CN104036525B - A kind of adaptive particle swarm optimization particle filter tracking method - Google Patents

Abstract

The present invention provides a kind of adaptive particle swarm optimization particle filter tracking method, said method comprising the steps of: S1, chooses candidate target, and the frame of the candidate target selected by setting is as the 1st frame；S2, sets up reference color rectangular histogram template for candidate target, and uses the similarity between color histogram and the reference color rectangular histogram template of Pasteur's coefficient tolerance candidate target；S3,tIn frame, int‑1 frame target location periphery presses gaussian random distributionMIndividual particle, and the parameter of particle swarm optimization algorithm is adjusted according to kinestate, wherein,t>1；S4,MIndividual particle scans for according to particle swarm optimization algorithm, is calculated target location by estimation.The present invention can be adapted to the video frequency object tracking of complex condition very well, including moving target, static target, and " walk and stop walking " class target etc., it is also possible to realize the tenacious tracking to above-mentioned compound movement target very 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 according to target state self-adaptative adjustment The visual target tracking method field of population, is specially a kind of adaptive particle swarm optimization particle filter tracking method.

Background technology

It is a urgently to be resolved hurrily and great challenge at computer vision field, robust and the most real-time visual tracking The task of property.Many breakthrough progress, such as pedestrian, the tracking etc. of face was had in recent years in terms of the tracking of concrete object. Zhang, Lu et al. are published in " IEEE Conference on Computer Vision and Pattern Recognition " on paper " Structure preserving object tracking " in mention for general object The most tired main cause of accurate real-time tracking be following some: 1) very few about the obtainable information of this target；2) The information obtained has ambiguity, it is thus achieved that target information in mix the information of this background；3) target outward appearance it may happen that Acute variation.For the above reasons, the accurate real-time tracking problem of general objectives is the most extremely challenging.

In recent years, particle filter method (PF) can the feature of non-gaussian and nonlinear problem be led at visual tracking due to it Territory is successful.But it faces a serious problem, i.e. himself suboptimum during importance sampling The samples impoverishment problem that Sampling Machine system causes.Particle filter method depends on importance sampling process and requires that suggestion is distributed energy Enough preferably approximate Posterior distrbutionp.Most common method is carried out according to the probabilistic model of state transfer (transfer prior probability) exactly Sampling.But when observed quantity occurs in the afterbody of prior probability or Posterior distrbutionp is if relatively prior distribution is more concentrated, this Method is the most inapplicable.

In order to solve this problem, there has been proposed several method to producing the most accurate Posterior distrbutionp.Such as, carry Going out without mark particle filter side device (UPF), UPF is to be combined by Unscented kalman filtering device (UKF) and particle filter (PF). The PF part of UPF provides general probability framework to process non-gaussian nonlinear system, and UKF part is by by nearest observed quantity Take into account and preferably advise distribution to produce.

Along with the rise of swarm intelligence algorithm, particle group optimizing thought (PSO) is due to its robust in complex dynamic environment Property and motility are incorporated into visual target tracking field by increasing researcher.PSO algorithm and PF algorithm are based on minimum Unit-particle, is inspired by this, and PSO is incorporated in PF and defines standard particle group's optimized particle filter algorithm by many researcheres (PSO-PF).However we have found that the PSO-PF algorithm of standard is not appropriate for following the tracks of the target with compound movement, such as " walk- Stop-walk " class target.During " walk-stop-walk " target static time and move, time and fast time and slow, time and the time of advance and retreat, also That is its kinestate is difficult to predict.Standard PSO-PF is due to himself parameter immutableness, it is impossible to process this classification very well Target compound movement characteristic to such an extent as to tracking effect is the best.

Summary of the invention

The shortcoming of prior art in view of the above, it is an object of the invention to provide a kind of adaptive particle swarm optimization grain Sub-filter tracking method, for solving the best the asking of tracking effect of the sensation target that kinestate change is complicated in prior art Topic.

For achieving the above object and other relevant purposes, the present invention provide a kind of adaptive particle swarm optimization particle filter with Track method, said method comprising the steps of: S1, chooses candidate target, and the frame of the candidate target selected by setting is as the 1st Frame；S2, sets up reference color rectangular histogram template for candidate target, and uses the color histogram of Pasteur's coefficient tolerance candidate target And the similarity between reference color rectangular histogram template；S3, in t frame, in t-1 frame target location, periphery is divided by gaussian random M particle of cloth, and the parameter of particle swarm optimization algorithm, wherein, t ＞ 1 is adjusted according to kinestate；S4, to M particle according to grain Subgroup optimized algorithm scans for, and is calculated target location by estimation.

Preferably, the color histogram of described candidate target is:

Wherein, x is candidate's mesh Target center, p_tX centered by (), position is the color histogram of the candidate target of x, p_t(n x) is p_tN-th dimension of (x) Value, 1≤n≤N, N are p_tThe interval sum of the dimension of (x), i.e. reference color rectangular histogram template；x_tFor candidate target at t frame Center, R (x_t) it is mesh target area, ω (μ-d) is weighting function, δ [b_t(μ)-n] it is dirichlet function, μ is R (x_t) The pixel coordinate in region, d is R (x_t) regional center point coordinates, b_t(μ) it is the p corresponding to μ pixel coordinate_tThe dimension of (x) Index, K is for ensureingNormalization coefficient.

Preferably, use between color histogram and the reference color rectangular histogram template of Pasteur's coefficient tolerance candidate target Similarity particularly as follows:Wherein, D_b[p_t, q] and it is p_tWith Pasteur's distance of q, p_tFor waiting Selecting the color histogram of target, q is reference color rectangular histogram template, p_t(n x) is p_tX n-th dimension value of (), q (n) is the of q N dimension value.

Preferably, the parameter of particle swarm optimization algorithm is adjusted particularly as follows: v according to kinestate_t=x_t-x_t-1；a_t=v_t- v_t-1；Wherein, v_tFor candidate target at the speed of t frame, v_t-1Exist for candidate target The speed of t-1 frame, x_tFor candidate target at the center of t frame, x_t-1For candidate target in the center of t-1 frame, a_tFor candidate target at the acceleration of t frame, d_tFor candidate target at the Prediction distance of t frame, w_tFor candidate target at t frame Inertia coeffeicent, h and l is respectively w_tHigher limit and lower limit.

Preferably, particle swarm optimization algorithm particularly as follows:

υ_i(t+1)=w_t·υ_i(t)+c₁·rand()·(p_i-x_i(t))+c₂·rand()·(p_g-x_i(t))；

x_i(t+1)=x_i(t)+υ_i(t+1)；Wherein, x_i(t)、υ_iT () is during i-th particle is when the t time iteration respectively Heart position and speed, x_i(t+1)、υ_i(t+1) it is the i-th particle center when the t+1 time iteration and speed, w respectively_t For candidate target at the inertia coeffeicent of t frame, p_iIt is the history optimum position of i-th particle, p_gIt it is the global history of all particles Optimum position, g is the numbering of the optimal particle of the overall situation, w_tParameter is inertia coeffeicent, c₁And c₂For coefficient, rand () is that [0,1] is interval Be uniformly distributed random function；The particle that the optimal i.e. fitness of particle of the overall situation is maximum, the fitness of particle is:

w^(m)=w^(m)·exp(-λ·D_b[p_m,q])；Wherein, w^(m)For the fitness of m-th particle, D_b[p_m, q] and it is p_mAnd q Pasteur's distance, p_mFor the color histogram of m-th particle, q is reference color rectangular histogram template, and λ is regulation Pasteur's distance Constant coefficient.

Preferably, be calculated target location particularly as follows: Wherein,The weight of m-th particle when being t frame,The power of m-th particle when being t-1 frame Weight,When being t frame, the color histogram of m-th particle, q are reference color rectangular histogram template, and λ is regulation Pasteur's distance Constant coefficient, x_tFor the candidate target that obtains according to the particle swarm optimization algorithm target location in t frame,When being t frame The present position of m particle, M is total number of particles, and A is for ensureingNormalization coefficient.

As it has been described above, a kind of adaptive particle swarm optimization particle filter tracking method of the present invention, there is following useful effect Really:

The present invention can be adapted to the video frequency object tracking of complex condition very well, including moving target, static target, with And " walk-stop-walk " class target etc., it is also possible to realize the tenacious tracking to above-mentioned compound movement target very well.

Accompanying drawing explanation

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

Detailed description of the invention

Below by way of specific instantiation, embodiments of the present invention being described, those skilled in the art can be by this specification Disclosed content understands other advantages and effect of the present invention easily.The present invention can also be by the most different concrete realities The mode of executing is carried out or applies, the every details in this specification can also based on different viewpoints and application, without departing from Various modification or change is carried out under the spirit of the present invention.

It is an object of the invention to provide a kind of adaptive particle swarm optimization particle filter tracking method, be used for solving existing The problem that in technology, the tracking effect of the sensation target that kinestate change is complicated is the best.The one of the present invention described in detail below Plant principle and the embodiment of adaptive particle swarm optimization particle filter tracking method, make those skilled in the art need not create Property work be i.e. appreciated that 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 it is shown in figure 1, described method bag Include following steps:

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

Step S2, sets up reference color rectangular histogram template for candidate target, and uses Pasteur's coefficient tolerance candidate target Similarity between color histogram and reference color rectangular histogram template.

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

Wherein, x is candidate's mesh Target center, p_tX centered by (), position is the color histogram of the candidate target of x, p_t(n x) is p_tN-th dimension of (x) Value, 1≤n≤N, N are p_tThe interval sum of the dimension of (x), i.e. reference color rectangular histogram template；x_tFor candidate target at t frame Center, R (x_t) it is mesh target area, ω (| μ-d |) is weighting function, δ [b_t(μ)-n] it is dirichlet function, μ is R (x_t) the pixel coordinate in region, d is R (x_t) regional center point coordinates, b_t(μ) it is the p corresponding to μ pixel coordinate_t(x) Dimension indexes, and K is for ensureingNormalization coefficient.

Specifically, in the present embodiment, the color histogram using Pasteur's coefficient tolerance candidate target is straight with reference color Side figure template between similarity particularly as follows:

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

Wherein, D_b[p_t, q] and it is p_tWith Pasteur's distance of q, p_tFor the color histogram of candidate target, q is that reference color is straight Side's figure template, p_t(n x) is p_tX n-th dimension value of (), q (n) is n-th dimension value of q.

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

We are at the target location x that previous frame obtains_tPlace, by Gauss distribution N (x_t, ∑) and random distribution M search grain Son, wherein ∑ is mean square deviation matrix, adjusts the parameter of particle swarm optimization algorithm afterwards according to kinestate.

Specifically, in the present embodiment, according to kinestate adjust particle swarm optimization algorithm parameter particularly as follows:

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(dt+2))；

Wherein, v_tFor candidate target at the speed of t frame, v_t-1For candidate target at the speed of t-1 frame, x_tFor candidate's mesh It is marked on the center of t frame, x_t-1For candidate target at the center of t-1 frame, a_tFor candidate target adding at t frame Speed, d_tFor candidate target at the Prediction distance of t frame, w_tFor candidate target at the inertia coeffeicent of t frame, h and l is respectively w_t Higher limit and lower limit.

Specifically, in the present embodiment, particle swarm optimization algorithm, namely M particle enters according to particle swarm optimization algorithm Line search particularly as follows:

υ_i(t+1)=w_t·υ_i(t)+c₁·rand()·(p_i-x_i(t))+c₂·rand()·(p_g-x_i(t))；

x_i(t+1)=x_i(t)+υ_i(t+1)；

Wherein, x_i(t)、υ_iT () is the i-th particle center when the t time iteration and speed, x respectively_i(t+1)、 υ_i(t+1) it is the i-th particle center when the t+1 time iteration and speed, w respectively_tFor candidate target being used at t frame Property coefficient, p_iIt is the history optimum position of i-th particle, p_gBeing the global history optimum position of all particles, g is that the overall situation is optimal The numbering of particle, w_tParameter is inertia coeffeicent, c₁And c₂For coefficient, rand () be [0,1] interval be uniformly distributed random function.

Wherein, the particle that the optimal i.e. fitness of particle of the overall situation is maximum, the fitness of particle is:

w^(m)=w^(m)·exp(-λ·D_b[p_m,q])；

Wherein, w^(m)For the fitness of m-th particle, D_b[p_m, q] and it is p_mWith Pasteur's distance of q, p_mFor m-th particle Color histogram, q is reference color rectangular histogram template, and λ is the constant coefficient of regulation Pasteur's distance.

Finally carry out estimation and obtain target location, in step s3, be calculated target location particularly as follows:

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

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

Wherein,The weight of m-th particle when being t frame,The weight of m-th particle when being t-1 frame,For During t frame, the color histogram of m-th particle, q are reference color rectangular histogram template, and λ is the constant coefficient of regulation Pasteur's distance, x_t For the candidate target that obtains according to the particle swarm optimization algorithm target location in t frame,M-th particle when being t frame Present position, M is total number of particles, and A is for ensureingNormalization coefficient.

In order to illustrate that the present invention has a superior performance, the most also by with the particle group optimizing particle filter of standard with Track method compares.The present embodiment uses following two index to carry out the quantitative assessment of performance:

1) average central error principle, definition follow the tracks of the center of target and the position of manually mark square Root.

2) success rate, is defined as:Wherein R_T、R_GIt is respectively track algorithm and manually marks The target area obtained.R_T∩R_GFor R_TAnd R_GIntersecting area.area(R_T∩R_G) represent R_TAnd R_GThe area of intersecting area. Score is the biggest, represents the most accurate of tracking, when Score is more than 0.7, then it is assumed that successfully track target.In the present embodiment, just The each parameter of beginningization is: M=10, N=110, c₁=c₂=1.0, λ=20.

As shown in table 1, standard particle group's optimized particle filter method (PSO-PF) and adaptive particle swarm optimization it are shown as The particle filter method (APSO-PF) the tracking center site error result to ball video.W is the inertial system of PSO-PF algorithm Number, wherein w=1.0 i.e. parameter is the PSO-PF algorithm (standard particle group's optimized particle filter algorithm) of w=1.0, APSO-PF generation Table inventive algorithm.

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, standard particle group's optimized particle filter method (PSO-PF) and adaptive particle swarm optimization it are shown as The particle filter method (APSO-PF) the tracking success rate result to ball video.W is the inertia coeffeicent of PSO-PF algorithm, wherein w =1.0 i.e. parameter is the PSO-PF algorithm of w=1.0, and APSO-PF represents inventive algorithm.

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 be adapted to the video frequency object tracking of complex condition very well, including moving target, quiet Only target, and " walk-stop-walk " class target etc., it is also possible to realize the tenacious tracking to above-mentioned compound movement target very well.Institute With, the present invention effectively overcomes various shortcoming of the prior art and has high industrial utilization.

The principle of above-described embodiment only illustrative present invention and effect thereof, not for limiting the present invention.Any ripe Above-described embodiment all can be modified under the spirit and the scope of the present invention or change by the personage knowing this technology.Cause This, have usually intellectual such as complete with institute under technological thought without departing from disclosed spirit in art All equivalences become are modified or change, and must be contained by the claim of the present invention.

Claims (5)

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

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

S2, sets up reference color rectangular histogram template for candidate target, and uses the color histogram of Pasteur's coefficient tolerance candidate target Similarity between figure and reference color rectangular histogram template；

S3, in t frame, presses gaussian random M particle of distribution in t-1 frame target location periphery, and adjusts according to kinestate The parameter of particle swarm optimization algorithm, wherein, t ＞ 1；According to kinestate adjust particle swarm optimization algorithm parameter particularly as follows:

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 at the speed of t frame, v_t-1For candidate target at the speed of t-1 frame, x_tExist for candidate target The center of t frame, x_t-1For candidate target at the center of t-1 frame, a_tFor candidate target t frame acceleration, d_tFor candidate target at the Prediction distance of t frame, w_tFor candidate target at the inertia coeffeicent of t frame, h and l is respectively w_tThe upper limit Value and lower limit；

S4, scans for according to particle swarm optimization algorithm M particle, is calculated target location by estimation.

Adaptive particle swarm optimization particle filter tracking method the most according to claim 1, it is characterised in that described candidate The color histogram of target is: p_t(x)={ p_t(n, x) }, n=1...N；

Wherein, x is the center of candidate target, p_tX centered by (), position is the color histogram of the candidate target of x, p_t(n, X) it is p_tX n-th dimension value of (), 1≤n≤N, N are p_tThe interval sum of the dimension of (x), i.e. reference color rectangular histogram template；x_t For candidate target in the center of t frame, R (x_t) it is mesh target area, ω (| μ-d |) is 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 coordinates, b_t(μ) it is that μ pixel is sat P corresponding to mark_tX the dimension index of (), K is for ensureingNormalization coefficient.

Adaptive particle swarm optimization particle filter tracking method the most according to claim 2, it is characterised in that use Pasteur Coefficient tolerance candidate target color histogram and reference color rectangular histogram template between similarity particularly as follows:

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

Wherein, D_b[p_t, q] and it is p_tWith Pasteur's distance of q, p_tFor the color histogram of candidate target, q is reference color Nogata artwork Plate, p_t(n x) is p_tX n-th dimension value of (), q (n) is n-th dimension value of q.

Adaptive particle swarm optimization particle filter tracking method the most according to claim 1, it is characterised in that population is excellent Change algorithm particularly as follows:

υ_i(t+1)=w_t·υ_i(t)+c₁·rand()·(p_i-x_i(t))+c₂·rand()·(p_g-x_i(t))；

x_i(t+1)=x_i(t)+υ_i(t+1)；

Wherein, x_i(t)、υ_iT () is the i-th particle center when the t time iteration and speed, x respectively_i(t+1)、υ_i(t+ 1) it is the i-th particle center when the t+1 time iteration and speed, w respectively_tFor candidate target in the inertial system of t frame Number, p_iIt is the history optimum position of i-th particle, p_gBeing the global history optimum position of all particles, g is the optimal particle of the overall situation Numbering, w_tParameter is inertia coeffeicent, c₁And c₂For coefficient, rand () be [0,1] interval be uniformly distributed random function；The overall situation The particle that the optimal i.e. fitness of particle is maximum, the fitness of particle is:

w^(m)=w^(m)·exp(-λ·D_b[p_m,q])；

Wherein, w^(m)For the fitness of m-th particle, D_b[p_m, q] and it is p_mWith Pasteur's distance of q, p_mColor for m-th particle is straight Fang Tu, q are reference color rectangular histogram template, and λ is the constant coefficient of regulation Pasteur's distance.

Adaptive particle swarm optimization particle filter tracking method the most according to claim 4, it is characterised in that estimate to calculate Obtain target location particularly as follows:

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

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

Wherein,The weight of m-th particle when being t frame,The weight of m-th particle when being t-1 frame,It is t During frame, the color histogram of m-th particle, q are reference color rectangular histogram template, and λ is the constant coefficient of regulation Pasteur's distance, x_tFor The candidate target obtained according to particle swarm optimization algorithm target location in t frame,The institute of m-th particle when being t frame Position, place, M is total number of particles, and A is for ensureingNormalization coefficient.