CN101408982A - Object-tracking method base on particle filtering and movable contour model - Google Patents

Abstract

The invention provides a target tracing method which is based on a particle filter and active contour model, and relates to a target tracing method for a gradient vector flow-parameter active contour model. The method firstly adopts a background difference method to obtain target initial profile, and the improved gradient vector flow-parameter active contour model is utilized, thus leading the parameter active contour model to be converged to true profile of a moving target; control points are increased or reduced according to the distance of the control points, so as to achieve the purpose of tracing moving and morph targets in a self-adapting way; then energetic particle filter target tracing algorithm is used for tracing the targets by combining the particle filter and the improved gradient vector flow-parameter active contour model; and tracing strategy when the target is sheltered is used, thus overcoming the effect of sheltering in the tracing process. In a whole tracing process, profile point information is not sheltered mostly by fully utilizing GVF-Snake model target profile, so as to effectively overcoming the effect such as complex environment and the like in the tracing process.

Description

A kind of method for tracking target based on particle filter and movable contour model

Technical field

The present invention be more particularly directed to particle filter and gradient vector flow-parameter movable contour model and carry out the method for target following, belong to the image processing and pattern recognition field.

Background technology

Pursuit movement and distortion target are to the tracking of rigidity and non-rigid target under the research difficult point, particularly complex environment of target following always.In recent years, a lot of to the research of the motion target tracking method of video image, be to emerge many methods that realistic meaning is arranged aspect the target following of background with active contour and intelligent transportation especially.Profile information has the unchangeability than robust, and is insensitive to illumination variation, and target is walked upwards insensitive to error in motion process at the edge.Curved profile based on the track algorithm utilization of active contour sealing is represented tracked target, and this profile can adaptive updates to realize Continuous Tracking to tracked target.

Movable contour model can be divided into parameter movable contour model and geometric active contour model, obtains application more and more widely in fields such as graphical analysis and computer visions.Target tracking algorism based on movable contour model can be divided into two classes:

Based on parameter movable contour model and Kalman filtering target tracking algorism, under complex environment to the motion and the distortion target follow the tracks of.These class methods are obtained at initial profile and are difficult to correct convergence under the bad situation, and this moment, tracking effect was relatively poor.

Based on the method for particle filter and level set cut apart the geometric active contour model method for tracking target, under complex environment to the motion and the distortion target follow the tracks of.Shortcomings such as there is the algorithm complexity in these these class methods, and calculated amount is big are unsuitable for real-time implementation.

Summary of the invention

Technical matters: the objective of the invention is to propose method for tracking target based on particle filter and gradient vector flow-parameter movable contour model.Tracking strategy when the target that the present invention proposes is blocked makes full use of the contour point information that GVF-Snake simulated target profile major part is not blocked in whole tracing process, effectively overcome the influence of complex environments such as blocking in the tracing process.

Technical scheme: the method for tracking target based on particle filter and movable contour model of the present invention is: at first adopt the background subtraction point-score to obtain the target initial profile, utilize improved gradient vector flow-parameter movable contour model, make the parameter movable contour model converge to the true profile of moving target; And according to the reference mark apart from the additions and deletions reference mark, reach pursuit movement and Amoebida target purpose adaptively; By in conjunction with particle filter and improved gradient vector flow-parameter movable contour model, use energy particle filtering target tracking algorism that target is followed the tracks of then, and the tracking strategy when using target to be blocked, overcome the influence of blocking in the tracing process.

Described improved gradient vector flow-parameter movable contour model, the polygon center of gravity of forming with each snake point on the parameter movable contour model is a control center, control energy is elected the absolute value that snake puts centroidal distance as.

Described reference mark adopt adaptive additions and deletions point algorithm apart from the additions and deletions reference mark, the number of snake point should be able to increase or reduce adaptively.

Described energy particle filtering target tracking algorism is promptly observed density function by the definition profile apart from observation density function and profile energy, obtains energy particle filtering target tracking algorism thus.

Tracking strategy when described target is blocked, promptly objective contour rectangular window algorithm is realized the tracking of the objective contour that is blocked is handled.

Beneficial effect: innovative point of the present invention is as follows:

The present invention proposes the method for tracking target based on particle filter and gradient vector flow-parameter movable contour model.In recent years, having become the research focus of target tracking domain based on the target tracking algorism of particle filter, also is the research difficult point in the non-linear filtering of self-adaptation field simultaneously.And in conjunction with particle filter target tracking algorism and parameter movable contour model, motion and distortion target are followed the tracks of, as far as we know, also there is not the achievement in research of this respect at present both at home and abroad.

The present invention at first by research GVF-Snake movable contour model, has proposed improved GVF-Snake method, utilizes the distinctive powerful search capability of GVF-Snake, makes Snake converge to the true profile of moving target.The singularity of following the tracks of under complex environment at target video stream by additions and deletions reference mark adaptively, changes situations such as (rigid deformation), shade (non-rigid deformation) and target are blocked to adapt to target sizes.

Under non-linear non-Gauss's situation, has tracking effect based on particle filter than robust, the present invention proposes a kind of new non-linear filtering algorithm of the self-adaptation in conjunction with particle filter and improved GVF-Snake movable contour model, energy particle filtering (EPF) target tracking algorism is followed the tracks of motion and distortion target.

Tracking strategy when the target that the present invention proposes is blocked makes full use of the contour point information that GVF-Snake simulated target profile major part is not blocked in whole tracing process, effectively handle tracked target in the tracking problem of blocking for a long time under the situation.

Embodiment

Method for tracking target based on particle filter and movable contour model of the present invention is: at first adopt the background subtraction point-score to obtain the target initial profile, utilize improved gradient vector flow-parameter movable contour model, make the parameter movable contour model converge to the true profile of moving target; And according to the reference mark apart from the additions and deletions reference mark, reach pursuit movement and Amoebida target purpose adaptively; By in conjunction with particle filter and improved gradient vector flow-parameter movable contour model, use energy particle filtering target tracking algorism that target is followed the tracks of then, and the tracking strategy when using target to be blocked, overcome the influence of blocking in the tracing process.

The present invention adopts the background subtraction point-score to obtain the target initial profile, by improved GVF-Snake model, utilizes the distinctive powerful search capability of GVF-Snake, makes Snake converge to the true profile of moving target.The singularity of following the tracks of under complex environment at target video stream by additions and deletions reference mark adaptively, changes to adapt to target sizes, thereby can reach fast and reliable ground pursuit movement and Amoebida target purpose under complex environment.

1. based on the object module of GVF-Snake

Under the static condition of video camera, the present invention adopts the background subtraction point-score to obtain the target initial profile.After having obtained initial profile, we use the GVF-Snake algorithm makes its convergence obtain the convergence profile of target.For restraining quickly and accurately, the characteristics of combining target video flowing on traditional GVF-Snake model basis have proposed the improvement algorithm at the GVF-Snake model of distortion target following.

The polygon center of gravity that goes up each snake point composition with Snake is a control center, and control energy is elected the absolute value that snake puts centroidal distance as, promptly

E _ctrl(v _j)＝π _j|v _j-C|

$C=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{v}_j---\left(1\right)$

π in the formula _jBe control coefrficient, v _jBe j snake point coordinate position, C is a center of gravity.Change the symbol of control coefrficient, can change the direction of motion of Snake curve, this to initial profile because light changes and shade influences to obtain still can restrain preferably under the bad situation and plays an important role.Consider the characteristics of traffic video, this paper method judges in conjunction with the above-mentioned background modeling whether current snake point is impact point, and then judges that the symbol of control coefrficient is positive and negative.If judging current snake point is the foreground point, then the control energy coefficient that will put be made as negative, thereby make this put outside turgor movement; Otherwise current snake point is a background dot, and then the control energy coefficient that will put just is made as, thereby makes this put inside contractile motion.Its formula is as follows:

Obtain the complete formula of Snake energy function thus:

E _snake＝E _int+E _ext+E _ctrl (3)

Be influences such as the distortion that adapts to target and target are blocked, the number of snake point should be able to increase adaptively or reduce, and this paper method adopts adaptive additions and deletions point algorithm.Its additions and deletions point principle is: 1. if | v _j-v _J-1| ²Excessive then should be at v _jWith v _J-1New snake point of middle increase; 2. if | v _J-1-2v _j+ v _J+1| ²Excessive then should be at v _jWith v _J-1Or v _J+1The middle snake point that increases; 3. if | v _J-1-2v _j+ v _J+1| ², | v _j-v _J-1| ²And | v _J+1-v _j| ²Snake point v in the middle of all less then deletion _j

2. based on the target tracking algorism of energy particle filtering (EPF)

For moving target is followed the tracks of, make Snake converge to the true profile of moving target fast, can be according to the continuity of Snake profile along time, spatial axes displacement and distortion, from the true profile of having followed the tracks of of target, analyze the variation tendency of objective contour, thereby dope the position and the shape of objective contour in the next frame image, with this as the initial profile of target at next frame.Target tracking algorism based on particle filter is to go on foot standard estimation procedures (prediction-correction) by two of recurrence, and promptly system state shifts and the observation model equation, and motion and distortion target are followed the tracks of.By in conjunction with improved GVF-Snake movable contour model, obtain a kind of new energy particle filtering (EPF) target tracking algorism, and the tracking strategy when having proposed that target is blocked.

Discuss respectively below from system state and shift and the observation model equation, set up the particle filter target tracking algorism, and the tracking strategy of target when being blocked.

2.1 system state shifts and observation model

The priori of target is that profile is described, and represents with the Snake movable contour model, just the coordinate position of the some control snake points on the known target profile.Because curve simulates by control snake dot information, obviously control snake point and should be positioned at the bigger position of objective contour turnover as far as possible.Get the snake j=1 that counts ..., N, population is i=1 ..., N _s, its weights ω _i, initial weight is 1/N _s, each particle is represented a possibility state of target, and their initial value is artificially given.

The variation of objective contour can be summed up as translation and distortion, and the translation of profile and distortion are regarded as the motion of reference mark under certain speed and acceleration, and according to the theory of dynamic system, the system state equation of transfer can adopt second order ARP model, promptly

$x_{k,j}^i={\mathrm{Ax}}_{k-2,j}^i+{\mathrm{Bx}}_{k-1,j}^i+{\mathrm{Cu}}_{k-1,j}^i---\left(4\right)$

U wherein _{K-1, j} ⁱRandom noise for system.Each particle carries out just can carrying out systematic observation to it after the state transitions, systematic observation is exactly the similarity degree of observing between the target of each particle representative possibility profile and the true profile of target, give bigger weights near the particle of the true profile of target, otherwise weights are less.

By observation data z _k, system state x _k, the definition profile apart from the observation density function is:

$p(z_k/x_k^{\left(i\right)})\∝\exp \left(\frac{-{\mathrm{DIS}}^{\left(i\right)2}}{{{\mathrm{\σ}}_d}^2}\right)---\left(5\right)$

Wherein, σ _dBe profile distance variance parameter, expression is carried out Gauss's modulation to profile apart from correlation.Because the ENERGY E of controlled target profile distortion _SnakeBe inversely proportional to Grad, promptly the Grad at the true profile of target place is big more, E _SnakeMore little, can further define profile energy observation density function p (z _k| x _k ⁽ⁱ⁾) be:

$p(z_k/x_k^{\left(i\right)})\∝\exp \left(\frac{-E_{\mathrm{snake}}^{\left(i\right)}}{{{\mathrm{\σ}}_s}^2}\right)---\left(6\right)$

Wherein, σ _sBe profile energy variance parameter, expression is carried out Gauss's modulation to profile energy correlation.Comprehensive above-mentioned discussion, weights further can be derived and obtain following formula:

${\mathrm{\ω}}_{k,j}^{\left(i\right)}={\mathrm{\ω}}_{k-1,j}^{\left(i\right)}p(z_k/x_k^{\left(i\right)})\∝{\mathrm{\ω}}_{k-1,j}^{\left(i\right)}\exp \left(\frac{-E_{\mathrm{snake}}^{\left(i\right)}}{{{\mathrm{\σ}}_s}^2}\right)\exp \left(\frac{{-\mathrm{DIS}}^{\left(i\right)2}}{{{\mathrm{\σ}}_d}^2}\right)---\left(7\right)$

Can be with profile distance and profile energy normalized, like this systematic observation probability density function redefine into:

$p\left(z_k\right|x_k^{\left(i\right)})\∝{\mathrm{\λ}}_1\left(\frac{e^{\frac{-E_{\mathrm{snake}}^{\left(i\right)}}{{{\mathrm{\σ}}_s}^2}}}{{\mathrm{\Σ}}_{i=1}^{N_S}{e}^{\frac{-E_{\mathrm{snake}}^{\left(i\right)}}{{{\mathrm{\σ}}_s}^2}}}\right)+{\mathrm{\λ}}_2(1-\frac{{\mathrm{DIS}}^{\left(i\right)2}}{{\mathrm{\Σ}}_{i=1}^{N_s}{\mathrm{DIS}}^{\left(i\right)2}})---\left(8\right)$

Following formula is energy particle filtering (EPF) equation, wherein λ ₁+ λ ₂=1.E _SnakeAnd DIS _jBe worth greatly more, dissimilar degree is just high more.The weights of each particle are still by formula like this ${\mathrm{\ω}}_k^{\left(i\right)}={\mathrm{\ω}}_{k-1}^{\left(i\right)}p\left(z_k\right|x_k^{\left(i\right)})$ Carrying out recursion calculates.

The tracking strategy when 2.2 target is blocked

When moving target during by partial occlusion, Partial Feature disappears owing to blocking, and the objective contour edge can't accurately be located and since the zone that is blocked to influence tracking accuracy very low.The present invention proposes a kind of objective contour rectangular window algorithm and profile and block update rule, realize the tracking of the objective contour that is blocked is handled.

Here we only consider the length and the width information of target, on the basis that accurately obtains the tracked target profile, directly from the objective contour image, extract length and width degree parameter, by image array f (i, j) cycle calculations goes out the objective contour point on the most upper and lower a, left side and the right, the coordinate that record is obtained, obtaining length is M, width is the distributed data of M * N objective contour rectangular window of N.

When target is blocked, according to the positional information of the objective contour that is blocked, target is blocked the tax of part value accordingly for invalid, promptly remove the false contouring point when blocking, as shown in Figure 1.By R _m(x _j, y _j) determine that point is effective in the window, point is invalid outside window, by the adaptive additions and deletions point algorithm of GVF-Snake, effectively handles the connectivity problem when blocking.(seeing shown in the accompanying drawing 1)

When moving target is blocked part when big, can adopt profile to block update rule, use R _pJudge whether to upgrade profile.Making the interior effective contour point of profile rectangular window is p with blocking front profile point ratio, according to the data that obtained, and setting threshold th2.If the p value is bigger, the part that is blocked is less, and the p value is upgraded profile greater than threshold value th2; If the p value is too small, the part that is blocked is bigger, and the p value is not upgraded profile less than threshold value th2.

$R_p=\left\{\begin{array}{cc}1& p>\mathrm{th}2\\ 0& p\≤\mathrm{th}2\end{array}\right.---\left(10\right)$

Example:

(1) initialization

K=0 is set, objective contour is chosen j=1 ..., N snake point produces N _sIndividual sample x _{0, j} ⁱ

(2) system state shifts k=1

Getting population is N _s, its weights ω _{K, j} ⁱInitial value be 1/N _s, state transition equation is: $x_{k,j}^i={\mathrm{Ax}}_{k-2,j}^i+{\mathrm{Bx}}_{k-1,j}^i+{\mathrm{Cu}}_{k-1,j}^i$

(3) systematic observation

The systematic observation probability density function is: $p\left(z_k\right|x_k^{\left(i\right)})\∝{\mathrm{\λ}}_1\left(\frac{e^{\frac{-E_{\mathrm{snake}}^{\left(i\right)}}{{{\mathrm{\σ}}_s}^2}}}{{\mathrm{\Σ}}_{i=1}^{N_S}{e}^{\frac{-E_{\mathrm{snake}}^{\left(i\right)}}{{{\mathrm{\σ}}_s}^2}}}\right)+{\mathrm{\λ}}_2(1-\frac{{\mathrm{DIS}}^{\left(i\right)2}}{{\mathrm{\Σ}}_{i=1}^{N_s}{\mathrm{DIS}}^{\left(i\right)2}})$

(4) posterior probability is calculated

To i=1 ..., N _sIndividual particle, the weights of calculating particle: ${\mathrm{\ω}}_k^{\left(i\right)}={\mathrm{\ω}}_{k-1}^{\left(i\right)}p\left(z_k\right|x_k^{\left(i\right)});$ Calculate the normalization weights of each target snake point: ${\widetilde{\mathrm{\ω}}}_{k,j}^{\left(i\right)}=\frac{{\mathrm{\ω}}_{k,j}^{\left(i\right)}}{{\mathrm{\Σ}}_{i=1}^{\mathrm{Ns}}{\mathrm{\ω}}_{k,j}^{\left(i\right)}}.$ The state estimation value

Adopt weighted criterion: ${\hat{x}}_{k,j}=\underset{i=1}{\overset{N_s}{\mathrm{\Σ}}}{x}_{k,j}^{\left(i\right)}{\widetilde{\mathrm{\ω}}}_{k,j}^{\left(i\right)}$

(5) profile rectangular window algorithm

Calculate the distributed data of M * N objective contour rectangular window.When target is blocked, target is blocked the tax of part value accordingly for invalid.Ratio calculated p determines whether to upgrade profile.

(6) if ${\hat{N}}_{\mathrm{eff}}<N_{\mathrm{th}},$ Right

Resample.

Claims (5)

1. method for tracking target based on particle filter and movable contour model, it is characterized in that, at first adopt the background subtraction point-score to obtain the target initial profile, utilize improved gradient vector flow-parameter movable contour model, make the parameter movable contour model converge to the true profile of moving target; And according to the reference mark apart from the additions and deletions reference mark, reach pursuit movement and Amoebida target purpose adaptively; By in conjunction with particle filter and improved gradient vector flow-parameter movable contour model, use energy particle filtering target tracking algorism that target is followed the tracks of then, and the tracking strategy when using target to be blocked, overcome the influence of blocking in the tracing process.

2. a kind of method for tracking target according to claim 1 based on particle filter and movable contour model, it is characterized in that described improved gradient vector flow-parameter movable contour model, the polygon center of gravity of forming with each snake point on the parameter movable contour model is a control center, and control energy is elected the absolute value that snake puts centroidal distance as.

3. a kind of method for tracking target according to claim 1 based on particle filter and movable contour model, what it is characterized in that described reference mark adopts adaptive additions and deletions point algorithm apart from the additions and deletions reference mark, and the number of snake point should be able to increase or reduce adaptively.

4. a kind of method for tracking target according to claim 1 based on particle filter and movable contour model, it is characterized in that described energy particle filtering target tracking algorism, promptly observe density function apart from observation density function and profile energy, obtain energy particle filtering target tracking algorism thus by the definition profile.

5. a kind of method for tracking target based on particle filter and movable contour model according to claim 1 is characterized in that the tracking strategy when described target is blocked, and promptly objective contour rectangular window algorithm is realized the tracking of the objective contour that is blocked is handled.