CN102682453A - Moving vehicle tracking method based on multi-feature fusion - Google Patents

Abstract

The invention provides a moving vehicle tracking method based on multi-feature fusion, which comprises the following steps: 1) a feature extraction stage: a) characteristics of detected moving vehicles are extracted and b) the step 1) is finished; and 2) a vehicle tracking stage: a) the state of each vehicle in the previous frame of image is updated, b) matched connection is established between moving vehicles in the current frame of image and the moving vehicles in the previous frame of image, c) a matching result is returned, and d) the step 2) is finished. By adopting the moving vehicle tracking method based on multi-feature fusion, which is provided by the invention, the tracking anti-occlusion performance is improved, and further the tracking accuracy is improved.

Description

Moving vehicle tracking based on multi-feature fusion

Technical field

The present invention relates to a kind of moving vehicle tracking, relate in particular to a kind of moving vehicle tracking based on multi-feature fusion.

Background technology

Intelligent traffic management systems (ITS:Intelligent Traffic System) is the focus that present countries in the world traffic and transport field is is competitively researched and developed.Its main target is to obtain road information and vehicle behavioural information, comprises vehicle flowrate, the speed of a motor vehicle, roadway occupancy, traffic hazard detection etc.The real-time follow-up of vehicle is one of important technology of intelligent transportation system, is the core methed of realizing the ITS target, also is that ITS realizes robotization, intellectuality and real-time key in application.

Motion target tracking is detected foreground target to be carried out real-time follow-up disappear in image sequence until it, thereby describes the movement locus of target, extracts its behavioral characteristics (like position, speed etc.).Tracking is equivalent to the object matching that parameters such as position-based, shape, color, size are carried out between sequence image.It is concrete an application of motion target tracking that moving vehicle is followed the tracks of.

Because in the video image of monitoring, the phenomenon that vehicle blocks often takes place, this has caused very big difficulty to vehicle tracking.Therefore find an anti-good vehicle tracking method of blocking property to have crucial meaning.

Tracking based on characteristic is wherein a kind of method of conventional motion method for tracking target; Utilize the characteristics such as barycenter, girth, curvature and each rank distance of moving target, adopt the method for pattern-recognition, speed is fast; Discrimination is high; But be to use single characteristic limited, especially after blocking the interference disappearance, be difficult to return to again correct tracking the processing power of blocking.

Summary of the invention

The present invention seeks to: technical matters to be solved by this invention provides a kind of moving vehicle tracking based on many Feature Fusion couplings, to improve the antijamming capability of following the tracks of blocking, improves the correctness to vehicle tracking.

For addressing the above problem, technical scheme of the present invention is: the moving vehicle tracking based on many Feature Fusion couplings comprises the steps:

1) feature extraction phases:

A) extract detected moving vehicle characteristic, said characteristic comprises the area of barycenter, region area and the boundary rectangle of vehicle.If the zone in the image of vehicle place is R, S is a region area, then

(x y) is the i.e. capable y row of x in the position of a certain pixel in image; The coordinate of barycenter does Then $\stackrel{\‾}{x}=\frac{1}{S}\underset{(x,y)\∈R}{\mathrm{\Σ}}x,$ $\stackrel{\‾}{y}=\frac{1}{S}\underset{(x,y)\∈R}{\mathrm{\Σ}}y;$ X is the wide of boundary rectangle, and Y is the length of boundary rectangle, then $X=\underset{(x,y)\∈R}{\mathrm{Max}}\left(x\right)-\underset{(x,y)\∈R}{\mathrm{Min}}\left(x\right),$ $Y=\underset{(x,y)\∈R}{\mathrm{Max}}\left(y\right)-\underset{(x,y)\∈R}{\mathrm{Min}}\left(y\right),$ The area of boundary rectangle is X * Y;

B) finish

2) the vehicle tracking stage

A) the history feature information of the target vehicle of each tracking forms a tracking chain.Search out the tracking chain of optimum matching as new vehicle target after, this vehicle target is joined in the tracking chain.And with the vehicle characteristics of up-to-date adding characteristic as whole tracking chain.This stage is for collecting the characteristic that all follow the tracks of chain;

B) new vehicle target come interim, for fresh target and former tracking chain carry out characteristic matching;

C) return matching result;

D) finish.

Said characteristic comprises the area of barycenter, region area and the boundary rectangle of vehicle among the step 1-a, and barycenter is the position of the row and column that in image, belongs to of vehicle barycenter, and area calculates the method for using the statistical pixel number.

After referring among the step 2-a that vehicle searches out corresponding coupling in the former frame image in for present frame during said update mode; With the characteristic of this vehicle in the present frame new feature, for participating in the required new feature of follow-up coupling as this object chain (object chain is meant the historical track of each the moving target vehicle that traces into)

The detailed process of step 2-b is following:

1) statistics obtained all follow the tracks of the characteristic of chains, the characteristic of each object chain is a standard with the up-to-date target signature that matches, and all characteristics is formed one do not mate formation Q ₁

2) characteristic of the initiate target vehicle of statistics, the vehicle target that all are new forms a formation Q to be matched ₂

3) use computes formation Q ₂Do not mate formation Q ₁In the adaptation function value E of all characteristics (m, n), (m n) is Q to E in the formula ₁In m characteristic and Q ₂In the adaptation function value of n characteristic, (m n) is the normalization centroid distance to Δ D, and (m n) is the normalization region area to Δ S, and (m n) is normalization boundary rectangle area to Δ R; α, β and γ are the weight of each characteristic;

E(m，n)＝αΔD(m，n)+βΔS(m，n)+γΔR(m，n)

4) find out in all adaptation function values minimum one and be current optimum matching, Q ₂In this target add in the corresponding tracking chain, the residue matching times Count of this trackings chain is set, the while is at Q ₁And Q ₂The target vehicle that deletion has been mated in the formation; If the formation of coupling and formation to be matched, i.e. Q are arranged ₁And Q ₂Formation all is not empty, then forwards 3 to), otherwise forward 5 to);

5) if do not mate formation Q ₁Be sky, then with formation Q to be matched ₂In all target vehicles that do not mate add to follow the tracks of in the chain as fresh target, the residue matching times Count of new tracking chain is set; If formation Q to be matched ₂Be sky, do not mate formation Q ₁Be not empty, upgrade and do not mate formation Q ₁In the residue matching times Count of each target, make Count subtract 1, if Count is 0, deletion should be followed the tracks of chain;

6) finish.

Beneficial effect of the present invention is: the moving vehicle tracking based on many Feature Fusion couplings of the present invention; Adopted many characteristic matching to follow the tracks of; Simultaneously be that current chain is provided with residue matching times mechanism when the current goal chain does not search out fresh target; After target occlusion recovers, still can successful match to target, improved the anti-performance of blocking of vehicle tracking method.

Description of drawings:

Fig. 1 is the Intelligent traffic management systems workflow diagram.

Fig. 2 is the process flow diagram of the moving vehicle detection method based on many Feature Fusion coupling of the present invention.

Fig. 3 is based on many Feature Fusion Matching Algorithm process flow diagram.

Embodiment

Below in conjunction with accompanying drawing the present invention is elaborated.

As shown in Figure 1; Intelligent traffic management systems is obtained vedio data through video image acquisition equipment; Through the image pre-service, moving vehicle detects, and carries out moving vehicle then and follows the tracks of; The result of vehicle tracking can analyze according to the pursuit path to vehicle, with this foundation as vehicle identification and vehicle behavior identification.

Vehicle tracking is the core procedure of Intelligent traffic management systems, and thinking of the present invention is exactly the antijamming capability through the raising vehicle tracking, thereby improves the performance of whole Intelligent traffic management systems.The vehicle tracking process is the process flow diagram of the moving vehicle detection method based on many Feature Fusion coupling of the present invention, and is as shown in Figure 2.

Step 0 is an initial state of the present invention;

(step 1) is extracted the characteristic that barycenter, region area and the boundary rectangle area of each moving vehicle in every frame picture are portrayed each moving vehicle in feature extraction phases.

At characteristic matching stage (step 2-4), step 2: the moving vehicle characteristic information of collecting current all object chains;

Step 3: collect the characteristic information of each moving vehicle in the current image frame, the characteristic information of collecting in the object chain in the integrating step 2 then carries out many Feature Fusion and calculates.

Step 4 forms the historical track information of each moving vehicle for the target that can mate adds object chain;

Step 5 is the end step of the moving vehicle tracking based on many Feature Fusion coupling of the present invention.

Fig. 3 is the detailed description of many characteristic matching stage algorithm

Step 30 is an initial step;

Step 31 is feature extraction;

Step 32 is to collect the vehicle characteristics of current goal chain to not mating formation Q ₁In, the vehicle characteristics that collection step 31 is extracted is to formation Q to be matched ₂In;

Step 33 is judged the Q of current queue ₁And Q ₂State, if all be not empty, then execution in step 34, otherwise execution in step 37;

Step 34 is according to formation Q ₁And Q ₂, calculate Q ₁In arbitrary vehicle and Q ₂In the adaptation function value of arbitrary vehicle;

Step 35 is sought one minimum in all adaptation function values, the Q that this adaptation function value is related to ₁In certain vehicle target and Q ₂In certain vehicle target be made as optimum matching, with Q ₂In this vehicle join Q ₁In the object chain of middle coupling vehicle correspondence;

Step 36 is for deleting at Q ₁And Q ₂Two target signatures that middle deletion matches through step 35;

Step 37 is for judging whether Q ₁Be empty (all object chains have found new coupling vehicle), Q ₂Be not empty (also having vehicle not join in the vehicle tracking chain in the present frame).If then turn to 38; Otherwise, turn to 39;

Step 38 is Q ₂In all targets set up new object chain, and the Count value is set;

Step 39 is for judging whether Q ₁Be not empty (also having object chain not find new coupling vehicle as yet), Q ₂Be empty (all vehicles all join in the vehicle tracking chain in the present frame); If then turn to 3a; Otherwise turn to 3b;

Step 3a is to Q ₁In can not find the coupling target object chain, its Count value is subtracted 1, as if the Count value be 0, then delete this object chain; Step 3b is the done state of algorithm.

Claims (1)

1. moving vehicle tracking based on multi-feature fusion is characterized in that comprising the steps:

1) feature extraction phases:

A) extract detected moving vehicle characteristic, said characteristic comprises the area of barycenter, region area and the boundary rectangle of vehicle; If the zone in the image of vehicle place is R, S is a region area, then

For the position of a certain pixel in image is the capable y row of x; The coordinate of barycenter does

Then

X is the wide of boundary rectangle, and Y is the length of boundary rectangle, then $X=\underset{(x,y)\∈R}{\mathrm{Max}}\left(x\right)-\underset{(x,y)\∈R}{\mathrm{Min}}\left(x\right),$ $Y=\underset{(x,y)\∈R}{\mathrm{Max}}\left(y\right)-\underset{(x,y)\∈R}{\mathrm{Min}}\left(y\right),$ The area of boundary rectangle is X * Y;

B) finish.

2) the vehicle tracking stage

A) the history feature information of the target vehicle of each tracking forms a tracking chain.Search out the tracking chain of optimum matching as new vehicle target after, this vehicle target is joined in the tracking chain.And with the vehicle characteristics of up-to-date adding characteristic as whole tracking chain; This stage is for collecting the characteristic that all follow the tracks of chain;

B) new vehicle target come interim, for fresh target and former tracking chain carry out characteristic matching;

C) return matching result;

D) finish;

Step 2)-detailed process of b is following:

1) statistics obtained all follow the tracks of the characteristic of chains, the characteristic of each object chain is a standard with the up-to-date target signature that matches, and all characteristics is formed one do not mate formation Q ₁

2) characteristic of the initiate target vehicle of statistics, the vehicle target that all are new forms a formation Q to be matched ₂

3) use computes formation Q ₂Do not mate formation Q ₁In the adaptation function value E of all characteristics (m, n), (m n) is Q to E in the formula ₁In m characteristic and Q ₂In the adaptation function value of n characteristic, (m n) is the normalization centroid distance to Δ D, and (m n) is the normalization region area to Δ S, and (m n) is normalization boundary rectangle area to Δ R; α, β and γ are the weight of each characteristic;

E(m，n)＝αΔD(m，n)+βΔS(m，n)+γΔR(m，n)

4) find out in all adaptation function values minimum one and be current optimum matching, Q ₂In this target add in the corresponding tracking chain, the residue matching times Count of this trackings chain is set, the while is at Q ₁And Q ₂The target vehicle that deletion has been mated in the formation; If the formation of coupling and formation to be matched, i.e. Q are arranged ₁And Q ₂Formation all is not empty, then forwards 3 to), otherwise forward 5 to);

5) if do not mate formation Q ₁Be sky, then with formation Q to be matched ₂In all target vehicles that do not mate add to follow the tracks of in the chain as fresh target, the residue matching times Count of new tracking chain is set; If formation Q to be matched ₂Be sky, do not mate formation Q ₁Be not empty, upgrade and do not mate formation Q ₁In the residue matching times Count of each target, make Count subtract 1, if Count is 0, deletion should be followed the tracks of chain;

6) finish.

Images