CN100587719C - Method for tracking dimension self-adaptation video target with low complex degree - Google Patents

Abstract

A scale self-adapting video target tracking method with low complexity in the video intelligent monitoring technical field, including: initializing the state of the particle sample; randomly generating two scale factors on each sampling point for the particle sample, computing the sample second autoegression center, storing as the mean drifting center; obtaining a mean drift field with neighborhood uniformity according to the second autoegression center of the sampling point in the sample set; building the important sampling density function for each sampling point, and obtaining the updatingstate X1' of the sample combining with the mean drifting method from the sampled probability angle of the Monte Carto, and updating the weight value w' under the state X1', then sampling again on thesample set {x1', w1'}, wherein 1=1...N, to obtain the posterior probability distribution dispersion estimation set {x1', w1'} of the object final state at the moment, wherein 1=1...N. The invention advances the tracking accuracy of the object scale space, reduces the computing complexity in the realtime video tracking.

Description

The method for tracking dimension self-adaption video target of low complex degree

Technical field

The present invention relates to a kind of method of video brainpower watch and control technical field, specifically is a kind of method for tracking dimension self-adaption video target of low complex degree.

Background technology

In many application of computer vision field,, all need video sequence interframe moving target is followed the tracks of as intelligent monitoring, robot vision, human-computer interaction interface.Because how the diversity of tracking target form and the uncertainty of target travel realize the real-time follow-up of robust under the various environment and change the focus that the reliable estimation that realizes its variable dimension is research with target range always.Sequential Monte Carlo filtering method is to use tracking widely in recent years, distributes with the posterior probability of target in state space and represents the state that the target most probable occurs, as position, size etc.Different with traditional Kalman filtering, this method is put to be similar to real dbjective state distribution function with one group of discrete sampling, so be also referred to as particle filter.The advantage of sequential Monte Carlo filtering method is that it is applicable to non-Gauss's nonlinear system, but because calculated amount is bigger, has limited its application in real-time system.Average drifting is another kind of method for tracking target commonly used, the state that its target most probable in the gradient direction local iteration searching image of dbjective state probability function has.1998, people such as Bradski improved basic average drifting method, and the High Order Moment with the dbjective state probability distribution in the ferret out position obtains accurate target scale size more, i.e. the average drifting method of dimension self-adaption.This class methods fast operation, but because may converge on the local optimum point of state space is in the phenomenon that BREAK TRACK occurs blocking, takes place easily under the situation such as similar background objects interference.

Based on the advantage separately of filtering of sequential Monte Carlo and average drifting method, people such as Shan have proposed to embed the particle filter method of average drifting.This method is carried out independently average drifting to each sampled point in the wave filter of sequential Monte Carlo, makes it more concentrate on local optimum, thereby has improved sampling efficiency, has reduced computational complexity.

Find by prior art documents; Similarly method also once was used in work in recent years and improved to some extent, " Object tracking by the mean-shiftof regional color distribution combined with the particle-filteralgorithm " (based on the average drifting of regional color distribution and the target following method of particle filtering) delivered 506 to 509 pages of " the 17 pattern-recognition international conference " (Proceedings of the 17th International Conference on PatternRecognition) the 3rd volumes such as the people such as Koichiro in 2004. In this method in conjunction with particle filter and average drifting, each particle through status predication has all been carried out the optimization of average drifting, make the big state space of the sampled point more close pattern probability of set that obtains, thereby improved sampling efficiency, reduced required number of particles.On the other hand, directly be decided by the number of particle the operation time of particle filter, therefore this improvement has improved the speed of tracker simultaneously.But these improve and also have some shortcomings:particle is too concentrated and is caused the sample scarcity, so that the diversity of sequential Monte Carlo filtering is destroyed.Method in conjunction with filtering of sequential Monte Carlo and average drifting in its these work that have its source in is too simple, do not consider to embed of the influence of average drifting optimization means, mainly show as the particle state space and present new probability distribution original sequential Monte Carlo filter frame.So propose complete average drifting associated methods from the probability angle of Monte Carlo sampling, and make its evaluated error that in tracing process, effectively reduces target scale with the variation of target sizes, still blank at present.

Summary of the invention

The present invention is directed to above-mentioned the deficiencies in the prior art, a kind of method for tracking dimension self-adaption video target of low complex degree has been proposed, make it dimension self-adaption average drifting method directly is fused to the probability system of sequential Monte Carlo filtering by design importance sampling function, being the average drifting method is embodied in the importance sampling function the optimization function of sampled point, and the estimation that has guaranteed the dbjective state probability distribution is no inclined to one side.

The present invention is achieved through the following technical solutions, the present invention includes following steps:

Step 1, initialization particle sample state: by detecting the original state x that determines target ₀, calculate the measurement vector of target simultaneously, the number N of given particle sampler point, the initial state vector of all sampled points is x ₀, and compose with unified weights

Step 2, the particle sample of step 1 is got equally distributed stochastic variable u～U[0,1 to each sampled point i=1:N), if

The scale factor k of this sampled point ⁱBe taken as k ₁, otherwise be taken as k ₂, wherein: k ₁=1.2, k ₂=0.9. has characterized the variation characteristic that the interframe target scale amplifies and dwindles respectively, calculates sample second order autoregression central point, and stores as the central point of average drifting;

Described calculating sample second order autoregression central point is specially:

Make the current sampling point set be { x _t ⁱ, w _t ⁱ} _I=1...N, each sampled point i=1:N is calculated its second order autoregression center

${\stackrel{\·}{x}}_t^i=2x_{t-1}^i-x_{t-2}^i$

In the formula, x _t ⁱThe state vector of i particle under the expression moment t;

Step 3 according to the second order autoregression center of the sampled point in the sample set in the step 2, obtains having the conforming average drifting of neighborhood field, and is specific as follows:

1. at first judge each sampled point i=1:N, if i==1 or

Neighborhood Ω outside, then carry out successively 2., 3., 5. with step 6., otherwise carry out 4. successively, 5. with step 6.; Wherein,

Be respectively sampling point i and i-1 at t second order autoregression center constantly;

2. reference state P wherein ^Ref, s ^RefBe respectively the sub-state of position and yardstick, s ^Ref={ l ^Ref, h ^Ref, l ^Ref, h ^RefDescribed with p respectively ^RefLength and height for the target rectangle profile at center; Zero setting rank square semi-invariant m ₀₀(x ^Ref, n=0) initial value is 1, n=0:I;

3. n=1:I is circulated:

At first, computation of mean values drift vector M _{R, g}(x ^Ref):

$M_{r,g}\left(x^{\mathrm{ref}}\right)=\frac{\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{p}_{j}m\left(p_j\right)g\left({\left|\right|\frac{p^{\mathrm{ref}}-p_j}{r}\left|\right|}^2\right)}{\underset{j=1}{\overset{M}{\mathrm{\Σ}}}m\left(p_j\right)g\left({\left|\right|\frac{p^{\mathrm{ref}}-p_j}{r}\left|\right|}^2\right)}-p^{\mathrm{ref}}$

Wherein, { p _j} _J=1...MBe reference state x ^RefThe pixel coordinate that image rectangular area, place covers, m (p _j) be position p _jThe possibility weights, when measurement vector is color histogram, by x ^RefMatching degree between the color of object histogram of state correspondence and the target signature template between chromatic zones is tried to achieve, and g () is a kernel function, and r is used for the factor of normalization kernel function envelope;

Then, upgrade the sub-state in position of reference state:

p ^ref＝p ^ref+M _r，g(x ^ref)

Calculate the zeroth order square density of reference state:

$m_{00}\left(x^{\mathrm{ref}}\right)=\frac{M_{00}\left(x^{\mathrm{ref}}\right)}{256\×l^{\mathrm{ref}}\×h^{\mathrm{ref}}}$

Wherein: M ₀₀(x ^Ref) be that target appears at state x ^RefZeroth order square in the pairing rectangular area, upgrade zeroth order square semi-invariant by zeroth order square density:

m ₀₀(x ^ref，n)＝m ₀₀(x ^ref，n-1)×m ₀₀(x ^ref)

4. average drifting vector, zeroth order square density, the zeroth order square semi-invariant of i sampled point are directly carried out assignment respectively by the corresponding information of i-1 sample point;

5. obtain final zeroth order square semi-invariant

And the new state of process average drifting optimization and dimension self-adaption adjustment:

${\hat{x}}_t^i={(p^{\mathrm{ref}},k^{\mathrm{iI}}\·\sqrt{{\stackrel{\·}{m}}_{00}^i}\·{\stackrel{\·}{s}}_t^i)}^T;$

6. store the final position of average drifting

With final zeroth order square semi-invariant

Step 4 is set up the importance sampling density function to each sampling point, combines with average drifting method the step 3 from the probability angle of Monte Carlo sampling, obtains the update mode x of sample _t ⁱ, and at state x _t ⁱFollowing refreshing weight w _t ⁱ, then by sampled point is gathered { x _t ⁱ, w _t ⁱ} _I=1...NResample, obtain the discrete estimation { x of the posterior probability distribution of current time target end-state _t ⁱ, w _t ⁱ} _I=1...N

Described each sampled point is set up the importance sampling density function, is specially:

4.1) particle assembly is divided into two groups randomly, wherein first group of ratio that accounts for sum is α, second group of ratio that accounts for sum is 1-α;

4.2) state of particle shifts according to second-order autoregressive model in first group, promptly on particle current state basis, add state transitions amount and a random quantity of deferring to Gaussian distribution of previous moment, this random quantity of deferring to Gaussian distribution has identical dimension with particle state, average is zero, and the variance ∑ _DProvide according to experience, or study obtains according to known sample;

4.3) when second group particle state is sampled, particle wherein is divided into A more at random with equiprobability, B two groups, the particle sampler process of every group is deferred to a kind of Gaussian distribution respectively and is carried out, but the location components at the Gaussian distribution center of two groups is the convergence position of particle through the average drifting method, and wherein: the scale component at A group Gaussian distribution center is to choose yardstick amplification factor k ₁The time self-adaptation two time scales approach the optimization yardstick, variance is amplified the second-order autoregressive model covariance matrix of correction for the stack yardstick

Add the probabilistic covariance matrix ∑ of simulation self-adaptation yardstick average drifting method _CAMThe scale component at B group Gaussian distribution center is then dwindled factor k for choosing yardstick ₂The time self-adaptation two time scales approach the optimization yardstick, variance is dwindled the second-order autoregressive model covariance matrix of correction for the stack yardstick

Add the probabilistic covariance matrix ∑ of simulation self-adaptation yardstick average drifting method _CAM, wherein, I is the maximum convergent iterations number of times of average drifting,

It is the final zeroth order square semi-invariant of i sampled point.

Described importance sampling density function, specific as follows:

$Q\left({\stackrel{\‾}{x}}_t^i\right|x_{0:t-1}^i,z_t)=\mathrm{\α}\·N({\stackrel{\‾}{x}}_t^i;2x_{t-1}^i-x_{t-2}^i,{\mathrm{\Σ}}_D)+(1-\mathrm{\α})\·$

$\{\frac{1}{2}N({\stackrel{\‾}{x}}_t^i;{({\stackrel{\·}{p}}^i,k_1^I\sqrt{{\stackrel{\·}{m}}_{00}^i}\·(2s_{t-1}^i-s_{t-2}^i))}^T,\mathrm{diag}(\mathrm{0,0},k_1^{2I}{\stackrel{\·}{m}}_{00}^i,k_1^{2I}{\stackrel{\·}{m}}_{00}^i)\×{\mathrm{\Σ}}_D+{\mathrm{\Σ}}_{\mathrm{CAM}})---\left(7\right)$

$+\frac{1}{2}N({\stackrel{\‾}{x}}_t^i;{({\stackrel{\·}{p}}^i,k_2^I\sqrt{{\stackrel{\·}{m}}_{00}^i}\·(2s_{t-1}^i-s_{t-2}^i))}^T,\mathrm{diag}(\mathrm{0,0}{k}_2^{2I}{\stackrel{\·}{m}}_{00}^i,k_2^{2I}{\stackrel{\·}{m}}_{00}^i)\×{\mathrm{\Σ}}_D+{\mathrm{\Σ}}_{\mathrm{CAM}})\}$

Wherein, the physical significance of each parameter is: i={1, and 2 ..., N}, N represent the number of particle sampler point; x _t ⁱThe state vector of i particle is only considered the sub-state p in position under the expression moment t ⁱWith the sub-state s of yardstick _t ⁱSituation under,

x _t ⁱExpression x _t ⁱIntermediateness before resampling; z _tMeasurement vector under the expression moment t; The intermediateness of i particle position after average drifting optimization is And

Represent the final zeroth order square semi-invariant under this intermediateness; Parameter I is the maximum convergent iterations number of times of embedded dimension self-adaption average drifting; α is the probability that the sampling particle state is optimized by average drifting; N () represents gauss of distribution function; ∑ _DWith ∑ _CAMThe variance of representing the state transition model predicated error that prediction brings to dbjective state of sequential Monte Carlo filtering second order autoregression state transition model and dimension self-adaption average drifting respectively; The error that embedded dimension self-adaption average drifting method produces when estimating the sub-state of yardstick is adjusted factor k by two kinds of yardsticks ₁And k ₂Carry out the best selective compensation; T representing matrix transposition symbol.

Described at state x _t ⁱFollowing refreshing weight w _Ti, be specially:

${\stackrel{\‾}{w}}_t^i=w_{t-1}^i\frac{p\left(z_t\right|{\stackrel{\‾}{x}}_t^i)p\left({\stackrel{\‾}{x}}_t^i\right|x_{0:t-1}^i)}{Q\left({\stackrel{\‾}{x}}_t^i\right|x_{0:t-1}^i,z_t)}---\left(8\right)$

Wherein, p (z _t| x _t ⁱ) be at hypothesis state x _t ⁱThe observation model probability of following target can adopt between color histogram and the target signature template matching degree between chromatic zones to try to achieve; P (x _t ⁱ| x _0:t-1 ^t) be the dynamic model probability of target, adopt second-order autoregressive model here; Q (x _t ⁱ| x _0:t-1 ⁱ, z _t) be the importance sampling probability.

Described to sampled point set { x _t ⁱ, w _t ⁱ} _I=1...NResample, be specially: choose the big sampled point of weights in the sample set, and the weights of all sampling points of normalization, obtain the discrete estimation { x that the posterior probability of current time target end-state distributes _t ⁱ, w _t ⁱ} _I=1...N

Principle of the present invention is, sequential Monte Carlo filtering time complexity is higher, therefore in the importance sampling function, introduce the optimization of dimension self-adaption average drifting on target location and yardstick,, make the approximate more accurate of target posterior probability to improve sampling efficiency.For further reducing complexity, when calculating the intermediateness space of sample set, the average drifting parameter of sample has been done the locally consistent processing, think that promptly the interior sample of a neighborhood can converge to same local optimum point after average drifting optimization.Wherein, sample has avoided all samples to concentrate on local mode maximum point place with the selected average drifting of doing of certain probability, has guaranteed the diversity of sample state.Simultaneously, adopt two kinds of different yardsticks to adjust the factor at random to sample, respectively the goal hypothesis zone is dwindled and amplify, weights calculating and resampling mechanism by the filtering of sequential Monte Carlo are automatically chosen best scale, have solved the not accurate enough and pervasive shortcoming of single dimension self-adaption average drifting method.

Compared with prior art, the present invention proposes the unified importance sampling function of design, propose complete average drifting associated methods from the probability angle of Monte Carlo sampling, utilized neighborhood states uniform convergence characteristic to reduce the computation complexity that average drifting is optimized; Only the part sample is carried out average drifting, avoided the deficient problem of sample; Choosing of the multiple dimensioned adjustment factor overcome the not accurate enough and pervasive problem of yardstick estimation in the state convergence process.At basic method for tracking target---filtering of sequential Monte Carlo and average drifting, under identical experiment condition, the present invention is the scale size of estimating target more accurately, and more tracking difficult problems such as target occlusion, the operation time of reducing tracking are effectively handled in robust ground.

Description of drawings

Fig. 1 is the workflow diagram of tracking of the present invention;

Fig. 2 is that the sequential Monte Carlo filtering method of embedding average drifting common in the prior art is adjusted the tracking effect figure of the sub-state of yardstick in change self-adaptation down of target scale;

Fig. 3 is the tracking effect figure that prior art mesoscale self-adaptation average drifting method self-adaptation under target scale changes is adjusted the sub-state of yardstick;

Fig. 4 is the tracking effect figure that the present invention's self-adaptation under target scale changes is adjusted the sub-state of yardstick;

Fig. 5 is the embodiment of the invention and embeds the comparing result of average drifting sequential Monte Carlo filter tracking method on computing time.

Embodiment

Below in conjunction with accompanying drawing embodiments of the invention are elaborated: present embodiment has provided detailed embodiment and process being to implement under the prerequisite with the technical solution of the present invention, but protection scope of the present invention is not limited to following embodiment.

In the present embodiment certain " football player " video sequence is carried out target following.Design parameter is set to: total number of particles N=50; Yardstick is adjusted factor k ₁=1.2, k ₂=0.9; For embedded average drifting optimization, its iterations I=2, probability of happening α=0.5, the uniform convergence neighborhood is defined as

The error variance that state optimization produces is about ∑ _CAM=diag{0.25,0.25,2 * 10 ^-4, 2 * 10 ^-4, the filtering of sequential Monte Carlo is about ∑ to the error variance that state estimation produces _D=diag{1.0,0.25, (5,4; 4,5) * 10 ^-4.

As shown in Figure 1, present embodiment comprises the steps:

Step 1, initialization particle sample state: at first frame of sequence, determine centroid position, the length and wide of target, thereby set up the initial state vector x of target by the automatic target detection method ₀, and the initial state vector of all sampled points all is made as x ₀, and compose with identical weights Simultaneously, calculate the measurement vector of target, as color histogram, as the reference template of target signature;

Step 2 produces two kinds of different scale factors at random to each sampled point, calculates sample second order autoregression central point, and stores as the central point of average drifting; Since second frame, recursively ask for the posterior probability sample distribution of present frame target with the measurement vector of the distributions of last frame sampling point and current frame image, specific as follows:

1. 1～50 sampled point is obtained separately second order autoregression center according to formula (1), these data all leave in the array as the central point of average drifting;

2. each sampling point i=1:N is got equally distributed stochastic variable u～U[0,1), if

The scale factor k of this sampling point ⁱBe taken as k ₁=1.2, otherwise be taken as k ₂=0.9, they have characterized the variation characteristic that the interframe target scale amplifies and dwindles respectively;

Step 3 according to the second order autoregression center of the sampled point in the sample set, obtains having the conforming average drifting of neighborhood field, and is specific as follows:

1. the average drifting vector of this dotted state, the sub-state of sampling point position, zeroth order square density, the zeroth order square semi-invariant after the drift are calculated successively according to formula (2)-(5) in the second order autoregression center of the 1st sampled point, so iterative computation is 2 times;

2. obtain the new state of the 1st sampled point according to formula (6) through average drifting and dimension self-adaption adjustment, and the final position and the zeroth order square semi-invariant of storing first sampled point under the new state;

3. to i sampled point (i={2,3 ... N}), if the Euclidean distance between the sub-state in position of sub-state in the position in its state vector and previous sampled point, carried out for the 4. step so greater than 2; Otherwise jumped to for the 5. step;

4. average drifting vector, zeroth order square density, the zeroth order square semi-invariant at this some place are calculated successively according to formula (2)-(5) in the second order autoregression center of i sampled point, so iterative computation is 2 times;

5. the average drifting vector of i sampled point, zeroth order square density, zeroth order square semi-invariant are directly carried out assignment respectively by the corresponding information of i-1 sample point;

6. obtain the new state of i sampled point according to formula (6) through average drifting and dimension self-adaption adjustment, and the final position and the zeroth order square semi-invariant of storing i sampled point under the new state.

7. repeated for the 3.-4. step, all processed until 50 sampled points.

Step 4 is at first set up the importance sampling density function to each sampling point, combines with the average drifting method from the probability angle of Monte Carlo sampling, obtains the update mode x of sample _t ⁱ, and at state x _t ⁱFollowing refreshing weight w _t ⁱ, then by sampled point is gathered { x _t ⁱ, w _t ⁱ} _I=1...NResample, obtain the discrete estimation { x of the posterior probability distribution of current time target end-state _t ⁱ, w _t ⁱ} _I=1...N, specific as follows:

1. set up the importance sampling function, obtain new sampling point set { x _t ⁱ, i={1,2 ..., 50}, the importance sampling function is specific as follows:

$Q\left({\stackrel{\‾}{x}}_t^i\right|x_{0:t-1}^i,z_t)=\mathrm{\α}\·N({\stackrel{\‾}{x}}_t^i;2x_{t-1}^i-x_{t-2}^i,{\mathrm{\Σ}}_D)+(1-\mathrm{\α})\·$

$\{\frac{1}{2}N({\stackrel{\‾}{x}}_t^i;{({\stackrel{\·}{p}}^i,k_1^I\sqrt{{\stackrel{\·}{m}}_{00}^i}\·(2s_{t-1}^i-s_{t-2}^i))}^T,\mathrm{diag}(\mathrm{0,0},k_1^{2I}{\stackrel{\·}{m}}_{00}^i,k_1^{2I}{\stackrel{\·}{m}}_{00}^i)\×{\mathrm{\Σ}}_D+{\mathrm{\Σ}}_{\mathrm{CAM}})---\left(7\right)$

$+\frac{1}{2}N({\stackrel{\‾}{x}}_t^i;{({\stackrel{\·}{p}}^i,k_2^I\sqrt{{\stackrel{\·}{m}}_{00}^i}\·(2s_{t-1}^i-s_{t-2}^i))}^T,\mathrm{diag}(\mathrm{0,0},k_2^{2I}{\stackrel{\·}{m}}_{00}^i,k_2^{2I}{\stackrel{\·}{m}}_{00}^i)\×{\mathrm{\Σ}}_D+{\mathrm{\Σ}}_{\mathrm{CAM}})\}$

Wherein, the physical significance of each parameter is: i={1, and 2 ..., N}, N represent the number of particle sampler point; x _t ⁱThe state vector of i particle is only considered the sub-state p in position under the expression moment t ⁱWith the sub-state s of yardstick _t ⁱSituation under,

x _t ⁱExpression x _t ⁱIntermediateness before resampling; z _tMeasurement vector under the expression moment t; The intermediateness of i particle position after average drifting optimization is

And

Represent the zeroth order square semi-invariant under this intermediateness; Parameter I is the iterations of embedded dimension self-adaption average drifting; α is the probability that the sampling particle state is optimized by average drifting; N () represents gauss of distribution function; ∑ _DWith ∑ _CAMThe variance of representing the state transition model predicated error that prediction brings to dbjective state of sequential Monte Carlo filtering second order autoregression state transition model and dimension self-adaption average drifting respectively; The error that embedded dimension self-adaption average drifting method produces when estimating the sub-state of yardstick is adjusted factor k by two kinds of yardsticks ₁And k ₂Carry out the best selective compensation; T representing matrix transposition symbol.

2. use formula (8) to upgrade the weights of each new sampled point, obtain { w _t ⁱ} _I=1...N

3. the up-to-date weights to 50 sampled points carry out normalization, obtain { w _t ⁱ} _I=1...N

4. to i (i={1,2 ... N}) individual sampled point resamples, and promptly by duplicating or delete the number of sampled point in sample set that makes this state is

Expression rounds operation;

If 5. the total number of sample points in sample set this moment is duplicated that sampled point of weights maximum so less than 50, equal 50 until total number of sample points, obtain the end-state set { x of sample _t ⁱ} _I=1...N

Implementation result

According to above-mentioned steps, the cycle tests of a plurality of CIF forms (352 * 288 pixel), 25fps is carried out target following, the length of each sequence does not wait at the 200-2000 frame, and the target sizes scope of being followed the tracks of is a 100-1000 pixel.

Be illustrated in figure 4 as the tracking results that adopts certain " football player " sequence that present embodiment obtains, target indicates with square frame.Compare with dimension self-adaption average drifting (as shown in Figure 3) with the sequential Monte Carlo filtering (as shown in Figure 2) that embeds average drifting, the present embodiment method is the change in location of tracking target well, more exactly captured target yardstick process from large to small.Wherein, Fig. 2, the subgraph among Fig. 3 and Fig. 4 (a)-(f) have portrayed the position and the size variation of target according to time sequencing.

In analyzing the operation time of the present embodiment that provides as Fig. 5, respectively four video sequences are tested with present embodiment method (CAMSGPF curve) and existing sequential Monte Carlo filtering method (MSEPF curve) in conjunction with the dimension self-adaption average drifting.Be 0.138ms the averaging time that method of the present invention spends on each particle in figure (a) expression " offroad vehicle " sequence, than existing method fast 23.63%; Be 0.043ms the averaging time that method of the present invention spends on each particle in figure (b) expression " ice hockey " sequence, than existing method fast 30.03%; Be 0.038ms the averaging time that method of the present invention spends on each particle in figure (c) expression " highway car " sequence, than existing method fast 45.49%; Be 0.239ms the averaging time that method of the present invention spends on each particle in figure (d) expression " football player " sequence, than existing method fast 34.47%.As seen present embodiment method (CAMSGPF curve) can be at 40ms with the interior processor active task of finishing a frame, improves about 30%-40% than the speed of existing sequential Monte Carlo filtering method (MSEPF curve) in conjunction with the dimension self-adaption average drifting.

Claims (4)

1, a kind of method for tracking dimension self-adaption video target of low complex degree is characterized in that, comprises the steps:

Step 1, initialization particle sample state: by detecting the original state x that determines target ₀, calculate the measurement vector of target simultaneously, the number N of given particle sampler point, the initial state vector of all sampled points is x ₀, and compose with unified weights

Step 2, the particle sample of step 1 is got equally distributed stochastic variable u～U[0,1 to each sampled point i=1:N), if $u<\frac{1}{2},$ The scale factor k of this sampled point ⁱBe taken as k ₁, otherwise be taken as k ₂, wherein: k ₁=1.2, k ₂=0.9, characterized the variation characteristic that the interframe target scale amplifies and dwindles respectively, calculate sample second order autoregression central point, and store as the central point of average drifting;

Described calculating sample second order autoregression central point is specially:

Make the current sampling point set be { x _t ⁱ, w _t ⁱ} _{I=1 ... N}, each sampled point i=1:N is calculated its second order autoregression center

${\stackrel{\&CenterDot;}{x}}_t^i=2x_{t-1}^i-x_{t-2}^i$

In the formula, x _t ⁱThe state vector of i particle under the expression moment t;

Step 3 according to the second order autoregression center of the sampled point in the sample set in the step 2, obtains having the conforming average drifting of neighborhood field;

Step 4 is set up the importance sampling density function to each sampled point, combines with average drifting method the step 3 from the probability angle of Monte Carlo sampling, obtains the update mode x of sample _t ⁱ, and at state x _t ⁱFollowing refreshing weight w _t ⁱ, then by sampled point is gathered { x _t ⁱ, w _t ⁱ} _{I=1 ... N}Resample, obtain the discrete estimation { x of the posterior probability distribution of current time target end-state _t ⁱ, w _t ⁱ} _{I=1 ... N}

Described each sampled point is set up the importance sampling density function, is specially:

4.1) particle assembly is divided into two groups randomly, wherein first group of ratio that accounts for sum is α, second group of ratio that accounts for sum is 1-α;

4.2) state of particle shifts according to second-order autoregressive model in first group, promptly on particle current state basis, add state transitions amount and a random quantity of deferring to Gaussian distribution of previous moment, this random quantity of deferring to Gaussian distribution has identical dimension with particle state, average is zero, and the variance ∑ _DProvide according to experience, or study obtains according to known sample;

4.3) when second group particle state is sampled, particle wherein is divided into A more at random with equiprobability, B two groups, the particle sampler process of every group is deferred to a kind of Gaussian distribution respectively and is carried out, but the location components at the Gaussian distribution center of two groups is the convergence position of particle through the average drifting method, and wherein: the scale component at A group Gaussian distribution center is to choose yardstick amplification factor k ₁The time self-adaptation two time scales approach the optimization yardstick, variance is amplified the second-order autoregressive model covariance matrix of correction for the stack yardstick $\mathrm{diag}(\mathrm{0,0}{,k}_1^{2I}{\stackrel{\&CenterDot;}{m}}_{00}^i,k_1^{2I}{\stackrel{\&CenterDot;}{m}}_{00}^i)\&times;{\mathrm{\&Sigma;}}_D$ Add the probabilistic covariance matrix ∑ of simulation self-adaptation yardstick average drifting method _CAMThe scale component at B group Gaussian distribution center is then dwindled factor k for choosing yardstick ₂The time self-adaptation two time scales approach the optimization yardstick, variance is dwindled the second-order autoregressive model covariance matrix of correction for the stack yardstick $\mathrm{diag}(\mathrm{0,0}{,k}_2^{2I}{\stackrel{\&CenterDot;}{m}}_{00}^i,k_2^{2I}{\stackrel{\&CenterDot;}{m}}_{00}^i)\&times;{\mathrm{\&Sigma;}}_D$ Add the probabilistic covariance matrix ∑ of simulation self-adaptation yardstick average drifting method _CAM, wherein, I is the maximum convergent iterations number of times of average drifting,

It is the final zeroth order square semi-invariant of i sampled point.

2, the method for tracking dimension self-adaption video target of low complex degree according to claim 1, it is characterized in that, described second order autoregression center according to the sampled point in the sample set in the step 2 obtains having the conforming average drifting of neighborhood field, and is specific as follows:

1. at first judge each sampled point i=1:N, if i==1 or

Neighborhood Ω outside, then carry out successively 2., 3., 5. with step 6., otherwise carry out 4. successively, 5. with step 6.; Wherein,

Be respectively sampling point i and i-1 at t second order autoregression center constantly;

2. reference state $x^{\mathrm{ref}}\&equiv;(p^{\mathrm{ref}},s^{\mathrm{ref}})={\stackrel{\&CenterDot;}{x}}_t^i,$ P wherein ^Ref, s ^RefBe respectively the sub-state of position and yardstick, s ^Ref={ l ^Ref, h ^Ref, l ^Ref, h ^RefDescribed with p respectively ^RefLength and height for the target rectangle profile at center; Zero setting rank square semi-invariant m ₀₀(x ^Ref, n=0) initial value is 1, n=0:I;

3. n=1:I is circulated:

At first, computation of mean values drift vector M _{R, g}(x ^Ref):

$M_{r,g}\left(x^{\mathrm{ref}}\right)=\frac{\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{p}_{j}m\left(p_j\right)g\left({\left|\right|\frac{p^{\mathrm{ref}}-p_j}{r}\left|\right|}^2\right)}{\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}m\left(p_j\right)g\left({\left|\right|\frac{p^{\mathrm{ref}}-p_j}{r}\left|\right|}^2\right)}-p^{\mathrm{ref}}$

Wherein, { p _j} _{J=1 ... M}Be reference state x ^RefThe pixel coordinate that image rectangular area, place covers, m (p _j) be position p _jThe possibility weights, when measurement vector is color histogram, by x ^RefMatching degree between the color of object histogram of state correspondence and the target signature template between chromatic zones is tried to achieve, and g () is a kernel function, and r is used for the factor of normalization kernel function envelope;

Then, upgrade the sub-state in position of reference state:

p ^ref＝p ^ref+M _r，g(x ^ref)

Calculate the zeroth order square density of reference state:

$m_{00}\left(x^{\mathrm{ref}}\right)=\frac{M_{00}\left(x^{\mathrm{ref}}\right)}{256\&times;l^{\mathrm{ref}}\&times;h^{\mathrm{ref}}}$

Wherein: M ₀₀(x ^Ref) be that target appears at state x ^RefZeroth order square in the pairing rectangular area, upgrade zeroth order square semi-invariant by zeroth order square density:

m ₀₀(x ^ref，n)＝m ₀₀(x ^ref，n-1)×m ₀₀(x ^ref)

4. average drifting vector, zeroth order square density, the zeroth order square semi-invariant of i sampled point are directly carried out assignment respectively by the corresponding information of i-1 sample point;

5. obtain final zeroth order square semi-invariant ${\stackrel{\&CenterDot;}{m}}_{00}^i=m_{00}(x^{\mathrm{ref}},I)$ And the new state of process average drifting optimization and dimension self-adaption adjustment:

${\hat{x}}_t^i={(p^{\mathrm{ref}},k^{\mathrm{iI}}\&CenterDot;\sqrt{{\stackrel{\&CenterDot;}{m}}_{00}^i}\&CenterDot;{\stackrel{\&CenterDot;}{s}}_t^i)}^T;$

6. store the final position of average drifting ${\stackrel{\&CenterDot;}{p}}^i=p^{\mathrm{ref}}$ With final zeroth order square semi-invariant

3, the method for tracking dimension self-adaption video target of low complex degree according to claim 1 is characterized in that, and is described at state x _t ⁱFollowing refreshing weight w _t ⁱ, be specially:

${\stackrel{\&OverBar;}{w}}_t^i=w_{t-1}^i\frac{p\left(z_t\right|{\stackrel{\&OverBar;}{x}}_t^i)p\left({\stackrel{\&OverBar;}{x}}_t^i\right|x_{0:t-1}^i)}{Q\left({\stackrel{\&OverBar;}{x}}_t^i\right|x_{0:t-1}^i,z_t)}$

Wherein, p (z _t| x _t ⁱ) be at state x _t ⁱThe observation model probability of following target, the matching degree between employing color histogram and the target signature template between chromatic zones is tried to achieve; P (x _t ⁱ| x _0:t-1 ⁱ) be the dynamic model probability of target, adopt second-order autoregressive model here; Q (x _t ⁱ| x _0:t-1 ⁱ, z _t) be the importance sampling probability.

4, the method for tracking dimension self-adaption video target of low complex degree according to claim 1 is characterized in that, and is described to sampled point set { x _t ⁱ, w _t ⁱ} _{I=1 ... N}Resample, be specially: choose the big sampled point of weights in the sample set, and the weights of all sampling points of normalization, obtain the discrete estimation { x that the posterior probability of current time target end-state distributes _t ⁱ, w _t ⁱ} _{I=1 ... N}

Images