme is that：The topological structure of non-overlapping visual field multiple-camera monitoring network is had with a weighting
To figure G=<V,E,W>To represent, and the three elements in study G respectively：The weight set of node set V, line set E and side
W.Enter, leave the position of the ken according to the target following result statistics target under haplopia domain, and by mixed Gauss model
The disappearance that (GMM, Gaussian Mixture Model) is set up under the ken-nodal analysis method occur.Disappearance under all kens-
There is nodal analysis method configuration node set V.Across ken node is calculated to (two nodes are respectively from the different cameras ken)
Intensity correlation function, so as to judge this pair of connectedness of node.Across the ken node of all connections is identical to regard to composition line set E
Node under domain is to being not added in line set E.In calculate node to (node pair hereinafter mentioned, unexplained reference all tables
Show across ken node to) Intensity correlation function when, it is contemplated that disappearance observation in this pair of node with there is combining for observation
Appearance similarity, so as to improve using Intensity correlation function decision node to the connective degree of accuracy.For the node pair for connecting,
It is distributed by standardizing this pair of transfer time of the Intensity correlation function calculate node pair of node.The interactive information for introducing node pair is come
This pair of transition probability of node is represented, disconnected node is 0, the interactive information of the node pair under the identical ken to interactive information
Also it is set to 0.Transition probability construction weight set W according to all nodes pair.Connected with weight set W constructions according to line set E
Property matrix, and by the matrix infer monitoring network topological structure." falseness connection " Solve Problem removal is utilized to be inferred to
" falseness connection " that may be present in topological structure.Policy update topological structure and parameter are updated using topology adaptation, while
Propose the quick response strategy to addition/removal camera situation so that the topological structure for learning can be well adapted to
Monitor the change in bad border.
The step that implements of the invention is followed successively by：
1) set up disappear-there is node set
2) line set of topological structure is calculated
3) the transfer time distribution of connection node pair is calculated
4) the weight set of topological structure is calculated
5) construct connective matrix and infer topological structure
6) exclude " falseness connection "
7) renewal of topological structure and parameter
Brief description of the drawings
Fig. 1 is the system stream of topology adaptation learning method of the present invention based on non-overlapping visual field multiple-camera monitoring network
Cheng Tu
Fig. 2 is the disappearance-node schematic diagram occur of the monitoring scene that the present invention is calculated
Fig. 3 is Intensity correlation function curve comparison schematic diagram in the present invention
Fig. 4 is connective matrix and topological structure schematic diagram in the present invention
Fig. 5 is the connective matrix and topological structure schematic diagram after removal in the present invention " falseness connection "
Fig. 6 is transfer time distributed update schematic diagram in the present invention
Fig. 7 is that transition probability of the present invention updates schematic diagram
Specific embodiment
Fig. 1 gives the system flow of the topology adaptation learning method based on non-overlapping visual field multiple-camera monitoring network
Figure：Use weighted digraph model G=<V,E,W>To represent that non-overlapping visual field multiple-camera monitors the topological structure of network, and
Learn the three elements in G respectively：Node set V, line set E and weight set W.The present invention only considers the different cameras ken
In node (i.e. across ken node) connectedness, the connection situation of identical ken lower node is not considered, therefore under the identical ken
The node of connection is to being not added in line set.The present invention enters target, leave position under camera field as opening up
Flutter the node of structure, and using mixed Gauss model to the entrance of target, leave position and be modeled, disappeared-occurred section
Point set V.This pair of connectedness of node is judged using the Intensity correlation function of node pair, by the connectedness for counting all nodes pair
Construction line set E.For the node pair for connecting, during by the transfer for standardizing this pair of Intensity correlation function calculate node pair of node
Between be distributed.In the Intensity correlation function of calculate node pair, it is contemplated that disappearance observation in this pair of node with there is observation
Joint appearance similarity, including domain color similarity, texture similarity and target conspicuousness similarity, so as to improve utilization
The Intensity correlation function of node pair judges this pair of connective degree of accuracy of node.For the node pair for connecting, using this pair of node
Interactive information represent the transition probability of node pair, the transition probability of other all nodes pair is 0.According to all nodes to turning
Move probability construction weight set W.Connective matrix is constructed according to line set E and weight set W, and monitoring is inferred by the matrix
The topological structure of network.When certain is probably " falseness connection " to node in topological structure, this pair of node is considered
Interactive information and transfer time distribution excluding " falseness connection ".Using topology adaptation update policy update topological structure and
Parameter, while proposing the quick response strategy to addition/remove camera situation so that the topological structure for learning can be compared with
The change in good adaptation monitoring bad border.
Concrete operation step of the invention
1) set up disappear-there is node set
Movement locus of the target under the different cameras ken is obtained using ripe camshift trackings, statistics is every
Target enters, leaves the position of the ken under the individual ken, and these positions are modeled using mixed Gauss model GMM, obtains
To the disappearance-node set V occur of topological structure.GMM parameters are Z={ z, λz,μz,σz 2}.Wherein z represents single Gauss in GMM
Disappear-occur the number of node under the number of distribution, i.e. the current camera ken.λzRepresent the weight of each Gauss, μz,σz 2For
The average and variance of corresponding Gaussian Profile.The number z of the gauss component in GMM is by bayesian information criterion (Bayesian
Information Criterion) automatically determine, other parameters { λz,μz,σz 2Pass through expectation maximization method
(Expectation-Maximization methods) is calculated.Disappearance-the appearance of the used monitoring network of present invention experiment
Node schematic diagram is as shown in Figure 2.
2) line set of topological structure is calculated
1. the Intensity correlation function of calculate nodes pair
The present invention adaptively judges whether certain connects to node using Intensity correlation function.Assuming that t disappears in node i
The target observation value sequence of mistake is XiT (), the target observation value sequence for mutually occurring in node j in the same time is Yj(t), then node i
Intensity correlation function with node j can be defined as
Wherein | | Xi(t) | | and | | Yj(t+T) | | X is represented respectivelyi(t) and YjThe number of observation in (t), when T is to postpone
Between.The disappearance observation in two nodes is only only accounted for due to Intensity correlation function and observation occurs, do not consider to disappear
Observation with occur whether observation belongs to same target, therefore comprising larger error.In order to improve using the mutual of node pair
The connective degree of accuracy of correlation function decision node, the present invention with the addition of joint appearance similarity in Intensity correlation function.This hair
The Intensity correlation function of bright proposed improved node i and node j is defined as follows：
s.t.Psim(Oa,i,Ob,j) ＞ δ
Wherein, Oa,iWith Ob,jMissing object sequence of observations X is represented respectivelyi(t) and there is target observation value sequence Yj(t+
T the observation in), Psim(Oa,i,Ob,j) represent observation Oa,iWith Ob,jJoint appearance similarity.Only work as Oa,iWith Ob,j's
When similarity is more than given threshold value δ, could be by observation Oa,iWith Ob,jIntensity correlation function is added as association observation, it is ensured that
The observation used in Intensity correlation function is to there is association higher.It is illustrated in figure 3 and is calculated using method proposed by the invention
The connection node pair for obtaining and the Intensity correlation function curve for not connecting node pair.In order to embody the superiority of the inventive method,
Simultaneously give without characteristic information Intensity correlation function curve and only add color similarity Intensity correlation function curve
Compare.Wherein (a), (b), (c) are the Intensity correlation function curve for connecting node pair, and (d), (e), (f) are not connect node pair
Intensity correlation function curve.
2. calculates joint appearance similarity
In order to the Intensity correlation function improved using node pair judges this pair of degree of accuracy of Connectivity, the present invention is being calculated
During the Intensity correlation function of node pair, it is contemplated that in this pair of node disappearance observation with there is observation combine appearance similarity,
So as to ensure that the observation for only having appearance similarity high in Intensity correlation function can just be associated.Outside joint proposed by the present invention
Table similarity includes domain color similarity, texture similarity and target conspicuousness similarity.Assuming that for any observation Oa,i
And Ob,j, their joint appearance similarity can be expressed as
Psim(Oa,i,Ob,j)=P (Oa,i,Ob,j|Oa=Ob)
=P (colora,i,colorb,j|Oa=Ob)
P(texa,i,texb,j|Oa=Ob)P(doma,i,domb,j|Oa=Ob)
=PcolorPtexPdom
Wherein, Oa=ObRepresent observation Oa,iAnd Ob,jBelong to same target.Separately below in introduction joint appearance similarity
Domain color similarity, texture similarity and target conspicuousness similarity computational methods：
For observation Oa,iAnd Ob,jDomain color similarity P (colora,i,colorb,j|Oa=Ob), the present invention is used and divided
The method of block weighting is calculated.It is people in view of the monitored object in this monitors environment, information will rule of thumb can observes
Value is divided into head, trunk, the part of leg three.For any observation, extracted using K mean cluster method respectively in each part
Primary color histogram.Two domain colors of observation corresponding part are calculated using the Bhattacharya distances of primary color histogram
Similarity.Hypothetic observation Oa,iAnd Ob,jHead, trunk, the domain color similarity of leg be expressed as Phead, PbodyAnd
Plegs, then observation Oa,iAnd Ob,jDomain color similarity can be expressed as
PColor=α Phead+βPbody+γPlegs
Wherein, α, beta, gamma is respectively head, trunk, the corresponding weight in leg, it is necessary to meet alpha+beta+γ=1.
For observation Oa,iAnd Ob,jTexture similarity P (texa,i,texb,j|Oa=Ob), the present invention is using observation
Gradient orientation histogram (HOG) feature is calculated.For observation Oa,iWith Ob,j, their HOG features and table are extracted respectively
It is shown as Ha,iAnd Hb,j.So, observation Oa,iAnd Ob,jTexture similarity PtexCan be using Ha,iAnd Hb,j's
Bhattacharya distances are represented, are defined as follows：
Ptex=exp (- DB/σt)
Wherein, σtIt is predefined bandwidth, DBRepresent Ha,iAnd Hb,jBhattacharya distances, kHOGIt is HOG features
Dimension.
For observation Oa,iAnd Ob,jConspicuousness similarity P (doma,i,domb,j|Oa=Ob), the present invention uses observation
Significant characteristics calculated.When the significant characteristics of observation are extracted, observation is divided into fritter first, using k most
Near neighbor method (KNN) automatically identifies maximum with other fritter distinctivenesses fritter in all fritters, used as the aobvious of the observation
Work property feature.Two conspicuousness similarity P of observationdomBy calculate the two observations significant characteristics Euclidean away from
From obtaining.The significant characteristics that are extracted of the present invention have good Shandong to attitudes vibration, lighting change of environment of target etc.
Rod.
3. calculates the line set of topological structure
The present invention judges this pair of connectedness of node using the Intensity correlation function of node pair, so as to calculate topological structure
Line set E.The present invention only considers the connectedness across ken node, and the connection situation of identical ken lower node is not considered, therefore
The node connected under the identical ken is to being not added in line set.Node is to the connective specific practice for judging：By node
To Intensity correlation function curve peak value and given threshold value thr be compared (judge Intensity correlation function curve peak value whether
Substantially), if the peak value of Intensity correlation function curve is more than thr, it is believed that this is connection to node.Otherwise it is assumed that this is to node
Do not connect.Threshold value thr can be calculated by the Intensity correlation function of node pair, be defined as follows：
Thr=mean (Rij(T))+ω·std(Rij(T))
Wherein ω is User Defined parameter.
3) the transfer time distribution of connection node pair is calculated
Node pair for connecting is distributed, it is necessary to calculate this pair of transfer time of node, used as monitoring network topology structure
Time-constrain.The transfer time distribution of node pair can be realized by standardizing the Intensity correlation function of the node pair.Assuming that section
Intensity correlation function of the point between i and j is Ri,j(T), then the transfer time distribution between the two nodes can be defined as
Pi,j(T)=Ri,j(T)/||Ri,j(T)||
The present invention represents transfer time distribution using Gauss model, therefore can be distributed transfer time and be expressed as one
The form of Gaussian Profile, it is as follows：
Pi,j(T)~N (μ, σ2)
Wherein μ, σ2The average and variance of Gauss transfer time model are represented respectively, can estimate to obtain by EM methods.
4) the weight set of topological structure is calculated
The present invention represents the transition probability of the node pair using the interactive information of node pair.Interactive information represents two sequences
Degree of dependence between row, it is assumed that the sequence of observations of node i is XiT (), the sequence of observations of node j is Yj(t+T), this two
The interactive information I (X, Y) of the individual sequence of observations is defined as follows:
Wherein, ρX,Y 2Represent sequence Xi(t) and Yj(t+T) the mutual correlation coefficient between, can be by the mutual of the two sequences
Correlation function Ri,j(T) it is calculated, is defined as follows：
Wherein TpeakThe value of time delay T when representing that Intensity correlation function gets most obvious peak value, denominator is respectively Xi(t)、Yj
(t+T) covariance of sequence.
For the node pair for connecting, the transition probability between node is equal to their interactive information value.For disconnected section
Point is right, and the transition probability between node is 0.For the node under the identical ken to (non-across ken node to), their transfer is general
Rate is 0.Transition probability according to all nodes pair can construct the weight set W of topological structure.
5) construct connective matrix and infer topological structure
Connective matrix can be constructed according to line set E and weight set W, monitoring network can be inferred that according to the matrix
Topological structure.The value of the m rows n row of connective matrix represents that target is general to the transfer of n-th node from m-th joint movements
Rate, the relevant position of disconnected node pair is 0.Under identical camera field, all nodes are to the correspondence in connective matrix
Value at position is 0.After obtaining connective matrix, for any row, the nonzero element inquired about in the row it is known that the row
Corresponding node can lead to which node (nonzero element arranges corresponding node).Be illustrated in figure 4 connective matrix and according to
The topological structure schematic diagram that this connective matrix is inferred to.
6) exclude " falseness connection "
" falseness connection " refers to inferring certain for obtaining to connection node, actual and non-immediate company by connective matrix
It is logical, but by other node indirect communications.After the disappearance of certain node destination, generally require to know that the target goes out next time
Which node is now likely located at, and the node where these nodes disappear with target is typically what is directly connected.If be inferred to
Topological structure in there is more " falseness connection " to exist, the complexity that target is recognized again can be increased, while reducing what is recognized again
Accuracy rate.Accordingly, it would be desirable to " falseness connection " in effectively removing topological structure.The present invention considers the friendship of connection node pair
Whether mutual information is distributed with transfer time, so as to judge the node to being " falseness connection ".
Relevance the interactive information size of connection node can reflect egress is strong and weak.Connect the interaction of node pair
Information is larger, represents between the two nodes there is very strong relevance, it is believed that be directly connection.Conversely, node pair
Interactive information is smaller, represents that the relevance between the two nodes is weaker, it is believed that the two nodes are by other nodes
Connection is realized, is belonged to " falseness connection ".
For cannot using interactive information judge whether be " falseness connection " node pair, it is possible to use the node is to turning
Shift time distribution is determined whether.For a pair of node i → j for being probably " falseness connection ", certainly existed between them
One path of non-" falseness connection " (k ..., m) so that can be connected by this paths between i → j, i.e. i → k →...→ m
→j.If there is no the path for meeting condition, then to i, j must not be " falseness connection " to node.If there is meeting condition
Path, it is possible to use the transfer time distribution of i → j carrys out decision node to i with the distribution of transfer time of i → k →...→ m → j
Whether → j is " falseness connection ".Because arbitrary node is known to the Annual distribution for meeting and obedience Gauss in topological structure
Distribution, then N (μ are obeyed according to the transfer time distribution that Gauss Adding law can obtain i → k →...→ m → jik+…+μmj,
σik 2+…+σmj 2).Meanwhile, N (μ are obeyed in transfer time distribution of the node to i → jij, σij 2).Calculate N (μik+…+μmj, σik 2+
σmj 2) and N (μij, σij 2) Kullback-Leibler divergences, and result is compared with predetermined threshold value th.If divergence is small
In th, then the transfer time distribution of two kinds of connection situations is similar, it is believed that i → j is indeed through i → k →...→ m → j realities
Existing, therefore node is " falseness connection " to i → j.Otherwise it is assumed that node is not " falseness connection " to i → j.It is illustrated in figure 5
The new connective matrix that the connective matrix of Fig. 4 carries out being obtained after " falseness is connected " excludes with topological structure is tied with topology
Structure, the topological structure shown in Fig. 5 truly reflects the connection situation of Fig. 2 Scenes.
7) topology adaptation updates
1. transfer time distributed updates
This pair of transfer time of node that updated by the time between connection node is distributed at certain using target.Assuming that tk
For k-th target at certain, by the time, N (μ, σ are obeyed in this pair of current transfer time distribution of node between connection node2), ρ is
Turnover rate, then the parameter renewal process of Annual distribution model is defined as below
μ*=(1- ρ) μ+ρ tk
σ*2=(1- ρ) σ2+ρ(μ*-tk)2
Be illustrated in figure 6 the schematic diagram of transfer time distributed update, transfer time be distributed by previous moment N (13.2,
3.2) N (17.8,3.85) is become.Less, essential difference is embodied in average the variance difference of the two Gaussian Profiles, Qian Zheping
The 17.8s of the latter is shorter than by time 13.2s, the movement velocity of target is very fast before showing.
2. transition probabilities update
The transition probability of the two nodes is updated using transfer case of the target between two nodes, it is assumed that for two nodes
I and j, turnover rate is κ, then the renewal of transition probability can be expressed as
wij *=(1- κ) wij+κPij(Tw)
Wherein Pij(Tw) represent in TwFrom node i to the target numbers of node j and the number of targets left from node i in time
Purpose ratio, TwA hour typically is taken as, i.e. each hour updates a transition probability.It is illustrated in figure 7 transfer general
The schematic diagram that rate updates, above the transition probability of three hours is increased slightly in figure, and transition probability is tapered into afterwards.
3. add/remove camera quick response
When the video camera configuration in monitors environment changes, it is necessary to readjust topology network architecture.Taken the photograph for removal
The situation of camera, it is only necessary to disconnect the pass of the node under node under original other camera fields and removed camera field
Connection.When increasing video camera, increased camera field is relearned by learning identical method with initial topology
Under connected relation between node, and these nodes and existing node of disappearance-occur, transfer time distribution and shift
Probability, and topological structure is updated accordingly.
With above-mentioned according to desirable embodiment of the invention as enlightenment, by above-mentioned description, relevant staff is complete
Various changes and amendments can be carried out without departing from the scope of the technological thought of the present invention' entirely.The technology of this invention
Property scope is not limited to the content on specification, it is necessary to its technical scope is determined according to right.
Claims (9)
1. a kind of non-overlapping visual field multiple-camera monitors network topology adaptation learning method, it is characterised in that had using weighting
Represented to figure non-overlapping visual field multiple-camera monitor network topological structure, weighted digraph by node set, line set with
And weight set is constituted；
Target under each camera field is entered, the position of the ken is left as topological node, is defined as disappearing-occurring
Node；
Target enters, leaves the position of the ken under counting each ken using haplopia domain method for tracking target, and uses mixing
The disappearance of Gauss model construction topological structure-node set occur；
Judge this pair of connectedness of node using the Intensity correlation function of across ken node pair, across the ken node is to referring to two sections
, respectively from different camera fields, across the ken node of all connections is to constituting line set for point；
When Intensity correlation function is calculated, it is considered to node centering disappearance observation with there is observation combine appearance similarity, bag
Domain color similarity, texture similarity and target conspicuousness similarity are included, for across the ken node pair for connecting, by standard
The Intensity correlation function for changing this pair of node calculates the distribution of its transfer time；
The interactive information for introducing node pair represents this pair of transition probability of node, so as to construct weight set；
Connective matrix is constructed according to the line set tried to achieve and weight set, and the topology knot of monitoring network is inferred by the matrix
Structure；
Judge whether this pair of node belongs to " falseness connection " using certain interactive information to node in topological structure, for being probably
The situation of " falseness connection ", " falseness connection " situation is excluded further with transfer time distribution；
Policy update topological structure and parameter are updated using topology adaptation, wherein parameter includes that transfer time is distributed and shifts
Probability；
Quick response strategy is performed to addition/removal camera situation simultaneously.
2. non-overlapping visual field multiple-camera according to claim 1 monitors network topology adaptation learning method, its feature
It is that the node set of topological structure is constructed using mixed Gauss model, key step is as follows：
A1) target enters, leaves the determination of camera field position
The movement locus of target is obtained using the method for tracking target under haplopia domain, certain time under certain camera field is obtained
After the movement locus of interior all targets, it is possible to count the entrance of target under the ken, position left, under the haplopia domain
Method for tracking target is：Camshift trackings；
A2) disappear-occur the determination of node
Being entered using target, leave the position of camera field as topological node, definition disappears-there is node：
For certain camera field, the position that target enters, leaves the ken is obtained using the method for tracking target under haplopia domain
Postpone, mixed Gauss model GMM is used based on positional information, construct the disappearance-node set occur of topological structure, wherein, GMM
Parameter be Z={ z, λz,μz,σz 2, wherein z represents the number of single Gaussian Profile in GMM, i.e., disappear under the current camera ken
Lose-occur the number of node, λzRepresent the weight of each Gauss, μz,σz 2It is the average and variance of corresponding Gaussian Profile, z is logical
Cross what bayesian information criterion was automatically determined, other parameters { λz,μz,σz 2Be calculated by expectation maximization method.
3. non-overlapping visual field multiple-camera according to claim 2 monitors network topology adaptation learning method, its feature
It is, the step a1) the middle target following carried out under haplopia domain using camshift trackings, determine the motion rail of target
Mark.
4. non-overlapping visual field multiple-camera according to claim 1 monitors network topology adaptation learning method, its feature
It is to construct the line set of topological structure using the Intensity correlation function of across ken node pair, key step is as follows：
B1) Intensity correlation function of across ken node pair is calculated
Using the disappearance-node set occur of mixed Gauss model GMM construction topological structures, using in node set to mutual pass
Connection function judges whether this pair of node connects；Assuming that t is X in the target observation sequence that node i disappearsi(t), mutually in the same time
The target observation sequence occurred in node j is Yj(t), then the Intensity correlation function of node i and j is defined as
Wherein | | Xi(t) | | and | | Yj(t+T) | | X is represented respectivelyi(t) and YjT the observation number of (), T is time delay；In order to
Intensity correlation function decision node is improved to the connective degree of accuracy, joint appearance similarity is with the addition of in Intensity correlation function, it is fixed
Justice is as follows：
s.t.Psim(Oa,i,Ob,j) ＞ δ
Wherein, Oa,iAnd Ob,jMissing object sequence X is represented respectivelyi(t) and there is target sequence Yj(t+T) observation, Psim
(OA, i, OB, j) two joint appearance similarities of observation are represented, only could be by when similarity is more than given threshold value δ
It adds Intensity correlation function, it is ensured that the observation for being used for associating in Intensity correlation function has probability higher to belong to identical mesh
Mark；
B2 the calculating of appearance similarity) is combined
Joint appearance similarity includes domain color similarity, texture similarity and target conspicuousness similarity, is easy to various
The combinations of features of extraction is used；
Assuming that there is two observation OA, iAnd OB, j, their joint appearance similarity is expressed as
Psim(OA, i, OB, j)=P (OA, i, OB, j|Oa=Ob)
P(colorA, i, colorB, j|Oa=Ob)×
P(texA, i, texB, j|Oa=Ob)×P(domA, i, domB, j|Oa=Ob)
=Pcolor×Ptex×Pdom
Wherein Oa=ObRepresent that two observations belong to same target；
For domain color similarity Pcolor, calculated using the method for divided group, it is contemplated that the monitoring in this monitors environment
Pair as if people, rule of thumb information observation is divided into head, trunk, the part of leg three, for any observation, in each portion
Divide and extract primary color histogram using K mean cluster method respectively, it is right that two some portion of domain color similarities of observation are used
The Bhattacharya distances of the primary color histogram of part are answered to represent, it is assumed that observation OA, iAnd OB, jHead, trunk, leg
The domain color similarity in portion is expressed as Phead, PbodyAnd Plegs, then observation OA, iAnd OB, jDomain color similarity table
It is shown as
PColor=α Phead+βPbody+γPlegs
Wherein, α, beta, gamma is respectively head, trunk, the corresponding weight in leg, it is necessary to meet alpha+beta+γ=1；
For texture similarity Ptex, the present invention is calculated using the gradient orientation histogram HOG features of observation, for seeing
Measured value OA, iWith OB, j, their HOG features are extracted respectively, it is expressed as HA, iAnd HB, j, then, observation OA, iAnd OB, jLine
Reason similarity PtexJust use HA, iAnd HB, jBhattacharya distance represent, be defined as follows：
Ptex=exp (- DB/σt)
Wherein, σtIt is predefined bandwidth, DBRepresent that HOG describes sub- Ha,iAnd Hb,jBhattacharya distances, kHOGIt is that HOG is retouched
State the dimension of son；
For the conspicuousness similarity P of targetdom, calculated using the significant characteristics of observation, extracting the aobvious of observation
During work property feature, observation is divided into fritter first, using k arest neighbors methods KNN automatically identify in all fritters with other
The maximum fritter of fritter distinctiveness, as the significant characteristics of the observation, two conspicuousness similarity P of observationdomPass through
The Euclidean distance for calculating their significant characteristics is obtained；
B3) across ken node is judged connectedness
This pair of connectedness of node is judged to the peak value of Intensity correlation function curve using node, following steps are specifically included：Will
Node is compared to the peak value of Intensity correlation function curve with threshold value thr, i.e. peak value of the decision node to Intensity correlation function curve
Whether substantially, if the peak value of Intensity correlation function curve is more than threshold value thr, it is believed that this is connection, the threshold value thr to node
Can be calculated by the Intensity correlation function of node pair, be defined as follows：
Thr=mean (Rij(T))+ω·std(Rij(T))
Wherein ω is User Defined parameter.
5. non-overlapping visual field multiple-camera according to claim 1 monitors network topology adaptation learning method, its feature
It is that the transfer time of computer connection node pair is distributed as time-constrain, is comprised the following steps that：
Node pair for connecting is distributed, it is necessary to calculate its transfer time, as the time-constrain of monitoring network topology structure,
This pair of transfer time distribution of node is calculated by the Intensity correlation function for standardizing node pair；Assuming that the mutual pass between node i and j
Connection function is Ri,j(T), then the transfer time distribution between the two nodes is defined as
Pi,j(T)=Ri,j(T)/||Ri,j(T)||
The present invention represents the transfer time distribution of node pair using Gauss model, as follows：
Pi,j(T):N(μ,σ2)
Wherein μ, σ2The average and variance of Gauss model are represented respectively, estimate to obtain by EM methods.
6. non-overlapping visual field multiple-camera according to claim 1 monitors network topology adaptation learning method, its feature
It is that the weight set of topological structure is constructed using interactive information, comprises the following steps：
The transition probability of node pair is represented using the interactive information of node pair, the transition probability construction weight sets according to node pair
Close；Interactive information represents the degree of dependence between two sequences, it is assumed that the sequence of observations of node i is Xi(t), the sight of node j
Measured value sequence is Yj(t+T), the interactive information I (X, Y) of the two sequence of observations is defined as follows:
Wherein, ρX,Y 2Represent X between two sequencesi(t) and Yj(t+T) mutual correlation coefficient, by the mutual correlation of the two sequences
Function Ri,j(T) it is calculated, is defined as follows：
Wherein TpeakThe value of time delay T when representing that Intensity correlation function gets most obvious peak value, denominator is respectively Xi(t)、Yj(t+
T) the covariance of sequence, for the node pair for connecting, the transition probability between node is equal to their interactive information value, for not connecting
Transition probability between logical node pair, node is 0, for the identical ken under all nodes pair, i.e., non-across ken node pair recognizes
For their transition probability is 0.
7. non-overlapping visual field multiple-camera according to claim 1 monitors network topology adaptation learning method, its feature
It is that the topological structure of monitoring network is inferred using connective matrix, comprises the following steps：
Connective matrix is constructed using line set and weight set, and the topological structure of monitoring network is inferred by the matrix；Even
The value of the m rows n row of general character matrix represents target from m-th joint movements to n-th transition probability of node, for not connecting
Node pair, the node number according to node pair sets to 0 in the relevant position of connective matrix, under identical camera field, Suo Youjie
Point is 0 to the value in the corresponding position of connective matrix, after obtaining connective matrix, has a correspondence for arbitrary node
OK, other nodes connected with the node can just be obtained by finding the nonzero element of the row.
8. non-overlapping visual field multiple-camera according to claim 1 monitors network topology adaptation learning method, its feature
It is the topological structure for using " falseness connection " Solve Problem to improve monitoring network, comprises the following steps：
" falseness connection " refers to certain being inferred to by connective matrix to connection node, actual and non-immediate connection, but passes through
Other node indirect communications；When target disappears in certain node, it is necessary to know that the target may occur in which node next time,
And the node that these nodes disappear with target should be connected directly, interactive information and the transfer of connection node pair are considered
Annual distribution, so as to judge the node to whether being " falseness connection "；
The interactive information for connecting node pair is larger, represents between the two nodes there is very strong relevance, it is believed that be directly connection
's；Conversely, the interactive information of node pair is smaller, represent that the relevance between the two nodes is weaker, it is believed that the two nodes are
Connection is realized by other nodes, is belonged to " falseness connection "；
For cannot using interactive information judge whether be " falseness connection " node pair, using the transfer time point of the node pair
Cloth is determined whether, for a pair of node i → j for being probably " falseness is connected ", a non-" void is certainly existed between them
Vacation connection " path (k ..., m) so that can be connected by this paths between i → j, i.e. i → k →...→ m → j；If
In the absence of the path for meeting condition, then to i, j must not be " falseness connection " to node, if there is the path for meeting condition,
Then using transfer time distribution and transfer time of i → k →...→ m → j of i → j be distributed come decision node to i → j whether be
" falseness connection "；N (μ are obeyed according to the transfer time distribution that Gauss Adding law obtains i → k →...→ m → jik+…+μmj,
σik 2+…+σmj 2)；Meanwhile, N (μ are obeyed in transfer time distribution of the node to i → jij, σij 2)；Calculate N (μik+…+μmj, σik 2+
σmj 2) and N (μij, σij 2) Kullback-Leibler divergences, and result is compared with predetermined threshold value th, if divergence is small
In th, then the transfer time distribution of two kinds of connection situations is similar, it is believed that i → j is indeed through i → k →...→ m → j realities
Existing, therefore node is " falseness connection " to i → j, otherwise it is assumed that node is not " falseness connection " to i → j.
9. non-overlapping visual field multiple-camera according to claim 1 monitors network topology adaptation learning method, its feature
It is that updating policy update using topology adaptation monitors network topology structure, comprises the following steps：
D1) transfer time distributed update strategy
This pair of transfer time of node that updated by the time between connection node is distributed at certain using target；Assuming that tkIt is kth
Individual target is at certain, by the time, N (μ, σ are obeyed in this pair of current transfer time distribution of node between connection node2), ρ is renewal
Rate, then the parameter renewal process of Annual distribution model is defined as below：
μ*=(1- ρ) μ+ρ tk
σ*2=(1- ρ) σ2+ρ(μ*-tk)2
D2) transition probability more new strategy
The transition probability of the two nodes is updated using transfer case of the target between two nodes, it is assumed that for two node is and
J, turnover rate is κ, then the renewal of transition probability can be expressed as
wij *=(1- κ) wij+κPij(Tw)
Wherein Pij(Tw) represent in TwFrom node i to the target numbers of node j and the target numbers left from node i in time
Ratio, TwA hour is taken as, i.e. each hour updates a transition probability；
D3) addition/removal camera
When the video camera configuration in monitors environment changes, it is necessary to readjust topology network architecture, for removing video camera
Situation, it is only necessary to disconnect associating for node under original other camera fields and the node under removed camera field,
When increasing video camera, relearned under increased camera field by learning identical method with initial topology
Disappear-occur connected relation between node, and these nodes and existing node, transfer time distribution and transfer is general
Rate, and topological structure is updated accordingly.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410266226.6A CN104010168B (en) | 2014-06-13 | 2014-06-13 | A kind of non-overlapping visual field multiple-camera monitors network topology adaptation learning method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410266226.6A CN104010168B (en) | 2014-06-13 | 2014-06-13 | A kind of non-overlapping visual field multiple-camera monitors network topology adaptation learning method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104010168A CN104010168A (en) | 2014-08-27 |
| CN104010168B true CN104010168B (en) | 2017-06-13 |
Family
ID=51370644
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410266226.6A Expired - Fee Related CN104010168B (en) | 2014-06-13 | 2014-06-13 | A kind of non-overlapping visual field multiple-camera monitors network topology adaptation learning method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104010168B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104599290B (en) * | 2015-01-19 | 2017-05-10 | 苏州经贸职业技术学院 | Video sensing node-oriented target detection method |
| CN105744223B (en) * | 2016-02-04 | 2019-01-29 | 北京旷视科技有限公司 | Video data handling procedure and device |
| CN106846378B (en) * | 2017-01-23 | 2019-12-03 | 中山大学 | A kind of across the video camera object matching and tracking of the estimation of combination topology of spacetime |
| CN107016374A (en) * | 2017-04-12 | 2017-08-04 | 电子科技大学 | Intelligent Measurement tracking and the generation method of space-time track towards specific objective |
| CN107689054B (en) * | 2017-08-24 | 2020-09-22 | 北京航空航天大学 | Multi-camera topological connectivity graph establishing and cross-camera target tracking method |
| CN107705327B (en) * | 2017-08-29 | 2020-08-11 | 电子科技大学 | Candidate target extraction method of multi-camera network space-time model |
| CN108614889B (en) * | 2018-05-04 | 2020-02-14 | 济南大学 | Moving object continuous k nearest neighbor query method and system based on Gaussian mixture model |
| CN110251076B (en) * | 2019-06-21 | 2021-10-22 | 安徽大学 | Method and device for detecting significance based on contrast and fusing visual attention |
| CN113313188B (en) * | 2021-06-10 | 2022-04-12 | 四川大学 | Cross-modal fusion target tracking method |
| CN114859300B (en) * | 2022-07-07 | 2022-10-04 | 中国人民解放军国防科技大学 | Radar radiation source data stream processing method based on graph connectivity |
| CN115578694A (en) * | 2022-11-18 | 2023-01-06 | 合肥英特灵达信息技术有限公司 | Video analysis computing power scheduling method, system, electronic equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5714997A (en) * | 1995-01-06 | 1998-02-03 | Anderson; David P. | Virtual reality television system |
| US6661918B1 (en) * | 1998-12-04 | 2003-12-09 | Interval Research Corporation | Background estimation and segmentation based on range and color |
| CN1738426A (en) * | 2005-09-09 | 2006-02-22 | 南京大学 | Video motion goal division and track method |
| CN101174327A (en) * | 2007-11-06 | 2008-05-07 | 北京邮电大学 | Implementing method for video sensor network covering and enhancing movement path of movable objective |
| CN101616309A (en) * | 2009-07-16 | 2009-12-30 | 上海交通大学 | Non-overlapping visual field multiple-camera human body target tracking method |
| CN101639935A (en) * | 2009-09-07 | 2010-02-03 | 南京理工大学 | Digital human serial section image segmentation method based on geometric active contour target tracking |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1657927A1 (en) * | 2004-11-12 | 2006-05-17 | Saab Ab | Image-based movement tracking |
-
2014
- 2014-06-13 CN CN201410266226.6A patent/CN104010168B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5714997A (en) * | 1995-01-06 | 1998-02-03 | Anderson; David P. | Virtual reality television system |
| US6661918B1 (en) * | 1998-12-04 | 2003-12-09 | Interval Research Corporation | Background estimation and segmentation based on range and color |
| CN1738426A (en) * | 2005-09-09 | 2006-02-22 | 南京大学 | Video motion goal division and track method |
| CN101174327A (en) * | 2007-11-06 | 2008-05-07 | 北京邮电大学 | Implementing method for video sensor network covering and enhancing movement path of movable objective |
| CN101616309A (en) * | 2009-07-16 | 2009-12-30 | 上海交通大学 | Non-overlapping visual field multiple-camera human body target tracking method |
| CN101639935A (en) * | 2009-09-07 | 2010-02-03 | 南京理工大学 | Digital human serial section image segmentation method based on geometric active contour target tracking |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104010168A (en) | 2014-08-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104010168B (en) | A kind of non-overlapping visual field multiple-camera monitors network topology adaptation learning method | |
| Sam et al. | Divide and grow: Capturing huge diversity in crowd images with incrementally growing cnn | |
| CN102436662B (en) | Human body target tracking method in nonoverlapping vision field multi-camera network | |
| CN106846378B (en) | A kind of across the video camera object matching and tracking of the estimation of combination topology of spacetime | |
| CN110414400A (en) | A kind of construction site safety cap wearing automatic testing method and system | |
| Chen et al. | Object tracking across non-overlapping views by learning inter-camera transfer models | |
| CN109508671A (en) | A kind of video accident detection system and method based on Weakly supervised study | |
| CN104717468B (en) | Cluster scene intelligent monitoring method and system based on the classification of cluster track | |
| CN103281477A (en) | Multi-level characteristic data association-based multi-target visual tracking method | |
| CN103390278A (en) | Detecting system for video aberrant behavior | |
| CN110427834A (en) | A kind of Activity recognition system and method based on skeleton data | |
| CN103971384B (en) | Node cooperation target tracking method of wireless video sensor | |
| CN103390151B (en) | Method for detecting human face and device | |
| CN109389044A (en) | More scene crowd density estimation methods based on convolutional network and multi-task learning | |
| CN109918995B (en) | Crowd abnormity detection method based on deep learning | |
| CN104065932B (en) | A kind of non-overlapping visual field target matching method based on amendment weighting bigraph (bipartite graph) | |
| CN108537181A (en) | A kind of gait recognition method based on the study of big spacing depth measure | |
| CN107146237A (en) | A kind of method for tracking target learnt based on presence with estimating | |
| CN103150552B (en) | A kind of driving training management method based on number of people counting | |
| CN108830882A (en) | Video abnormal behaviour real-time detection method | |
| CN108093213A (en) | A kind of target trajectory Fuzzy Data Fusion method based on video monitoring | |
| Barkhade et al. | Flood extent mapping with unmanned aerial vehicles data using deep convolutional neural network | |
| CN106991400A (en) | A kind of fire hazard smoke detecting method and device | |
| Wang et al. | Correlated time series self-supervised representation learning via spatiotemporal bootstrapping | |
| CN110334743A (en) | A kind of progressive transfer learning method based on the long memory network in short-term of convolution |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170613 Termination date: 20200613 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |