CN102497495B - Target association method for multi-camera monitoring system - Google Patents

Abstract

The invention discloses a target association method for a multi-camera monitoring system. The method mainly comprises the following steps that: a minimum cost flow module based on increment realizes real-time tracking of a target object in a multi-camera covering domain; when a target appears in a camera, a target detection module of a computer connected with the camera detects the target, extracts the characteristics of the target, and finally uploads the characteristics to a server; a target tracking module on the computer continually tracks the target inside a single camera scene, extractsthe characteristics of the target at the same moment once again at certain time intervals, and uploads the characteristics to the server; and the server only calculates whether target characteristicstransmitted by the target detection module include characteristics which can be associated with the target characteristics and are transmitted by the target tracking module and the target tracking module, and returns a result to the computer. The target association method can be widely applied to an intelligent monitoring system in indoor and outdoor scenes, and has a wide market prospect and high application value.

Description

A kind of target association method for the multiple-camera supervisory control system

Technical field

The present invention relates to method such as moving object detection, target following in the intelligent video monitoring system, mainly be applicable to the multiple-camera supervisory control system, belong to the technical field of video monitoring, be specifically related to a kind of target association method for the multiple-camera supervisory control system.

Background technology

Along with the acceleration of global urban process, a large amount of video cameras has been installed in the public place.Though exist very big correlation between video camera, related application also only rests on the basis of single video camera related algorithm.At present, work in coordination with the technology of monitoring between multiple-camera and still be in the starting stage.The multiple-camera tracking mainly comprises at present:

1, based on the multiple-camera target tracking algorism of three-dimensional information

If Camera calibration information and three-dimensional environment coordinate information are known, just can be mapped to these information unifications under the same coordinate system by certain mapping function, at last just can obtain corresponding relation between the correct multiple-camera by comparatively simple one dimension parameter.But this method needs equipment that higher precision is arranged.

2, the multiple-camera target tracking algorism of based target model

This basic idea is that hypothesis assert that certain feature of people is not easy to change along with the change in time and space (as people's gait etc.), these features are set up unified model, when target detection, directly detect this feature of target, with the model comparison, draw the identity of target.But the method process complexity is difficult to realize real-time.

3, the multiple-camera target tracking algorism that merges based on feature

The basic thought of this method is some simple feature of selecting target in the initial period for use, as color, profile, positional information etc., utilizes methods such as statistics or probability to draw last corresponding relation then.

Summary of the invention

The technical problem to be solved in the present invention: a kind of target association method for the multiple-camera supervisory control system is provided, and this method adopts the least cost flow model based on increment, makes the raising degree of system effectiveness fairly obvious.

The technical scheme that the present invention solves the problems of the technologies described above: a kind of target association method for the multiple-camera supervisory control system comprises the steps:

Step (1) will be applied to the association between the target in the multiple-camera supervisory control system based on the least cost flow model of increment;

Step (2) is by repeatedly training, and the Adjustment System parameter improves the accuracy of this target association method;

Used similarity measurement when step (3) target association calculates is taken all factors into consideration feature that target is detected in the target detection stage and the feature several times in time and the target tracking stage afterwards.

Wherein, described least cost flow model based on increment when time target association calculates, does not directly rebulid whole least cost flow model, but takes full advantage of a preceding target association result calculated, reduces amount of calculation, improves running efficiency of system.

Wherein, described by repeatedly training, utilize repeatedly to train to obtain the reasonable threshold interval of a plurality of systems, these interval logarithmic axis are covered, and number axis is divided into a plurality of intervals, the intermediate value in the interval that coverage is the highest is as the threshold value of system.

Wherein, used similarity measurement when described target association calculates, the utility function in target association calculates in conjunction with the module of target detection in the supervisory control system and target tracking module, extracts feature simultaneously in two modules, improve the related accuracy of calculating.

Wherein, utility function during described target association calculates, target signature similarity are calculated and are got the feature that first object appearing is partly sampled in target detection stage and target tracking stage on the feature in target detection stage and time order of back object appearing on the time order and carry out similarity measurement.

Principle of the present invention is:

The target association method based on the least cost flow model of increment that adopts in this method is new object observing to occur, when server end will carry out target association calculating, directly the remaining network of the least cost flow network that calculates at last once target association adds two nodes and a spot of limit, and carrying out iteration then, to look for weights be positive ring.After finding positive ring, directly revise remaining network.When do not exist just encircle after, finish based on the part of increment.Constantly look for least cost road augmentation afterwards, after certain augmentation, the threshold value that the added value of expense is obtained by the training stage less than system.The expense of least cost flow network top, the characteristics determined of being extracted by the target of target detection and target tracking module mark.

The present invention's advantage compared with prior art is:

In multiple-camera relay tracking research field, the thought of Combinatorial Optimization is used for association between a plurality of targets, it is the method for relatively using always, wherein the least cost flow model is a kind of simple, intuitive and comparatively ripe a kind of target association method, but each efficient of directly using the least cost flow model to carry out target association calculating is too low.And the minimum cost flow simulated target correlating method that is based on increment that uses in this patent has taken full advantage of last target association result calculated, thereby makes computational efficiency be greatly improved.

Description of drawings

Fig. 1 is system software module figure;

Fig. 2 is the system hardware Organization Chart;

Fig. 3 is the client-side program flow chart;

Fig. 4 is based on the target association method flow diagram of increment;

Fig. 5 is effective threshold interval schematic diagram.

Embodiment

The present invention is described in detail below in conjunction with drawings and the specific embodiments.

As shown in Figure 1, the native system software module comprises: target association module and data transmission module are arranged on the server end, image capture module, module of target detection, target tracking module and data transmission module are arranged on the client.

As shown in Figure 2, the native system hardware device comprises: a station server, multiple cameras and the computer that links to each other with every video camera.

In system, work as a new target appears in the video camera at every turn, the module of target detection of the computer that links to each other with this video camera will detect this target and it is extracted feature, the rectangular area mark at detected target place is come out, color histogram is obtained in this rectangular area, and feature uploads onto the server the most at last.Because the rectangular area of each observation size is also inequality, thus need be to color histogram normalization, to the data of each dimension of feature, divided by data on all dimensions and, its meaning accounts for the proportion of all data for the data on this one dimension.To in i video camera, detected a target O of target detection stage _{I, a}Color histogram be designated as f _{I, a, 0}, the time is designated as t _{I, a}And the g of color histogram feature dimension is designated as f _{I, a, 0, g}

Then, target tracking module can carry out continuing in the single camera to follow the tracks of to the target that is detected, in tracing process, sampled in the rectangular area at this target place at regular intervals, and obtain color histogram after the normalization of this new rectangular area again.To in i video camera, detected a target O of target detection stage _{I, a}The color histogram that obtains in the h time sampling of target tracking stage afterwards is designated as f _{I, a, h}(h＞0), and this color histogram feature g dimension be designated as f _{I, a, h, g}Like this, this object observing just has the feature f of many group color histograms in this video camera _{I, a, 0}, f _{I, a, 1}, f _{I, a, 2}..., f _{I, a, h}

The feature that server only transmits module of target detection, by the target association algorithm based on the least cost flow model of increment in the target association module, whether have comprise the feature that target detection and target tracking module transmit of with it association, and the result is returned to computer if calculating this target signature.As a new target O _{J, b}Appear in another video camera, ask target O this moment _{J, b}With target O _{I, a}Association

Effectiveness

Be that this association is twice continuous possibility that occurs of the same target in the real world, adopt current object observing O _{J, b}Color histogram feature f _{J, b, 0}With object observing O _{I, a}The feature f of q+1 color histogram _{I, a, 0}, f _{I, a, 1}, f _{I, a, 2}..., f _{I, a, q}Ask Euclidean distance respectively, and the Euclidean distance that these are tried to achieve is got minimum value as object observing O _{J, b}With object observing O _{I, a}Similarity distance between the color histogram feature, that is:

$\mathrm{dis}\tan \mathrm{ce}(f_{i,a},f_{j,b})=\underset{h=0}{\overset{q}{\min }}\sqrt{\underset{g=1}{\overset{r}{\mathrm{\Σ}}}{(f_{j,b,0,g}-f_{i,a,h,g})}^2}$

In system, between per two video cameras that directly can reach a parameter is set and is called the mean transferred time, namely under the photographed scene of a video camera without the scene of the 3rd video camera and directly arrive the expectation of the photographed scene required time of another video camera.In utility function, the mean transferred time is only done the accessibility judgement,, if the mean transferred time is more than 10 times of two time differences between the observation, thinks that then this association is infeasible that is.

The feature of the final color histogram of in system, tieing up for q+1 r, the utility function of target association algorithm adopts:

ξ wherein _{I, j}=1 represents between two video cameras and directly can reach ξ _{I, j}=0 represents between two video cameras and directly can not reach τ _{I, j}It is two mean transferred times between the video camera that directly can reach.

If do not satisfy condition t _{J, b}＞t _{I, a}, and ξ _{I, j}=1, and

Be in the formula " otherwise " situation; This represents this association can not need not consider this association for the continuous appearance of the same target in the real world in system, so will

Value is-∞.

The software of native system is divided into server end and client.Server software operates on the server, on the computer that client software operates in video camera links to each other.Client-side program flow chart such as Fig. 3, the client-side program flow process is as follows:

Step 1 is extracted view data, and this step is finished by image capture module.

Step 2, target detection utilizes the aims of systems detection module to carrying out target detection in certain zone, and detected fresh target is sent clarification of objective to server end.

Step 3, target following utilizes the aims of systems module that detected target in the step 2 is carried out lasting tracking in the multiple-camera scope, and after at set intervals, sends the current feature of tracking target to server end.

Server is embodied as, when receiving clarification of objective that a computer is sent, judge at first whether this feature is by the detected fresh target of module of target detection, if fresh target, just as Fig. 4, use the minimum cost flow simulated target correlating method based on increment to carry out a target association calculating.

The target association method modeling pattern of least cost flow model is: increase by two node s at the fee flows network, t is as source point and meeting point.To each target O _{I, a}, increase by two nodes at the fee flows network

Add the limit Expense is 0, and flow restriction is 1.To each effectiveness

Be not-association of ∞

Add the limit

Expense is

Flow restriction is 1.This flow network is constantly asked the shortest augmenting path, after certain augmentation, flow is that incidence number is increased to m+1 from m, expense is relevant effectiveness and is increased to w (m+1) from w (m), and effective increment function w (m+1)-w (m) is less than threshold value, at this moment, flow m before the augmentation is the optimal relevance number, be real related number in the real world, the stream before the augmentation, the institute that flows through tangible as

The limit, represent target O _{I, a}With O _{J, b}Related.

Minimum cost flow simulated target correlating method flow process based on increment is as follows:

Step 1, the remaining network that calculates at last once target association increases node and the limit of current newly-increased target correspondence.

Step 2 is positive ring if there is not cost metric in the residual network, forwards step 4 to, otherwise changes the step to rapid 3.

Step 3 is positive ring augmentation along the cost metric that finds, and changes step 2.

Step 4: continue to look for the shortest augmenting path in current network, if can not find the shortest augmenting path or effective increment less than threshold value, then finish.Otherwise forward step 4 to.

At this moment, as the method for non-increment, the flow m before the last augmentation is the optimal relevance number, the stream before the augmentation, the institute that flows through tangible as The limit, represent target O _{I, a}With O _{J, b}Related.

The reasonable threshold value of system utility increment is chosen the employing training method, and flow process is as follows: at first each group sample is carried out the calculating of effective increment function w (m) in the training, count M according to the true optimal relevance of sample, can obtain two value w (M) and w (M+1).W (M+1)＜w (M)≤1 wherein must be arranged.As shown in Figure 5, so just obtained one to the effective threshold interval of this group sample (w (M+1), w (M)).

This interval meaning is: the threshold value t of effective increment function gets this interval any one interior number, can both make this group sample obtain a correct incidence number after the algorithm operation.

Each group sample is all calculated its effective threshold interval (w _i(M+1), w _i(M)), and then try to achieve an interval (p, q) make satisfied

Interval quantity maximum.Then train the threshold value t that obtains to get

The meaning of doing so directly perceived is that the threshold value of the effective increment function selected can make algorithm when sample set moves, and the samples of many as far as possible groups are arranged, and what its optimal relevance number obtained is correct value.

The flow process in the interval of concrete judgement degree of covering maximum is as follows:

Step 1 is marked at the coordinate that occurs in all effective threshold intervals on the reference axis X-axis, thereby X-axis is divided into a plurality of intervals, and with these intervals counter of each auto correlation all, initial value all is made as 0.

Step 2, for each effective threshold interval, the Counter Value that the interval that the X-axis that it is covered splits is associated with all adds 1.

Step 3 is found out the interval of counter values maximum, calculate this interval left and right sides end points and half threshold value as system.

The part that the present invention does not elaborate belongs to techniques well known.

Claims (2)

1. a target association method that is used for the multiple-camera supervisory control system is characterized in that: comprise the steps:

Step (1) will be applied to the association between the target in the multiple-camera supervisory control system based on the least cost flow model of increment;

Step (2) is by repeatedly training, and the Adjustment System parameter improves the accuracy of this target association method;

Used similarity measurement when step (3) target association calculates is taken all factors into consideration feature that target is detected in the target detection stage and the feature several times in time and the target tracking stage afterwards; Described least cost flow model based on increment when time target association calculates, does not directly rebulid whole least cost flow model, but takes full advantage of a preceding target association result calculated, reduces amount of calculation, improves running efficiency of system; Minimum cost flow simulated target correlating method flow process based on increment is as follows:

Step 1, the remaining network that calculates at last once target association increases node and the limit of current newly-increased target correspondence;

Step 2 is positive ring if there is not cost metric in the residual network, forwards step 4 to, otherwise forwards step 3 to;

Step 3 is positive ring augmentation along the cost metric that finds, and forwards step 2 to;

Step 4: continue to look for the shortest augmenting path in current network, if can not find the shortest augmenting path or effective increment less than threshold value, then finish, otherwise forward step 4 to.

2. a kind of target association method for the multiple-camera supervisory control system according to claim 1, it is characterized in that: described by repeatedly training, utilizing repeatedly, training obtains the reasonable threshold interval of a plurality of systems, these interval logarithmic axis are covered, and number axis is divided into a plurality of intervals, the intermediate value in the interval that coverage is the highest is as the threshold value of system.

Images