CN104010168A - Non-overlapping vision field multi-camera monitoring network topology self-adaptation learning method - Google Patents
Non-overlapping vision field multi-camera monitoring network topology self-adaptation learning method Download PDFInfo
- Publication number
- CN104010168A CN104010168A CN201410266226.6A CN201410266226A CN104010168A CN 104010168 A CN104010168 A CN 104010168A CN 201410266226 A CN201410266226 A CN 201410266226A CN 104010168 A CN104010168 A CN 104010168A
- Authority
- CN
- China
- Prior art keywords
- node
- target
- correlation function
- ken
- communicated
- 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
Abstract
The invention provides a non-overlapping vision field multi-camera monitoring network topology self-adaptation learning method, and relates to the field of computer vision. A weighted directed graph G=<V, E and W> is used, and the topology of a monitoring network is represented. According to the non-overlapping vision field multi-camera monitoring network topology self-adaptation learning method, the leaving position and the entering position of a target in a single-camera vision field are used as topological nodes V, and a Gaussian mixture model is utilized for modeling. The cross-correlation function computing method based on united surface similarity is provided, the connectivity of a certain pair of nodes is judged through a cross-correlation function, and therefore an edge set E is obtained. As for the connected node pair, transfer time distribution is calculated through the standardization cross-correlation function. Mutual information of the node pair is utilized for representing the transfer probability of the nodes, and therefore the weight set W is obtained. According to the non-overlapping vision field multi-camera monitoring network topology self-adaptation learning method, the false connection removal strategy is provided for removing probable false connection in the topology, the topology self-adaptation updating strategy is provided for ensuring the higher robustness of the topological structure to environmental changes.
Description
Technical field
The invention belongs to computer vision field, be specifically related to field of intelligent monitoring, particularly a kind of non-overlapping visual field multiple-camera monitor network topology adaptation learning method.
Background technology
Along with the development of camera supervised technology, extensive area is monitored and become the important means assuring the safety for life and property of the people.But, for the larger monitoring occasion in region, use video camera to cover all guarded regions very unrealistic.Therefore the method that, conventionally adopts key area to cover is built the multiple-camera supervisory control system that comprises non-overlapping visual field.Compared with traditional single camera supervisory control system or overlapping ken multiple-camera supervisory control system, non-overlapping visual field multiple-camera supervisory control system is because its observed object is all discrete on time, space, so it is more difficult that target is carried out to Continuous Tracking.Common processing method is the topological structure of study camera network, and utilizes the space-time restriction information that topological structure provides to help mate the target under the different kens, realizes the Continuous Tracking to target.
Non-overlapping visual field multiple-camera topology of networks learning method is mainly divided into two kinds of supervision formula and non-supervisory formulas.Supervision formula learning method is followed the tracks of by the target to artificial mark, thereby the topological structure of study camera network, modal as the space-time restriction information of utilizing Parzen window learning network topological structure of O.Javed, 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 document " 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 supervision formula learning method needs a large amount of data of mark conventionally, and not strong for the variation adaptive capacity of monitoring environment in actual use, is therefore difficult to be applied to actual monitored system.
Non-supervisory formula learning method does not need artificial flag data, system can according to data adaptive that in monitor network, each child node detects learn the topological structure of supervisory control system.Whether non-supervisory formula learning method adopts Intensity correlation function to infer between certain two node to be communicated with conventionally, 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 only comprises the information that two nodes disappear and occur at certain section of time internal object, the similarity information that does not comprise target, therefore error is larger.Having afterwards scholar to propose two kinds improves one's methods: colouring information is fused in Intensity correlation function, 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 ", effectively improve Intensity correlation function by the colouring information of interpolation target and judged connective ability, meanwhile, utilize connective matrix can infer the topological structure of monitor network.On the basis of color characteristic, to utilize the identity information that face detection obtains to be fused in Intensity correlation function, 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 adding, greatly improve and only adopted the Intensity correlation function of colouring information to judge connective ability.Accurately obtaining under the prerequisite of identity information, the Intensity correlation function that the method calculates can obtain the most connective judgement.But face detection all there are certain requirements for imaging precision and angle, be not suitable for promoting.
Therefore, utilize the topological structure of unsupervised learning method adaptive learning multiple-camera monitor network, can be applied to actual most important for multiple-camera supervisory control system.In without supervision topology learning method, mainly containing at following 3 should be noted that: (1) is in the time being used node right Intensity correlation function to judge this to node connective, how in Intensity correlation function, to be added with representativeness and to be easy to the target signature of extracting, to improve the accuracy of utilizing Intensity correlation function decision node connectedness; (2), in the topological structure tentatively obtaining, how to get rid of " false connection "; (3) in actual use, how adaptive updates topological structure, to adapt to the variation of monitoring environment.
Summary of the invention
The object of the invention is to propose a kind of non-overlapping visual field multiple-camera monitor network topology adaptation learning method.The method has very strong robustness to indoor variation, complicated monitoring environment.
Technical scheme of the present invention is: by a weighted digraph G=<V for the topological structure of non-overlapping visual field multiple-camera monitor network, E, W> represents, and learns respectively the weight set W on the three elements in G: node set V, limit set E and limit.Enter, leave the position of this ken according to the target following result statistics target under haplopia territory, and set up disappearance under this ken-occur nodal analysis method by mixed Gauss model (GMM, Gaussian Mixture Model).Disappearance under all kens-occur nodal analysis method configuration node set V.Calculate across the ken node Intensity correlation function to (two nodes are respectively from the different cameras ken), thereby judge this connectedness to node.All connections across ken node to forming limit set E, the node under the identical ken is not to adding in limit set E.At computing node to the (node pair of mentioning hereinafter, unexplained reference all represent across ken node to) Intensity correlation function time, considered this to the disappearance measured value in node with there is measured value combine resemblance in appearance degree, utilize Intensity correlation function decision node to connective accuracy thereby improved.For the node pair being communicated with, by standardization, this distributes to the Intensity correlation function computing node of node right transfer time.Introduce the right interactive information of node and represent this transition probability to node, disconnected node is 0 to interactive information, and the right interactive information of node under the identical ken is also made as 0.The transition probability structure weight set W right according to all nodes.Construct connective matrix according to limit set E and weight set W, and infer the topological structure of monitor network by this matrix.Utilize " false connection " to get rid of strategy and remove " the false connection " that in the topological structure of inferring, may exist.Utilize topology adaptation update strategy to upgrade topological structure and parameter, proposed the quick response policy to adding/remove camera situation simultaneously, make the topological structure of learning can adapt to preferably monitor the change in bad border.
Specific implementation step of the present invention is followed successively by:
1) set up the node set that disappears-occur
2) gather on the limit of calculating topological structure
3) calculate and be communicated with right distribution transfer time of node
4) the weight set of calculating topological structure
5) construct connective matrix and infer topological structure
6) get rid of " false connection "
7) renewal of topological structure and parameter
Brief description of the drawings
Fig. 1 is the system flow chart that the present invention is based on the topology adaptation learning method of non-overlapping visual field multiple-camera monitor network
Fig. 2 is disappearance-the occur node schematic diagram of the monitoring scene that calculates of the present invention
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 removes " false connection " connective matrix and topological structure schematic diagram afterwards in the present invention
Fig. 6 be in the present invention transfer time distributed update schematic diagram
Fig. 7 is that transition probability of the present invention upgrades schematic diagram
Embodiment
Fig. 1 has provided the system flow chart of the topology adaptation learning method based on non-overlapping visual field multiple-camera monitor network: use weighted digraph model G=<V, E, W> represents the topological structure of non-overlapping visual field multiple-camera monitor network, and learns respectively the three elements in G: node set V, limit set E and weight set W.The present invention only considers the connectedness of the node (across ken node) in the different cameras ken, does not consider the connection situation of identical ken lower node, and the node being therefore communicated with under the identical ken is not to adding in the set of limit.The present invention enters, leaves position under the video camera ken node as topological structure using target, and uses mixed Gauss model to carry out modeling to the position that enters, leaves of target, and node set V is disappeared-occur.Right Intensity correlation function judges this connectedness to node to utilize node, by adding up the right connectedness structure limit set E of all nodes.For the node pair being communicated with, by standardization, this distributes to the Intensity correlation function computing node of node right transfer time.In the time of the right Intensity correlation function of computing node, considered this to disappearance measured value in node with there is measured value combine resemblance in appearance degree, comprise main color similarity, texture similarity and target conspicuousness similarity, utilize node right Intensity correlation function judges this connective accuracy to node thereby improved.For the node pair being communicated with, utilize this interactive information to node to represent the transition probability that node is right, the right transition probability of other all nodes is 0.The transition probability structure weight set W right according to all nodes.Construct connective matrix according to limit set E and weight set W, and infer the topological structure of monitor network by this matrix.For in topological structure, certain may be the situation of " false connection " to node, consider this to the interactive information of node and transfer time distributing to get rid of " false connection ".Utilize topology adaptation update strategy to upgrade topological structure and parameter, proposed the quick response policy to adding/remove camera situation simultaneously, make the topological structure of learning can adapt to preferably monitor the change in bad border.
Concrete operation step of the present invention
1) set up the node set that disappears-occur
Utilize ripe camshift tracking to obtain the movement locus of target under the different cameras ken, add up the position that target under each ken entered, left this ken, and utilize mixed Gauss model GMM to carry out modeling to these positions, obtain the disappearance of topological structure-occur node set V.GMM parameter is Z={z, λ
z, μ
z, σ
z 2.Wherein z represents the number of single Gaussian Profile in GMM, the number of the node that disappears-occur under the current video camera ken.λ
zrepresent each Gauss's weight, μ
z, σ
z 2for average and the variance of corresponding Gaussian Profile.The number z of the gauss component in GMM determines automatically by bayesian information criterion (Bayesian Information Criterion), other parameter { λ
z, μ
z, σ
z 2calculate by expectation maximization method (Expectation-Maximization method).The present invention tests the disappearance of monitor network used-occur node schematic diagram as shown in Figure 2.
2) gather on the limit of calculating topological structure
1.. the Intensity correlation function that computing node is right
The present invention adopts Intensity correlation function to judge adaptively whether certain is communicated with node.The target observation value sequence that is engraved in node i disappearance while supposing t is X
i(t) the target observation value sequence, occurring at node j in the same time is mutually Y
j(t), the Intensity correlation function of node i and node j can be defined as so
Wherein || X
i(t) || and || Y
j(t+T) || represent respectively X
iand Y (t)
j(t) number of measured value in, T is time of delay.Because Intensity correlation function has only been considered the disappearance measured value in two nodes and occurred measured value, do not consider disappearance measured value and occur that whether measured value belongs to same target, therefore comprises larger error.In order to improve the accuracy of utilizing the right Intensity correlation function decision node connectedness of node, the present invention has added associating resemblance in appearance degree in Intensity correlation function.Improved node i proposed by the invention and the Intensity correlation function of node j are defined as follows:
s.t.P
sim(O
a,i,O
b,j)＞δ
Wherein, O
a,iwith O
b,jrepresent respectively disappearance target observation value sequence X
iand occur target observation value sequence Y (t)
j(t+T) measured value in, P
sim(O
a,i, O
b,j) expression measured value O
a,iwith O
b,jassociating resemblance in appearance degree.Only has the O of working as
a,iwith O
b,jsimilarity while being greater than given threshold value δ, could be by measured value O
a,iwith O
b,jadd Intensity correlation function as associated measured value, ensured that the measured value of using in Intensity correlation function is to there being higher association.Be illustrated in figure 3 and use the connection node that calculates of method proposed by the invention to the Intensity correlation function curve right with not being communicated with node.In order to embody the superiority of the inventive method, provide the Intensity correlation function curve that does not add the Intensity correlation function curve of characteristic information and only add color similarity simultaneously and compared.Wherein (a), (b), (c) be for being communicated with the right Intensity correlation function curve of node, (d), (e), (f) be not for being communicated with the right Intensity correlation function curve of node.
2.. calculate associating resemblance in appearance degree
To utilize node right Intensity correlation function judges this accuracy to Connectivity in order to improve, the present invention is in the time of the right Intensity correlation function of computing node, considered this in node, disappear measured value with there is measured value combine resemblance in appearance degree, thereby ensured to only have in Intensity correlation function the measured value that resemblance in appearance degree is high just can be associated.The associating resemblance in appearance degree that the present invention proposes comprises main color similarity, texture similarity and target conspicuousness similarity.Suppose for any measured value O
a,iand O
b,j, their associating resemblance in appearance degree can be expressed as
P
sim(O
a,i,O
b,j)＝P(O
a,i,O
b,j|O
a＝O
b)
＝P(color
a,i,color
b,j|O
a＝O
b)
P(tex
a,i,tex
b,j|O
a＝O
b)P(dom
a,i,dom
b,j|O
a＝O
b)
＝P
colorP
texP
dom
Wherein, O
a=O
brepresent measured value O
a,iand O
b,jbelong to same target.Introduce respectively the computational methods of main color similarity, texture similarity and target conspicuousness similarity in associating resemblance in appearance degree below:
For measured value O
a,iand O
b,jmain color similarity P (color
a,i, color
b,j| O
a=O
b), the present invention adopts the method for divided group to calculate.Consider that the monitored object in this monitoring environment is people, rule of thumb measured value is divided into head, trunk, shank three parts by information.For any measured value, use respectively K means clustering method to extract primary color histogram in each part.Utilize the Bhattacharya of primary color histogram apart from the main color similarity that calculates two measured value corresponding parts.Hypothetic observation O
a,iand O
b,jhead, trunk, the main color similarity of shank be expressed as P
head, P
bodyand P
legs, measured value O so
a,iand O
b,jmain color similarity can be expressed as
P
Color＝αP
head+βP
body+γP
legs
Wherein, α, beta, gamma is respectively head, trunk, weight that shank is corresponding, must meet alpha+beta+γ=1.
For measured value O
a,iand O
b,jtexture similarity P (tex
a,i, tex
b,j| O
a=O
b), the present invention adopts gradient orientation histogram (HOG) feature of measured value to calculate.For measured value O
a,iwith O
b,j, extract respectively their HOG feature and be expressed as H
a,iand H
b,j.So, measured value O
a,iand O
b,jtexture similarity P
texjust can use H
a,iand H
b,jbhattacharya apart from expression, be defined as follows:
P
tex＝exp(-D
B/σ
t)
Wherein, σ
tpredefined bandwidth, D
brepresent H
a,iand H
b,jbhattacharya distance, k
hOGit is the dimension of HOG feature.
For measured value O
a,iand O
b,jconspicuousness similarity P (dom
a,i, dom
b,j| O
a=O
b), the present invention adopts the significant characteristics of measured value to calculate.In the time extracting the significant characteristics of measured value, first measured value is divided into fritter, use k arest neighbors method (KNN) to automatically identify in all fritters and the fritter of other fritter distinctiveness maximums, as the significant characteristics of this measured value.The conspicuousness similarity P of two measured values
domthe Euclidean distance of the significant characteristics by calculating these two measured values obtains.The significant characteristics that the present invention extracts has good robustness to attitude variation, the lighting change of environment etc. of target.
3.. calculate the limit set of topological structure
The Intensity correlation function that it is right that the present invention utilizes node judges this connectedness to node, thereby calculates the limit set E of topological structure.The present invention only considers the connectedness across ken node, does not consider the connection situation of identical ken lower node, and the node being therefore communicated with under the identical ken is not to adding in the set of limit.Node to the specific practice of connectedness judgement is: the peak value of Intensity correlation function curve right node and given threshold value thr are compared (whether the peak value that judges Intensity correlation function curve is obvious), if the peak value of Intensity correlation function curve is greater than thr, think that this is communicated with node.Otherwise, think that this is not communicated with node.Threshold value thr can calculate by the right Intensity correlation function of node, is defined as follows:
thr＝mean(R
ij(T))+ω·std(R
ij(T))
Wherein ω is User Defined parameter.
3) calculate and be communicated with right distribution transfer time of node
For the node pair being communicated with, need to calculate and distribute this transfer time to node, as the time-constrain of monitor network topological structure.Node distributes and can realize by the right Intensity correlation function of this node of standardization right transfer time.Suppose that the Intensity correlation function between node i and j is R
i,j(T), distribute the transfer time between these two nodes and can be defined as so
P
i,j(T)＝R
i,j(T)/||R
i,j(T)||
The present invention adopts Gauss model to represent to distribute transfer time, therefore distribution table transfer time can be shown to the form of a Gaussian Profile, as follows:
P
i,j(T)～N(μ,σ
2)
Wherein μ, σ
2represent respectively average and the variance of Gauss's model transfer time, can estimate to obtain by EM method.
4) the weight set of calculating topological structure
The interactive information that it is right that the present invention adopts node represents the transition probability that this node is right.Interactive information represents two degrees of dependence between sequence, and the sequence of observations of supposing node i is X
i(t), the sequence of observations of node j is Y
j(t+T), the interactive information I (X, Y) of these two sequence of observations is defined as follows:
Wherein, ρ
x,Y 2represent sequence X
iand Y (t)
j(t+T) the mutual correlation coefficient between, can pass through the Intensity correlation function R of these two sequences
i,j(T) calculate, be defined as follows:
Wherein T
peakrepresent that Intensity correlation function postpones the value of time T while getting obvious peak value, denominator is respectively X
i(t), Y
j(t+T) covariance of sequence.
For the node pair being communicated with, internodal transition probability equals their interactive information value.For disconnected node pair, internodal transition probability is 0.For the node under the identical ken, to (non-across ken node to), their transition probability is 0.Can construct the weight set W of topological structure according to the right transition probability of all nodes.
5) construct connective matrix and infer topological structure
Connective matrix can be constructed according to limit set E and weight set W, the topological structure of monitor network can be inferred according to this matrix.The value representation target of the capable n row of m of connective matrix moves to the transition probability of n node from m node, the right relevant position of disconnected node is 0.Under the identical video camera ken, all nodes are 0 to the value of the corresponding position at connective matrix.Obtain after connective matrix, for any row, the nonzero element of inquiring about in this row just can know the node that this row is corresponding can lead to which node (nonzero element is listed as corresponding node).The topological structure schematic diagram that is illustrated in figure 4 connective matrix and infers according to this connective matrix.
6) get rid of " false connection "
" false connect " refer to by connective matrix infer obtain certain to being communicated with node, be actually not directly communicated with, but pass through other node indirect communication.After certain node target disappears, often need to know this target occurs being positioned at which node next time, and these nodes are normally directly communicated with the node at target disappearance place.If there be more " false connection " to exist in the topological structure of inferring, can increase the complexity that target is identified again, reduce the accuracy rate of identification more simultaneously.Therefore, need to effectively remove " the false connection " in topological structure.The present invention considers the right interactive information of connection node and distributes transfer time, thereby judges whether this node is to being " false connection ".
Be communicated with the right interactive information size of node and can reflect internodal relevance power.Be communicated with the right interactive information of node larger, representing has very strong relevance between these two nodes, can think to be directly communicated with.Otherwise the right interactive information of node is less, a little less than representing that relevance between these two nodes is, can think that these two nodes are realized and being communicated with by other nodes, belong to " false connection ".
Be the node pair of " false connection " for using interactive information to judge whether, right distribution transfer time further judges can to utilize this node.For a pair of node i → j that may be " false connection ", the path of certain existence one non-" false connection " between them (k ..., m), make can be communicated with by this paths between i → j, i.e. i → k → ... → m → j.If there is no the path satisfying condition, node is to i so, and j must not be " false connection ".If there is the path that satisfies condition, can utilize distribute transfer time of i → j with i → k → ... whether the decision node that distributes the transfer time of → m → j is " false connection " to i → j.Known and Gaussian distributed because arbitrary node in topological structure distributed to satisfied time, so according to Gauss's Adding law can obtain i → k → ... distribute the transfer time of → m → j and obey N (μ
ik+ ... + μ
mj, σ
ik 2+ ... + σ
mj 2).Meanwhile, node is obeyed N (μ to distributing the transfer time of i → j
ij, σ
ij 2).Calculate N (μ
ik+ ... + μ
mj, σ
ik 2+ σ
mj 2) and N (μ
ij, σ
ij 2) Kullback-Leibler divergence, and result and predetermined threshold value th are compared.If divergence is less than th, so two kinds be communicated with situations transfer time distribute similar, think i → j be actually by i → k → ... → m → j realizes, and therefore node is " false connection " to i → j.Otherwise, think that node is not " false connection " to i → j.Be illustrated in figure 5 the connective matrix of Fig. 4 is carried out to new connective matrix and the topological structure that after " false connection " eliminating, obtain with topological structure, the topological structure shown in Fig. 5 has truly reflected the connection situation of Fig. 2 Scene.
7) topology adaptation upgrades
1.. transfer time distributed update
Utilize target internodally to upgrade and distribute transfer time to node this by the time being communicated with at certain.Suppose t
kbe that k target is internodal by the time to being communicated with at certain, this is to current distribution transfer time of node obedience N (μ, σ
2), ρ is turnover rate, the parameter renewal process of time distributed model is as given a definition so
μ
*＝(1-ρ)μ+ρt
k
σ
*2＝(1-ρ)σ
2+ρ(μ
*-t
k)
2
Be illustrated in figure 6 the schematic diagram of distributed update transfer time, transfer time, distribution became N (17.8,3.85) by the N (13.2,3.2) of previous moment.The variance difference of these two Gaussian Profile is little, and main difference is embodied in average, and the former mean transit time 13.2s is shorter than the latter's 17.8s, and before showing, the movement velocity of target is very fast.
2.. transition probability upgrades
Utilize target to upgrade the transition probability of these two nodes in two internodal transfer case, suppose that, for two node i and j, turnover rate is κ, the renewal of transition probability can be expressed as so
w
ij *＝(1-κ)w
ij+κP
ij(T
w)
Wherein P
ij(T
w) be illustrated in T
wthe ratio of the target numbers from node i to node j in the time and the target numbers of leaving from node i, T
wgenerally be taken as one hour, within each one hour, upgrade a transition probability.Be illustrated in figure 7 the schematic diagram that transition probability upgrades, in figure, the transition probability of three hours slightly increases above, and transition probability diminishes gradually afterwards.
3.. add/remove camera quick response
When the camera arrangement in monitoring environment changes, need to readjust topology network architecture.For the situation that removes video camera, only need to disconnect node under original other video camera kens and be removed the associated of node under the video camera ken.For the situation that increases video camera, by learning disappearance under the video camera ken that identical method relearns increase-occur node with initial topology, and these nodes and existed connected relation between node, transfer time to distribute and transition probability, and topological structure is upgraded accordingly.
Taking above-mentioned foundation desirable embodiment of the present invention as enlightenment, by above-mentioned description, relevant staff can, not departing from the scope of this invention technological thought, carry out various change and amendment completely.The technical scope of this invention is not limited to the content on specification, must determine its technical scope according to claim scope.
Claims (7)
1. a non-overlapping visual field multiple-camera monitor network topology adaptation learning method, it is characterized in that, represent the topological structure of non-overlapping visual field multiple-camera monitor network with weighted digraph, weighted digraph is made up of node set, limit set and weight set;
The position that target under each video camera ken was entered, left this ken, as topological node, is defined as the node that disappears-occur;
Use haplopia territory method for tracking target to add up target under each ken and enter, leave the position of this ken, and use the disappearance of mixed Gauss model structure topological structure-occur node set;
Utilize and judge this connectedness to node across the right Intensity correlation function of ken node, described across ken node to referring to that two nodes are respectively from the different video camera kens, all connections across ken node to forming limit set;
In the time calculating Intensity correlation function, consider node centering disappearance measured value with there is measured value combine resemblance in appearance degree, comprise main color similarity, texture similarity and target conspicuousness similarity, for be communicated with across ken node pair, calculate and distribute its transfer time by this Intensity correlation function to node of standardization;
Introduce the right interactive information of node and represent this transition probability to node, thus the set of structure weight;
Construct connective matrix according to the limit set of trying to achieve with weight set, and infer the topological structure of monitor network by this matrix;
For in topological structure, certain may be the situation of " false connection " to node, consider this to the interactive information of node and transfer time distributing to get rid of " false connection ";
Utilize topology adaptation update strategy to upgrade topological structure and parameter;
Carry out quick response policy to adding/remove camera situation simultaneously.
2. non-overlapping visual field multiple-camera topology adaptation learning method according to claim 1, is characterized in that, uses the node set of mixed Gauss model structure topological structure, and key step is as follows:
A1) target enters, leaves determining of video camera ken position
Utilize method for tracking target under haplopia territory to obtain the movement locus of target, obtain after the movement locus of certain section of interior all targets of time under certain video camera ken, what just can count target under this ken enters, leaves position, and the method for tracking target under described haplopia territory includes but not limited to: camshift tracking, kalman tracking, TLD tracking;
A2) the definite of node disappears-occurs
The position that enters, leaves the video camera ken using target, as topological node, defines the node that disappears-occur:
For certain video camera ken, utilize method for tracking target under haplopia territory to obtain behind position that target entered, left this ken, adopt mixed Gauss model GMM (Gaussian Mixture Model) to carry out modeling to these positions, construct the node that disappears-occur, wherein, the parameter of GMM is Z={z, λ
z, μ
z, σ
z 2, wherein z represents the number of single Gaussian Profile in GMM, the number of the node that disappears-occur under the current video camera ken, λ
zrepresent each Gauss's weight, μ
z, σ
z 2for average and the variance of corresponding Gaussian Profile, the number z of the gauss component in GMM determines automatically by bayesian information criterion (Bayesian Information Criterion), other parameter { λ
z, μ
z, σ
z 2calculate by expectation maximization method (EM).
3. non-overlapping visual field multiple-camera topology adaptation learning method according to claim 2, it is characterized in that, described step a1) in adopt camshift tracking to carry out the target following under haplopia territory, and determine the movement locus of target according to the state of each frame of target.
4. non-overlapping visual field multiple-camera topology adaptation learning method according to claim 1, is characterized in that, uses the limit set of the Intensity correlation function structure topological structure right across ken node, and key step is as follows:
B1) calculate across the right Intensity correlation function of ken node
Adopt the topological structure without supervision formula learning method study monitor network, right Intensity correlation function judges whether this is communicated with node to utilize node; The target observation sequence that is engraved in node i disappearance while supposing t is X
i(t) the target observation sequence, occurring at node j in the same time is mutually Y
j(t), the Intensity correlation function of node i and j can be defined as so
Wherein || X
i(t) || and || Y
j(t+T) || represent respectively X
iand Y (t)
j(t) measured value number, T is time of delay; In order to improve Intensity correlation function decision node to connective accuracy, in Intensity correlation function, add associating resemblance in appearance degree, be defined as follows:
s.t.P
sim(O
a,i,O
b,j)＞δ
Wherein, O
a,iand O
b,jrepresent respectively disappearance target sequence X
iand occur target sequence Y (t)
j(t+T) measured value, P
sim(O
a,i, O
b,j) represent the associating resemblance in appearance degree of two measured values, only have in the time that similarity is greater than given threshold value δ and could be added Intensity correlation function, having ensured to be used in Intensity correlation function associated measured value has higher probability to belong to identical target;
B2) calculating of associating resemblance in appearance degree
Associating resemblance in appearance degree comprises main color similarity, texture similarity and target conspicuousness similarity, and the multiple Feature Combination that is easy to extract is used;
Suppose to have two measured value O
a,iand O
b,j, their associating resemblance in appearance degree can be expressed as
P
sim(O
a,i,O
b,j)＝P(O
a,i,O
b,j|O
a＝O
b)
＝P(color
a,i,color
b,j|O
a＝O
b)·
P(tex
a,i,tex
b,j|O
a＝O
b)·P(dom
a,i,dom
b,j|O
a＝O
b)
＝P
color·P
tex·P
dom
Wherein O
a=O
brepresent that two measured values belong to same target;
For main color similarity P
coloradopt the method for divided group to calculate, consider that the monitored object in this monitoring environment is people, rule of thumb measured value is divided into head, trunk, shank three parts by information, for any measured value, use respectively K means clustering method to extract primary color histogram in each part, the main color similarity of two measured value parts can represent by the Bhattacharya distance of the primary color histogram of corresponding part, hypothetic observation O
a,iand O
b,jhead, trunk, the main color similarity of shank be expressed as P
head, P
bodyand P
legs, measured value O so
a,iand O
b,jmain color similarity can be expressed as
P
Color＝αP
head+βP
body+γP
legs
Wherein, α, beta, gamma is respectively head, trunk, weight that shank is corresponding, must meet alpha+beta+γ=1;
For texture similarity P
tex, the present invention adopts the gradient orientation histogram HOG feature of measured value to calculate, for measured value O
a,iwith O
b,j, extract respectively their HOG feature, be expressed as H
a,iand H
b,j, so, measured value O
a,iand O
b,jtexture similarity P
texjust can use H
a,iand H
b,jbhattacharya apart from expression, be defined as follows:
P
tex＝exp(-D
B/σ
t)
Wherein, σ
tpredefined bandwidth, D
brepresent HOG descriptor H
a,iand H
b,jbhattacharya distance, k
hOGit is the dimension of HOG descriptor;
For the conspicuousness similarity P of target
domadopt the significant characteristics of measured value to calculate, in the time extracting the significant characteristics of measured value, first measured value is divided into fritter, use k arest neighbors method KNN to automatically identify in all fritters and the fritter of other fritter distinctiveness maximums, as the significant characteristics of this measured value, the conspicuousness similarity P of two measured values
domobtain by the Euclidean distance of calculating their significant characteristics;
B3) across ken node, connectedness is judged
Utilize node to judge this connectedness to node to the peak value of Intensity correlation function curve, specifically comprise the following steps: the peak value by node to Intensity correlation function curve and threshold value thr compare, be whether decision node is obvious to the peak value of Intensity correlation function curve, if the peak value of Intensity correlation function curve is greater than threshold value thr, think that this is communicated with node, described threshold value thr can calculate by the right Intensity correlation function of node, is defined as follows:
thr＝mean(R
ij(T))+ω·std(R
ij(T))
Wherein ω is User Defined parameter.
B4) right distribution transfer time is calculated to be communicated with node
For the node pair being communicated with, distribute the transfer time that need to calculate it, as the time-constrain of monitor network topological structure, calculated and distributed this transfer time to node by the right Intensity correlation function of standardization node; Suppose that the Intensity correlation function between node i and j is R
i,j(T), distribute the transfer time between these two nodes and can be defined as so
P
i,j(T)＝R
i,j(T)/||R
i,j(T)||
The present invention utilizes Gauss model to represent that node distributes right transfer time, as follows:
P
i,j(T)～N(μ,σ
2)
Wherein μ, σ
2represent respectively average and the variance of Gauss model, can estimate to obtain by EM method.
5. non-overlapping visual field multiple-camera topology adaptation learning method according to claim 1, is characterized in that, uses the weight set of interactive information structure topological structure, comprises the following steps:
Adopt the right interactive information of node to represent the transition probability that node is right, the transition probability structure weight set right according to node; Interactive information represents two degrees of dependence between sequence, and the sequence of observations of supposing node i is X
i(t), the sequence of observations of node j is Y
j(t+T), the interactive information I (X, Y) of these two sequence of observations is defined as follows:
Wherein, ρ
x,Y 2represent X between two sequences
iand Y (t)
j(t+T) mutual correlation coefficient, can pass through the Intensity correlation function R of these two sequences
i,j(T) calculate, be defined as follows:
Wherein T
peakrepresent that Intensity correlation function postpones the value of time T while getting obvious peak value, denominator is respectively X
i(t), Y
j(t+T) covariance of sequence, for the node pair being communicated with, internodal transition probability equals their interactive information value, for disconnected node pair, internodal transition probability is 0, for all nodes pair under the identical ken, non-across ken node pair, think that their transition probability is 0.
6. non-overlapping visual field multiple-camera topology adaptation learning method according to claim 1, is characterized in that, uses connective matrix to infer the topological structure of monitor network, comprises the steps:
C1) determining of connective matrix
Utilize limit set to construct connective matrix with weight set, and infer the topological structure of monitor network by this matrix; The value representation target of the capable n row of m of connective matrix moves to the transition probability of n node from m node, for the node pair that can not be communicated with, the node number right according to node sets to 0 in the relevant position of connective matrix, under the identical video camera ken, all nodes are 0 to the value of the corresponding position at connective matrix, obtain after connective matrix, have a corresponding row for arbitrary node, just can obtain by the nonzero element of finding this row other nodes that are communicated with this node;
C2) " false connection " eliminating strategy
" false connect " refer to infer by connective matrix certain to being communicated with node, be actually not directly communicated with, but pass through other node indirect communication; When target disappears at certain node, need to know which node is this target may occur at next time, and these nodes should directly be communicated with the node that target disappears, consider and be communicated with the right interactive information of node and distribution transfer time, thereby judge whether this node is to being " false connection ";
Be communicated with the right interactive information of node larger, representing has very strong relevance between these two nodes, can think to be directly communicated with; Otherwise the right interactive information of node is less, a little less than representing that relevance between these two nodes is, can think that these two nodes are realized and being communicated with by other nodes, belong to " false connection ";
It is the node pair of " false connection " for using interactive information to judge whether, right distribution transfer time further judges can to utilize this node, for a pair of node i → j that may be " false connection ", certain path (k that has non-" false connection " between them, m), make can be communicated with by this paths between i → j, i.e. i → k → ... → m → j.If there is no the path satisfying condition, node is to i so, j must not be " false connect ", if there is the path satisfying condition, can utilize distribute transfer time of i → j with i → k → ... whether the decision node that distributes the transfer time of → m → j is " false connection " to i → j; According to Gauss's Adding law can obtain i → k → ... distribute the transfer time of → m → j and obey N (μ
ik+ ... + μ
mj, σ
ik 2+ ... + σ
mj 2); Meanwhile, node is obeyed N (μ to distributing the transfer time of i → j
ij, σ
ij 2); Calculate N (μ
ik+ ... + μ
mj, σ
ik 2+ σ
mj 2) and N (μ
ij, σ
ij 2) Kullback-Leibler divergence, and result and predetermined threshold value th are compared, if divergence is less than th, distribute similar two kinds of transfer times that are communicated with situation so, think i → j be actually by i → k → ... → m → j realizes, therefore node is " false connection " to i → j, otherwise, think that node is not " false connection " to i → j.
7. non-overlapping visual field multiple-camera topology adaptation learning method according to claim 1, is characterized in that, uses topology adaptation update strategy to upgrade monitor network topological structure, comprises the steps:
D1) transfer time distributed update strategy
Utilize target internodally to upgrade and distribute transfer time to node this by the time being communicated with at certain; Suppose t
kbe that k target is internodal by the time to being communicated with at certain, this is to current distribution transfer time of node obedience N (μ, σ
2), ρ is turnover rate, the parameter renewal process of time distributed model is as given a definition so:
μ
*＝(1-ρ)μ+ρt
k
σ
*2＝(1-ρ)σ
2+ρ(μ
*-t
k)
2
D2) transition probability update strategy
Utilize target to upgrade the transition probability of these two nodes in two internodal transfer case, suppose that, for two node i and j, turnover rate is κ, the renewal of transition probability can be expressed as so
w
ij *＝(1-κ)w
ij+κP
ij(T
w)
Wherein P
ij(T
w) be illustrated in T
wthe ratio of the target numbers from node i to node j in the time and the target numbers of leaving from node i, T
wgenerally be taken as one hour, within each one hour, upgrade a transition probability;
D3) add/remove camera
When the camera arrangement in monitoring environment changes, need to readjust topology network architecture, for the situation that removes video camera, only need to disconnect node under original other video camera kens and be removed the associated of node under the video camera ken, for the situation that increases video camera, by learning disappearance under the video camera ken that identical method relearns increase-occur node with initial topology, and these nodes and existed connected relation between node, transfer time to distribute and transition probability, and topological structure is upgraded 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 true CN104010168A (en) | 2014-08-27 |
| CN104010168B 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) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104599290A (en) * | 2015-01-19 | 2015-05-06 | 苏州经贸职业技术学院 | Video sensing node-oriented target detection method |
| CN105744223A (en) * | 2016-02-04 | 2016-07-06 | 北京旷视科技有限公司 | Video data processing method and apparatus |
| CN106846378A (en) * | 2017-01-23 | 2017-06-13 | 中山大学 | Across video camera object matching and tracking that a kind of combination topology of spacetime is estimated |
| 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 |
| CN107689054A (en) * | 2017-08-24 | 2018-02-13 | 北京航空航天大学 | A kind of polyphaser topology connected graph is established and across camera subject tracking |
| CN107705327A (en) * | 2017-08-29 | 2018-02-16 | 电子科技大学 | The candidate target extracting method of polyphaser network space-time model |
| CN108614889A (en) * | 2018-05-04 | 2018-10-02 | 济南大学 | Mobile object Continuous k-nearest Neighbor based on mixed Gauss model and system |
| CN110251076A (en) * | 2019-06-21 | 2019-09-20 | 安徽大学 | Merge conspicuousness detection method and device of the visual attention based on contrast |
| CN113313188A (en) * | 2021-06-10 | 2021-08-27 | 四川大学 | Cross-modal fusion target tracking method |
| CN114859300A (en) * | 2022-07-07 | 2022-08-05 | 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 (7)
| 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 |
| US20070296815A1 (en) * | 2004-11-12 | 2007-12-27 | Saab Ab | Image-Based Movement Tracking |
| 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 |
-
2014
- 2014-06-13 CN CN201410266226.6A patent/CN104010168B/en not_active Expired - Fee Related
Patent Citations (7)
| 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 |
| US20070296815A1 (en) * | 2004-11-12 | 2007-12-27 | Saab Ab | Image-Based Movement Tracking |
| 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 |
Cited By (17)
| 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 |
| CN104599290A (en) * | 2015-01-19 | 2015-05-06 | 苏州经贸职业技术学院 | Video sensing node-oriented target detection method |
| CN105744223A (en) * | 2016-02-04 | 2016-07-06 | 北京旷视科技有限公司 | Video data processing method and apparatus |
| CN106846378A (en) * | 2017-01-23 | 2017-06-13 | 中山大学 | Across video camera object matching and tracking that a kind of combination topology of spacetime is estimated |
| 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 |
| CN107689054A (en) * | 2017-08-24 | 2018-02-13 | 北京航空航天大学 | A kind of polyphaser topology connected graph is established and across camera subject tracking |
| 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 |
| CN107705327A (en) * | 2017-08-29 | 2018-02-16 | 电子科技大学 | The candidate target extracting method of polyphaser network space-time model |
| CN108614889A (en) * | 2018-05-04 | 2018-10-02 | 济南大学 | Mobile object Continuous k-nearest Neighbor based on mixed Gauss model and system |
| 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 |
| CN110251076A (en) * | 2019-06-21 | 2019-09-20 | 安徽大学 | Merge conspicuousness detection method and device of the visual attention based on contrast |
| 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 |
| CN113313188A (en) * | 2021-06-10 | 2021-08-27 | 四川大学 | Cross-modal fusion target tracking method |
| CN113313188B (en) * | 2021-06-10 | 2022-04-12 | 四川大学 | Cross-modal fusion target tracking method |
| CN114859300A (en) * | 2022-07-07 | 2022-08-05 | 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104010168B (en) | 2017-06-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104010168A (en) | Non-overlapping vision field multi-camera monitoring network topology self-adaptation learning method | |
| CN102436662B (en) | Human body target tracking method in nonoverlapping vision field multi-camera network | |
| CN104504400B (en) | A kind of driver's anomaly detection method based on online behavior modeling | |
| Elmas et al. | A data fusion framework with novel hybrid algorithm for multi-agent Decision Support System for Forest Fire | |
| US11757723B2 (en) | Distributed secure state reconstruction method based on double-layer dynamic switching observer | |
| CN103390278A (en) | Detecting system for video aberrant behavior | |
| CN104717468B (en) | Cluster scene intelligent monitoring method and system based on the classification of cluster track | |
| CN112788699B (en) | Method and system for determining network topology of self-organizing network | |
| CN103971384B (en) | Node cooperation target tracking method of wireless video sensor | |
| CN110414400A (en) | A kind of construction site safety cap wearing automatic testing method and system | |
| CN109389044A (en) | More scene crowd density estimation methods based on convolutional network and multi-task learning | |
| WO2018227491A1 (en) | Method and device for association of fuzzy data of multiple targets in video | |
| CN104065932B (en) | A kind of non-overlapping visual field target matching method based on amendment weighting bigraph (bipartite graph) | |
| Shakeri et al. | Real-time bird detection based on background subtraction | |
| Fox et al. | Hippocampus as unitary coherent particle filter | |
| CN114265913A (en) | Space-time prediction algorithm based on federal learning on industrial Internet of things edge equipment | |
| CN104091332A (en) | Method for optimizing multilayer image segmentation of multiclass color texture images based on variation model | |
| CN106991400A (en) | A kind of fire hazard smoke detecting method and device | |
| Chau | Rainfall-runoff correlation with particle swarm optimization algorithm | |
| Falcon et al. | Behavioral learning of vessel types with fuzzy-rough decision trees | |
| CN109309586A (en) | A kind of food processing tele-control system intrusion detection method | |
| Li et al. | Abnormal data detection in sensor networks based on DNN algorithm and cluster analysis | |
| Chu et al. | Self-organized and scalable camera networks for systematic human tracking across nonoverlapping cameras | |
| CN116563953B (en) | Bottom-up weak supervision time sequence action detection method, system, equipment and medium | |
| Naït-Abdesselam et al. | Data Quality Improvements for Internet of Things Using Artificial Neural Networks |
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 |