CN104036525A - Self-adaptive particle swarm optimization (PSO)-based particle filter tracking method - Google Patents

Abstract

The invention provides a self-adaptive particle swarm optimization (PSO)-based particle filter tracking method. The method comprises the following steps: S1, selecting a candidate target, and setting the frame of the selected candidate target as a first frame; S2, creating a reference color histogram template for the candidate target, and measuring the similarity between the color histogram of the candidate target and the reference color histogram template according to the bhattacharyya coefficients; S3, randomly distributing M particles in the periphery of the t-1 frame target in the t frame according to the Gauss, and adjusting the parameters of PSO algorithm according to the movement state, wherein t is more than 1; S4, searching the M particle according to the PSO algorithm, and then performing estimation calculation to obtain the position of the target. With the adoption of the self-adaptive PSO-based particle filter packing method, video targets can be tracked well in a complex conditions, including a moving target, a static target and moving-stopping-moving target and the like, and stable tracking of the complex movement targets above can also be realized well.

Description

A kind of adaptive particle swarm optimization particle filter tracking method

Technical field

The present invention relates to computer vision tracking technique field, particularly relate to the visual target tracking method field of adjusting population according to target state self-adaptation, be specially a kind of adaptive particle swarm optimization particle filter tracking method.

Background technology

At computer vision field, robust and precisely real-time vision follow the tracks of be one urgently to be resolved hurrily and have a challenging task.Aspect the tracking of concrete object, there had been many breakthrough progress, for example pedestrian, the tracking of face etc. in recent years.Zhang, the people such as Lu be published in the paper " Structure preserving object tracking " on " IEEE Conference on Computer Vision and Pattern Recognition ", mention for the accurate real-time follow-up of general object still extremely tired main cause for following some: 1) very few about the obtainable information of this target; 2) information obtaining has ambiguity, mixes the information of this background in the target information of acquisition; 3) may there is acute variation in the outward appearance of target.For the above reasons, the accurate real-time follow-up problem of general objectives has challenge all the time.

In recent years, particle filter method (PF) due to its can non-gaussian sum nonlinear problem feature be successful in vision tracking field.But it faces a serious problem, the samples impoverishment problem that himself suboptimum Sampling Machine system in importance sampling process causes.Particle filter method depends on importance sampling process and requires suggestion to distribute and can be similar to preferably posteriority distribution.The most frequently used method is sampled according to the probability model of state transitions (transfer prior probability) exactly.But in the time that observed quantity appears at the afterbody of prior probability or posteriority distribute more concentrated compared with prior distribution, this method is just extremely inapplicable.

In order to address this problem, people have proposed several method and have distributed to producing more accurate posteriority.For example, propose without mark particle filter side device (UPF), UPF is combined by Unscented kalman filtering device (UKF) and particle filter (PF).The PF part of UPF provides general probability framework to process non-Gauss's nonlinear system, and UKF part distributes by nearest observed quantity being taken into account to produce better suggestion.

Along with the rise of swarm intelligence algorithm, particle group optimizing thought (PSO) is because its robustness and dirigibility in complex dynamic environment is incorporated into visual target tracking field by increasing researcher.PSO algorithm and PF algorithm, all based on least unit-particle, are inspired by this, and many researchers are incorporated into PSO in PF, to have formed standard particle group and optimize particle filter algorithm (PSO-PF).But we find the PSO-PF algorithm of standard and are not suitable for tracking to have the target of compound movement, for example, " walk-stop-walk " class target.While " walking-stop-walk " target and when static and move, time and when fast and slow, time and in time of advancing and retreating, that is to say that its motion state is difficult to prediction.Standard P SO-PF is due to himself parameter immutableness, compound movement characteristic that can not fine this type of target of processing to such an extent as to tracking effect is not good.

Summary of the invention

The shortcoming of prior art in view of the above, the object of the present invention is to provide a kind of adaptive particle swarm optimization particle filter tracking method, changes the not good problem of tracking effect of complicated sensation target for solving prior art motion state.

For achieving the above object and other relevant objects, the invention provides a kind of adaptive particle swarm optimization particle filter tracking method, said method comprising the steps of: S1, choose candidate target, and the frame of setting selected candidate target is the 1st frame; S2, for candidate target is set up reference color histogram template, and adopts Pasteur's coefficient to measure the similarity between color histogram and the reference color histogram template of candidate target; S3, in t frame, presses M particle of gaussian random distribution in t-1 frame target location periphery, and adjusts the parameter of particle swarm optimization algorithm according to motion state, wherein, and t > 1; S4, searches for according to particle swarm optimization algorithm M particle, calculates target location by estimation.

Preferably, the color histogram of described candidate target is:

p _t(x)＝{p _t(n,x)},n＝1...N； $p_t(n,x)=K\×\underset{\mathrm{\μ}\∈R\left(x_t\right)}{\mathrm{\Σ}}\mathrm{\ω}\left(\right|\mathrm{\μ}-d\left|\right)\mathrm{\δ}[b_t\left(\mathrm{\μ}\right)-n];$ Wherein, the center that x is candidate target, p _t(x) color histogram of the candidate target that centered by, position is x, p _t(n, x) is p _t(x) n dimension value, 1≤n≤N, N is p _t(x) dimension, i.e. the interval sum of reference color histogram template; x _tfor candidate target is in the center of t frame, R (x _t) be order target area, ω (| μ-d|) be weighting function, δ [b _t(μ)-n] be dirichlet function, μ is R (x _t) the pixel coordinate in region, d is R (x _t) regional center point coordinate, b _t(μ) be the corresponding p of μ pixel coordinate _t(x) dimension index, K is for ensureing ${\mathrm{\Σ}}_{n=1}^N{p}_t(n,x_t)=1$ Normalization coefficient.

Preferably, the similarity that adopts Pasteur's coefficient to measure between color histogram and the reference color histogram template of candidate target is specially: wherein, D _b[p _t, q] and be p _twith Pasteur's distance of q, p _tfor the color histogram of candidate target, q is reference color histogram template, p _t(n, x) is p _t(x) n dimension value, the n dimension value that q (n) is q.

Preferably, be specially according to the parameter of motion state adjustment particle swarm optimization algorithm: v _t=x _t-x _t-1; a _t=v _t-v _t-1; wherein, v _tfor candidate target is in the speed of t frame, v _t-1for candidate target is in the speed of t-1 frame, x _tfor candidate target is at center, the x of t frame _t-1for candidate target is at center, a of t-1 frame _tfor candidate target is at acceleration, the d of t frame _tfor candidate target is at prediction distance, the w of t frame _tfor candidate target is at the inertial coefficient of t frame, h and l are respectively w _thigher limit and lower limit.

Preferably, particle swarm optimization algorithm is specially:

υ _i(t+1)＝w _t·υ _i(t)+c ₁·rand()·(p _i-x _i(t)+c ₂·rand()·(p _g-x _i(t))；

X _i(t+1)=x _i(t)+υ _i(t+1); Wherein, x _i(t), υ _i(t) be respectively center and the speed of i particle in the time of the t time iteration, x _i(t+1), υ _i(t+1) be respectively center and the speed of i particle in the time of the t+1 time iteration, w _tfor candidate target is at the inertial coefficient of t frame, p _ithe historical optimum position of i particle, p _gthe global history optimum position of all particles, the numbering that g is overall best particle, w _tparameter is inertial coefficient, c ₁and c ₂for coefficient, rand () is the random function that is uniformly distributed in [0,1] interval; The best particle of the overall situation is the particle of fitness maximum, and the fitness of particle is:

W ^(m)=w ^(m)exp (λ D _b[p _m, q]); Wherein, w ^(m)be the fitness of m particle, D _b[p _m, q] and be p _mwith Pasteur's distance of q, p _mbe the color histogram of m particle, q is reference color histogram template, and λ is the constant coefficient that regulates Pasteur's distance.

Preferably, calculating target location is specially: wherein, the weight of m particle while being t frame, the weight of m particle while being t-1 frame, while being t frame, color histogram, the q of m particle are reference color histogram template, and λ is the constant coefficient that regulates Pasteur's distance, x _tfor the candidate target that obtains according to the particle swarm optimization algorithm target location in t frame, the present position of m particle while being t frame, M is total number of particles, A is for ensureing normalization coefficient.

As mentioned above, a kind of adaptive particle swarm optimization particle filter tracking method of the present invention, has following beneficial effect:

The present invention can finely be adapted to the video frequency object tracking of complex condition, comprises moving target, static target, and " walk-stop-walk " class target etc., also can the tenacious tracking of fine realization to above-mentioned compound movement target.

Brief description of the drawings

Fig. 1 is shown as the schematic flow sheet of adaptive particle swarm optimization particle filter tracking method of the present invention.

Embodiment

Below, by specific instantiation explanation embodiments of the present invention, those skilled in the art can understand other advantages of the present invention and effect easily by the disclosed content of this instructions.The present invention can also be implemented or be applied by other different embodiment, and the every details in this instructions also can be based on different viewpoints and application, carries out various modifications or change not deviating under spirit of the present invention.

The object of the present invention is to provide a kind of adaptive particle swarm optimization particle filter tracking method, change the not good problem of tracking effect of complicated sensation target for solving prior art motion state.Below will elaborate a kind of adaptive particle swarm optimization particle filter tracking side's ratio juris of the present invention and embodiment, and make those skilled in the art not need creative work can understand a kind of adaptive particle swarm optimization particle filter tracking method of the present invention.

The present invention is providing a kind of adaptive particle swarm optimization particle filter tracking method, as shown in Figure 1, said method comprising the steps of:

Step S1, chooses candidate target, and the frame of setting selected candidate target is the 1st frame.

Step S2, for candidate target is set up reference color histogram template, and adopts Pasteur's coefficient to measure the similarity between color histogram and the reference color histogram template of candidate target.

Particularly, in the present embodiment, the color histogram of described candidate target is:

p _t(x)＝{p _t(n,x)},n＝1...N； $p_t(n,x)=K\×\underset{\mathrm{\μ}\∈R\left(x_t\right)}{\mathrm{\Σ}}\mathrm{\ω}\left(\right|\mathrm{\μ}-d\left|\right)\mathrm{\δ}[b_t\left(\mathrm{\μ}\right)-n];$ Wherein, the center that x is candidate target, p _t(x) color histogram of the candidate target that centered by, position is x, p _t(n, x) is p _t(x) n dimension value, 1≤n≤N, N is p _t(x) dimension, i.e. the interval sum of reference color histogram template; x _tfor candidate target is in the center of t frame, R (x _t) be order target area, ω (| μ-d|) be weighting function, δ [b _t(μ)-n] be dirichlet function, μ is R (x _t) the pixel coordinate in region, d is R (x _t) regional center point coordinate, b _t(μ) be the corresponding p of μ pixel coordinate _t(x) dimension index, K is for ensureing ${\mathrm{\Σ}}_{n=1}^N{p}_t(n,x_t)=1$ Normalization coefficient.

Particularly, in the present embodiment, the similarity that adopts Pasteur's coefficient to measure between color histogram and the reference color histogram template of candidate target is specially:

$D_b[p_t,q]={[1-\underset{n=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{p_t(n,x)q\left(n\right)}]}^{\frac{1}{2}};$

Wherein, D _b[p _t, q] and be p _twith Pasteur's distance of q, p _tfor the color histogram of candidate target, q is reference color histogram template, p _t(n, x) is p _t(x) n dimension value, the n dimension value that q (n) is q.

Step S3, in t frame, presses M particle of gaussian random distribution in t-1 frame target location periphery, and adjusts the parameter of particle swarm optimization algorithm according to motion state, wherein, and t > 1; S4, M particle searched for according to particle swarm optimization algorithm, calculates target location by estimation.

The xt place, target location that we obtain at previous frame, by Gaussian distribution N (x _t, Σ) stochastic distribution M search particle, wherein Σ is mean square deviation matrix, adjusts the parameter of particle swarm optimization algorithm afterwards according to motion state.

Particularly, in the present embodiment, the parameter of adjusting particle swarm optimization algorithm according to motion state is specially:

v _t＝x _t-x _t-1；

a _t＝v _t-v _t-1；

$d_t=v_t+\frac{1}{2}{a}_t;$

$w_t=(h-l^{\log \left(\right|d_t|+2)});$

Wherein, v _tfor candidate target is in the speed of t frame, v _t-1for candidate target is in the speed of t-1 frame, x _tfor candidate target is at center, the x of t frame _t-1for candidate target is at center, a of t-1 frame _tfor candidate target is at acceleration, the d of t frame _tfor candidate target is at prediction distance, the w of t frame _tfor candidate target is at the inertial coefficient of t frame, h and l are respectively w _thigher limit and lower limit.

Particularly, in the present embodiment, particle swarm optimization algorithm, namely M particle is specially according to particle swarm optimization algorithm search:

υ _i(t+1)＝w _t·υ _i(t)+c ₁·rand()·(p _i-x _i(t)+c ₂·rand()·(p _g-x _i(t))；

x _i(t+1)＝x _i(t)+υ _i(t+1)；

Wherein, x _i(t), υ _i(t) be respectively center and the speed of i particle in the time of the t time iteration, x _i(t+1), υ _i(t+1) be respectively center and the speed of i particle in the time of the t+1 time iteration, w _tfor candidate target is at the inertial coefficient of t frame, p _ithe historical optimum position of i particle, p _gthe global history optimum position of all particles, the numbering that g is overall best particle, w _tparameter is inertial coefficient, c ₁and c ₂for coefficient, rand () is the random function that is uniformly distributed in [0,1] interval.

Wherein, overall best particle is the particle of fitness maximum, and the fitness of particle is:

w ^(m)＝w ^(m)·exp(-λ·D _b[p _m,q])；

Wherein, w ^(m)be the fitness of m particle, D _b[p _m, q] and be p _mwith Pasteur's distance of q, p _mbe the color histogram of m particle, q is reference color histogram template, and λ is the constant coefficient that regulates Pasteur's distance.

Finally estimate to obtain target location, in step S3, calculate target location and be specially:

$w_t^{\left(m\right)}=w_{t-1}^{\left(m\right)}\·\exp (-\mathrm{\λ}\·D_b[p_t^{\left(m\right)},q\left]\right);$

$x_t=A\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{w}_t^{\left(m\right)}{x}_t^{\left(m\right)};$

Wherein, the weight of m particle while being t frame, the weight of m particle while being t-1 frame, while being t frame, color histogram, the q of m particle are reference color histogram template, and λ is the constant coefficient that regulates Pasteur's distance, x _tfor the candidate target that obtains according to the particle swarm optimization algorithm target location in t frame, the present position of m particle while being t frame, M is total number of particles, A is for ensureing normalization coefficient.

In order to illustrate that the present invention has superior performance, in this enforcement, also will compare with the particle group optimizing particle filter tracking method of standard.The present embodiment adopts following two kinds of indexs to carry out the quantitative evaluation of performance:

1) average central error principle, the center of definition tracking target and the square root of the position of manual mark.

2) success ratio, is defined as: wherein R _t, R _gbe respectively track algorithm and manually mark the target area obtaining.R _t∩ R _gfor R _tand R _gintersecting area.Area (R _t∩ R _g) expression R _tand R _gthe area of intersecting area.Score is larger, and what represent to follow the tracks of is more accurate, in the time that Score is greater than 0.7, thinks successful tracking target.In the present embodiment, the each parameter of initialization is: M=10, N=110, c ₁=c ₂=1.0, λ=20.

As shown in table 1, be shown as standard particle group and optimize particle filter method (PSO-PF) and the tracking center error result of adaptive particle swarm optimization particle filter method (APSO-PF) to ball video.W is the inertial coefficient of PSO-PF algorithm, and wherein w=1.0 is that parameter is the PSO-PF algorithm (standard particle group optimizes particle filter algorithm) of w=1.0, and APSO-PF represents algorithm of the present invention.

Table 1

Video sequence	w＝1.0	w＝0.8	w＝0.5	w＝0.3	w＝0.0	APSO-PF
							MountainBike	105.5	74.5	15.0	13.5	79.5	7.5
Woman	20.0	17.5	13.0	14.5	16.0	8.5
							Walking	25.5	18.5	15.0	18.5	20.5	12.5
Walking2	15.0	13.5	10.5	11.5	13.0	6.5
							ball	18.5	13.0	45.5	50.0	53.5	6.5

As shown in table 2, be shown as standard particle group and optimize particle filter method (PSO-PF) and the tracking success ratio result of adaptive particle swarm optimization particle filter method (APSO-PF) to ball video.W is the inertial coefficient of PSO-PF algorithm, and wherein w=1.0 is that parameter is the PSO-PF algorithm of w=1.0, and APSO-PF represents algorithm of the present invention.

Table 2

Video sequence

w＝1.0

w＝0.8

w＝0.5

w＝0.3

w＝0.0

APSO-PF

MountainBike

75.0％

80.0％

85.0％

83.0％

79.0％

95.0％

Woman

35.5％

56.5％

55.5％

38.5％

30.5％

72.5％

Walking

20.0％

26.0％

25.0％

20.0％

16.0％

45.0％

Walking2

45.0％

53.5％

60.5％

51.5％

33.0％

76.5％

ball

90.0％

89.0％

33.0％

33.0％

33.0％

96.0％

In sum, the present invention can finely be adapted to the video frequency object tracking of complex condition, comprises moving target, static target, and " walk-stop-walk " class target etc., also can the tenacious tracking of fine realization to above-mentioned compound movement target.So the present invention has effectively overcome various shortcoming of the prior art and tool high industrial utilization.

Above-described embodiment is illustrative principle of the present invention and effect thereof only, but not for limiting the present invention.Any person skilled in the art scholar all can, under spirit of the present invention and category, modify or change above-described embodiment.Therefore, such as in affiliated technical field, have and conventionally know that the knowledgeable, not departing from all equivalence modifications that complete under disclosed spirit and technological thought or changing, must be contained by claim of the present invention.

Claims (6)

1. an adaptive particle swarm optimization particle filter tracking method, is characterized in that, said method comprising the steps of:

S1, chooses candidate target, and the frame of setting selected candidate target is the 1st frame;

S2, for candidate target is set up reference color histogram template, and adopts Pasteur's coefficient to measure the similarity between color histogram and the reference color histogram template of candidate target;

S3, in t frame, presses M particle of gaussian random distribution in t-1 frame target location periphery, and adjusts the parameter of particle swarm optimization algorithm according to motion state, wherein, and t > 1;

S4, searches for according to particle swarm optimization algorithm M particle, calculates target location by estimation.

2. adaptive particle swarm optimization particle filter tracking method according to claim 1, is characterized in that, the color histogram of described candidate target is: p _t(x)={ p _t(n, x) }, n=1...N; $p_t(n,x)=K\×\underset{\mathrm{\μ}\∈R\left(x_t\right)}{\mathrm{\Σ}}\mathrm{\ω}\left(\right|\mathrm{\μ}-d\left|\right)\mathrm{\δ}[b_t\left(\mathrm{\μ}\right)-n];$

Wherein, the center that x is candidate target, p _t(x) color histogram of the candidate target that centered by, position is x, p _t(n, x) is p _t(x) n dimension value, 1≤n≤N, N is p _t(x) dimension, i.e. the interval sum of reference color histogram template; x _tfor candidate target is in the center of t frame, R (x _t) be order target area, ω (| μ-d|) be weighting function, δ [b _t(μ)-n] be dirichlet function, μ is R (x _t) the pixel coordinate in region, d is R (x _t) regional center point coordinate, b _t(μ) be the corresponding p of μ pixel coordinate _t(x) dimension index, K is for ensureing normalization coefficient.

3. adaptive particle swarm optimization particle filter tracking method according to claim 2, is characterized in that, the similarity that adopts Pasteur's coefficient to measure between color histogram and the reference color histogram template of candidate target is specially:

$D_b[p_t,q]={[1-\underset{n=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{p_t(n,x)q\left(n\right)}]}^{\frac{1}{2}};$

Wherein, D _b[p _t, q] and be p _twith Pasteur's distance of q, p _tfor the color histogram of candidate target, q is reference color histogram template, p _t(n, x) is p _t(x) n dimension value, the n dimension value that q (n) is q.

4. adaptive particle swarm optimization particle filter tracking method according to claim 1, is characterized in that, the parameter of adjusting particle swarm optimization algorithm according to motion state is specially:

v _t＝x _t-x _t-1；

a _t＝v _t-v _t-1；

$d_t=v_t+\frac{1}{2}{a}_t;$

$w_t=(h-l^{\log \left(\right|d_t|+2)});$

Wherein, v _tfor candidate target is in the speed of t frame, v _t-1for candidate target is in the speed of t-1 frame, x _tfor candidate target is at center, the x of t frame _t-1for candidate target is at center, a of t-1 frame _tfor candidate target is at acceleration, the d of t frame _tfor candidate target is at prediction distance, the w of t frame _tfor candidate target is at the inertial coefficient of t frame, h and l are respectively w _thigher limit and lower limit.

5. according to the adaptive particle swarm optimization particle filter tracking method described in claim 1 or 4, it is characterized in that, particle swarm optimization algorithm is specially:

υ _i(t+1)＝w _t·υ _i(t)+c ₁·rand()·(p _i-x _i(t)+c ₂·rand()·(p _g-x _i(t))；

x _i(t+1)＝x _i(t)+υ _i(t+1)；

Wherein, x _i(t), υ _i(t) be respectively center and the speed of i particle in the time of the t time iteration, x _i(t+1), υ _i(t+1) be respectively center and the speed of i particle in the time of the t+1 time iteration, w _tfor candidate target is at the inertial coefficient of t frame, p _ithe historical optimum position of i particle, p _gthe global history optimum position of all particles, the numbering that g is overall best particle, w _tparameter is inertial coefficient, c ₁and c ₂for coefficient, rand () is the random function that is uniformly distributed in [0,1] interval; The best particle of the overall situation is the particle of fitness maximum, and the fitness of particle is:

w ^(m)＝w ^(m)·exp(-λ·D _b[p _m,q])；

Wherein, w ^(m)be the fitness of m particle, D _b[p _m, q] and be p _mwith Pasteur's distance of q, p _mbe the color histogram of m particle, q is reference color histogram template, and λ is the constant coefficient that regulates Pasteur's distance.

6. adaptive particle swarm optimization particle filter tracking method according to claim 5, is characterized in that, calculates target location and is specially:

$w_t^{\left(m\right)}=w_{t-1}^{\left(m\right)}\·\exp (-\mathrm{\λ}\·D_b[p_t^{\left(m\right)},q\left]\right);$

$x_t=A\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{w}_t^{\left(m\right)}{x}_t^{\left(m\right)};$

Wherein, the weight of m particle while being t frame, the weight of m particle while being t-1 frame, while being t frame, color histogram, the q of m particle are reference color histogram template, and λ is the constant coefficient that regulates Pasteur's distance, x _tfor the candidate target that obtains according to the particle swarm optimization algorithm target location in t frame, the present position of m particle while being t frame, M is total number of particles, A is for ensureing normalization coefficient.