CN102063727B - Covariance matching-based active contour tracking method - Google Patents

Abstract

The invention relates to a covariance matching-based active contour tracking method and belongs to the technical field of visual tracking. In the covariance matching-based active contour tracking method, an image area energy term is modeled by using non-Euclidean geometry. The method comprises the following steps of: manually initializing a curve surrounding an objective and establishing a covariance matrix as a template of an objective contour for an area surrounded by the curve in a first frame; after the contour of the objective is obtained, recording a level set function value of the template to make preparation for a prior shape and calculating a symbolized distance function of the template; from the image of the next frame, deducing a gradient descent flow from a result of the previous frame according to the established energy functional and updating the level set function; and checking whether iteration stops or not. In the method, the tracking result is more accurate; meanwhile, the covariance matrix is used as an area descriptor and all kinds of information in an image sequence and the correlation between all kinds of information are considered comprehensively, and the method does not depend on foreground and background information distribution, so that the tracking method has universality.

Description

A kind of active contour tracing method based on the covariance coupling

Technical field

The present invention relates to a kind of active contour tracing method, belong to the vision track technical field based on the covariance coupling.

Background technology

Motion target tracking is one of classical problem of computer vision field, has important use to be worth.In the real world applications, owing to image quality is not good, ambient lighting variation, shade, block with reasons such as target distortion and make tracking problem become a generally acknowledged challenging difficult problem.

Visual target tracking can be divided into following several big classifications: point target is followed the tracks of, and nuclear is followed the tracks of and profile is followed the tracks of.Point target is followed the tracks of and is adopted one or more points to come the modeling sensation target; Tracking based on nuclear has the advantage of computing simple and fast; In vision track, obtained in recent years to use widely; But target shape was expressed and is usually adopted such as ellipse in the middle of nuclear was followed the tracks of, and therefore simple geometric primitives such as rectangle are difficult to accurately extract Amoebida target complex outline.And the accurate extraction of sensation target shape is above-mentioned many key in application place of mentioning, has only profile to follow the tracks of curve tracking in other words and can provide this information.

The basic thought that skeleton pattern is followed the tracks of is to be plane curve with the Target Modeling in the sequence image---the active profile of promptly evolving, according to the correlativity of video information on room and time, confirm position and the attitude of target in video.Can skeleton pattern be divided into two types on parameter skeleton pattern and geometric profile model, the former curve representation has adopted the explicit parament description, and the latter then is embedded in the higher-dimension level of function with curve representation and concentrates (Level set).The former advantage is that computational complexity is lower, and shortcoming is not have topological adaptivity, and the latter's advantage is to possess topological adaptivity and calculation stability, and deficiency is that calculation cost is higher.

Terzopoulos has defined kinetic energy, potential energy, damping term respectively from the lagrangian dynamics principle; And the dynamic deformation model of deriving unified shape and moving and describe; Be expressed as that become and tool inertia for the moment dynamic outline, the equation of motion itself has been expressed a kind of follow-up mechanism based on dynamic balance, and shape constraining is general flatness constraint; Shortcoming is that system's dimension is too high, receives noise effect easily; Peterfreund has proposed the speed snake, and light stream estimation, active contour model, Kalman filtering are combined, and can effectively handle noise to a certain extent, block and complex background, and deficiency is that the distortion dimension is too high, calculation of complex; In order to solve the profile tracking problem of multi-peak observation model under the complex backgrounds such as foreign material; Isard has proposed the conditional probability propagation algorithm; Be the Condensation algorithm, the conditional probability propagation algorithm combines Stochastic Dynamics model and weight sampling along with advancing of time propagated about the complete probability distribution of shape and position, the target travel of the wave filter that is produced under can the tracking complex background of stalwartness; The deficiency of Condensation algorithm is: compare with kalman filter method; Calculation cost is still too high, if the measurement model modeling is unreasonable, then increases sample number in any case and also can not improve tracking performance.

The geometric profile model is proposed respectively by Malladi and Caselles the earliest, and thought comes from solution that Oshier and Sethian propose with the Level Set Method of curvature relevant speed along the border propagation problem that normal direction is advanced.Curve is the zero level set of higher-dimension function, and curve evolution is evolved by the higher-dimension function and tried to achieve.The equation of motion that produces can approach calculating by the finite difference method that satisfies entropy condition, and the speed term that from image, obtains is used for the propagation in image border termination border.

Level Set Method is to be propagated in the research by the interface that J.Sethian and S.Osher propose progressively to grow up the earliest, and it is to handle sealing moving interface effective computational tool of how much change in topology in the evolutionary process in time.The basic thought of Level Set Method is the level set that is expressed as two-dimentional toroidal function with plane closed curve implicit expression, and the point set that promptly has the same functions value is through the motion of finding the solution curve of curved surface evolution implicit expression.The advantage of this method is the processing changes of topology structure that it can nature, and stable numerical algorithm is provided.

The basic process of covariance track algorithm is following:

1) to each two field picture of video sequence, extracts corresponding characteristic pattern.To an any given two field picture I (x, y) (image can be one dimension gray level image, three-dimensional color image or the like) established its width and highly is respectively W and H; To figure in each pixel, for its set up a proper vector f (x, y); It can be by pixel coordinate, image color or half-tone information, the image gradient of image; Elements such as image edge direction constitute, and for example, can get remember it dimension be the d dimension; Thus; Can from image I (x, extract in y) a W * H * d dimension characteristic pattern F (x, y)

F(x,y)＝(I,x,y)

Obviously, f (x, y) ∈ F (x, y).

2), set up a d * d dimension covariance matrix C to target area R given in the image _RAs region description

$C_R=\frac{{\∫}_R(f(x,y)-\stackrel{\‾}{f_R}){(f(x,y)-\stackrel{\‾}{f_R})}^T\mathrm{dxdy}}{{\∫}_R\mathrm{dxdy}}$

Wherein,

is the average of all pixel character pair vectors of region R, is designated as

3) a given template covariance matrix C _TCovariance matrix C with the present frame candidate target region _R, can find correct target area through the two covariance distance of minimization.

Summary of the invention

The objective of the invention is on the basis that existing active profile is followed the tracks of; The characteristic covariance matrix is sub as region description; Proposed a kind of active contour tracing method, utilized non-euclidean geometry modeling image-region energy term, made tracking results more accurate based on the covariance coupling.

The objective of the invention is to realize through following technical proposals.

A kind of active contour tracing method of the present invention based on the covariance coupling, the detailed process of this method is:

When template:

Utilize non-euclidean geometry method to establish the image energy model under the level set framework of template, derive the gradient katabatic drainage, upgrade level set function then according to the image energy model functional of setting up;

Image energy model functional under the level set framework is defined as:

${\mathrm{\ρ}}_{i,1}\left(\mathrm{\φ}\right)={\mathrm{\λ}}_1{\left|\right|\log C_R\left(\mathrm{\φ}\right)-\log C_T\left|\right|}_F-{\mathrm{\λ}}_2{\left|\right|\log C_{R^C}\left(\mathrm{\φ}\right)-\log C_T\left|\right|}_F$

λ ₁, λ ₂Be the weight of control prospect and background information, C _R(φ) with

The covariance of representing inside and outside zone of curve respectively, C _TThe covariance of representation template; Obviously, minimization energy functional ρ _{I, 1}The covariance that (φ) just means minimization candidate target region and template is apart from the covariance distance of candidate background zone with template that maximize simultaneously; Complete energy functional is:

ρ(φ)＝λ ₁ρ _i，11(φ)-λ ₂ρ _i,12(φ)+αρ _s(φ)+βρ _r(φ)

ρ in the formula _{I, 11}(φ), ρ _{I, 12}(φ), ρ _s(φ), ρ _r(φ) difference representative image energy foreground items, image energy background item, shape energy term and punishment energy term, λ ₁, λ ₂, α, β representes weight.

Concrete steps are following:

1) at first frame, manually the curve of target is surrounded in initialization, for the curve area surrounded is set up the template of covariance matrix as objective contour;

2) after the profile that has obtained target, the level set function value of logging template is that prior shape is prepared and the symbolism distance function of calculation template;

3) begin from the next frame image,, derive the gradient katabatic drainage, upgrade level set function then according to the energy functional of setting up by the result of previous frame;

Whether 4) detect iteration stops; Here the iteration stop criterion that adopts: (a) distance between the covariance matrix of the covariance matrix of the current zero level collection of calculating inside and priori objective contour inside; If less than a threshold value, iteration stops, otherwise continues iteration; (b) restriction of iterations;

5) forward step 3) to.

Above-mentioned steps 2) can also do following processing before: because possibly there is error in manual selection, can utilize traditional geometry active contour model (like the C-V model), make profile more near real target to the objective contour dividing processing.

When template does not exist:

Our can directly maximize degree of not matching of prospect covariance matrix and background covariance matrix, thus realize target following based on image segmentation.Thus, our second kind of image energy model may be defined as

${\mathrm{\ρ}}_{i,2}\left(\mathrm{\φ}\right)={\left|\right|\log C_R\left(\mathrm{\φ}\right)-\log C_{R^C}\left(\mathrm{\φ}\right)\left|\right|}_F$

According to following formula, when the covariance distance in candidate target region and candidate background zone reaches maximum, just can the target area be extracted from the background area.

Concrete steps are following:

1) read picture, the curve of target is surrounded in manual or semi-automatic initialization, and curve is rule or irregular, and can not be too far away from target;

2) be respectively the inside and outside covariance matrix of setting up of curve, and the compute sign distance function;

3) by the iteration result of last time,, upgrade level set function according to setting up the gradient katabatic drainage that energy functional is derived;

4) detect iteration and whether stop, the iteration stop criterion that adopts here: (a) distance between the covariance matrix of the covariance matrix of the current zero level collection of calculating inside and outside, if greater than a threshold value, iteration stops, otherwise continues iteration; (b) restriction of iterations (the capping number of times guarantees travelling speed).

5) forward step 3) to.

Beneficial effect

Algorithm proposed by the invention utilizes the image energy quantifier in the logarithm euclidian metric criterion modeling active profile trace model; This makes tracking results more accurate; Sub with covariance matrix simultaneously as region description; Taken all factors into consideration various information in the image sequence and correlativity each other, do not relied on distribution preceding, background information, made tracking have more universality.

Description of drawings

Fig. 1 is the tracking results figure of the embodiment of the invention.

Embodiment

Below in conjunction with embodiment the application of present technique scheme is specified, and provide the result.

Embodiment

Active contour tracing method based on the covariance coupling;

Adopting certain street automobile image sequence is tested object, and resolution is 180x135, whenever follows the tracks of once at a distance from a frame; Total tracking frame number is 17 frames, and the extracting waste car is as the target of following the tracks of, and the characteristics of this sequence are; There is certain rotational transform in the shape of target; In the automobile turning process, utilize this method not only can converge to the border of target accurately, the angle that can also estimate simultaneously to rotate, change in size etc.

Concrete steps are following:

1) at first frame, manually the curve of white car is surrounded in initialization, and for the curve area surrounded is set up the template of covariance matrix as white car profile, shown in (a) among Fig. 1, the symbolism distance function of template is shown in (b) among Fig. 1;

2) because there is error in manual the selection, utilize the C-V model, make profile more near real target to the objective contour dividing processing;

3) after the profile that has obtained white car, the level set function value of logging template is that prior shape is prepared and the symbolism distance function of calculation template;

4) begin from the next frame image,, derive the gradient katabatic drainage, upgrade level set function then according to the energy functional of setting up by the result of previous frame;

Whether 5) detect iteration stops; Here the iteration stop criterion that adopts: (a) distance between the covariance matrix of the covariance matrix of the current zero level collection of calculating inside and priori objective contour inside; If less than a threshold value, iteration stops, otherwise continues iteration; (b) restriction of iterations;

6) forward step 4) to;

The tracking results of the 5th frame that obtains according to said method, the 7th frame, the 9th frame, the 11st frame, the 13rd frame, the 17th frame like (c) among Fig. 1 to shown in (h).

Above-described specific descriptions; Purpose, technical scheme and beneficial effect to invention have carried out further explain, and institute it should be understood that the above is merely specific embodiment of the present invention; And be not used in qualification protection scope of the present invention; All within spirit of the present invention and principle, any modification of being made, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (3)

1. the active contour tracing method based on covariance coupling is characterized in that: establish the level set framework image energy model functional down of template, derive the gradient katabatic drainage according to the image energy model functional of foundation, upgrade level set function then;

Have the image energy model functional under the level set framework of template to be defined as:

${\mathrm{\ρ}}_{i,1}\left(\mathrm{\φ}\right)={\mathrm{\λ}}_1{\left|\right|\log C_R\left(\mathrm{\φ}\right)-\log C_T\left|\right|}_F-{\mathrm{\λ}}_2{\left|\right|\log C_{R^c}\left(\mathrm{\φ}\right)-\log C_T\left|\right|}_F$

λ ₁, λ ₂Be the weight of control prospect and background information, C _R(φ) with

The covariance of representing inside and outside zone of curve respectively, C _TThe covariance of representation template; Minimization energy functional ρ _{I, 1}(φ) be the covariance distance of the covariance of minimization candidate target region and template apart from maximize simultaneously candidate background zone and template;

Concrete steps are following:

1) at first frame, manually the curve of target is surrounded in initialization, sets up the covariance matrix of template for the curve area surrounded, utilizes how much active contour models to the objective contour dividing processing then, makes profile near real target;

2) after the profile that has obtained target, the level set function value of logging template and the symbolism distance function of calculation template;

3) begin from the next frame image,, derive the gradient katabatic drainage, upgrade level set function then according to the image energy model functional of setting up by the result of previous frame;

4) detect iteration and whether stop, the iteration stop criterion that adopts here: (a) distance between the covariance matrix of the covariance matrix of the current zero level collection of calculating inside and template, if less than a threshold value, iteration stops, otherwise continues iteration; (b) restriction of iterations;

5) forward step 3) to.

2. a kind of active contour tracing method based on covariance coupling according to claim 1 is characterized in that: utilize non-euclidean geometry method to establish the image energy model functional under the level set framework of template.

3. a kind of active contour tracing method based on the covariance coupling according to claim 2 is characterized in that: how much active contour models are the C-V model.

Images