CN105913457B - Based on the particle filter method for tracking target for reinforcing particle group optimizing - Google Patents
Based on the particle filter method for tracking target for reinforcing particle group optimizing Download PDFInfo
- Publication number
- CN105913457B CN105913457B CN201610230000.XA CN201610230000A CN105913457B CN 105913457 B CN105913457 B CN 105913457B CN 201610230000 A CN201610230000 A CN 201610230000A CN 105913457 B CN105913457 B CN 105913457B
- Authority
- CN
- China
- Prior art keywords
- particle
- value
- target
- time
- optimized
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 239000002245 particle Substances 0.000 title claims abstract description 195
- 238000000034 method Methods 0.000 title claims abstract description 74
- 230000003014 reinforcing effect Effects 0.000 title abstract 3
- 238000005457 optimization Methods 0.000 claims abstract description 21
- 238000012952 Resampling Methods 0.000 claims abstract description 12
- 238000004364 calculation method Methods 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 238000012546 transfer Methods 0.000 claims description 5
- 230000003247 decreasing effect Effects 0.000 claims description 4
- 230000002708 enhancing effect Effects 0.000 claims description 4
- 230000017105 transposition Effects 0.000 claims description 3
- 238000004422 calculation algorithm Methods 0.000 abstract description 16
- 238000000605 extraction Methods 0.000 abstract description 3
- 238000012544 monitoring process Methods 0.000 abstract description 2
- 238000001914 filtration Methods 0.000 description 10
- 238000010586 diagram Methods 0.000 description 9
- 238000004088 simulation Methods 0.000 description 8
- 230000000694 effects Effects 0.000 description 4
- 230000001965 increasing effect Effects 0.000 description 4
- 238000011160 research Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 2
- 235000003642 hunger Nutrition 0.000 description 2
- 230000037351 starvation Effects 0.000 description 2
- 238000000342 Monte Carlo simulation Methods 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000019637 foraging behavior Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the particle filter method for tracking target for reinforcing particle group optimizing, mainly solves the problems, such as the particle degeneracy that particle filter video tracking algorithm occurs during tracking.Implementation step is：1. predicting to obtain candidate target by dbjective state under particle filter frame；2. by the local binary feature LBP statistic histogram of extraction candidate target as candidate target feature；3. obtaining corresponding weight by calculating the distance between candidate target feature and template；4. being optimized by reinforcing particle swarm optimization algorithm to prediction particle；5. the particle after pair optimization carries out resampling；6. the particle after resampling is merged to obtain the state estimation of target, the reliable tracking to target is realized.The present invention can effectively improve particle to the descriptive power of dbjective state, increase particle diversity, solve the problems, such as the particle degeneracy during particle filter, can be used for intelligent video monitoring, robot navigation, road traffic control system.
Description
Technical Field
The invention belongs to the technical field of target tracking, and particularly relates to a video target tracking method which can be used for intelligent video monitoring, robot navigation and road traffic control systems.
Background
In video tracking research, a target tracking method based on particle filtering is widely applied to video tracking due to the excellent characteristics of the target tracking method in solving the non-linearity problem and the compatibility of a relatively open tracking framework to various feature description methods. However, the particle filter algorithm has a particle starvation phenomenon in the tracking process, and once the particle starvation occurs, the diversity of the particles is greatly reduced, which greatly affects the tracking accuracy and the calculation efficiency of the tracking algorithm. Therefore, how to solve the problem of particle shortage in the particle filter algorithm has become a difficult point in video tracking research.
Particle swarm optimization PSO algorithm was proposed by Dr. Kennedy and Eberhart in 1995, particle swarm optimization was derived from simulation of a simple social model, inspired from bird foraging behavior model, and used to solve optimization problem. In particle swarm optimization, the potential solution to each optimization problem is a bird in the search space, called a "particle". All particles have an adaptation value determined by the optimisation function and each particle has a speed which determines their direction of transfer and displacement. The particles then search in space following the current optimal particle. The particle swarm optimization algorithm can find the global optimal solution of the problem with a larger probability, the calculation efficiency is higher than that of the traditional random method, once the particle swarm optimization algorithm is put forward, the particle swarm optimization algorithm immediately draws the wide attention of researchers in the field of evolution calculation, and a large number of research results including the reinforced particle swarm optimization algorithm ELPSO, the hybrid planning algorithm GAPSO and the like are shown in a short period of several years.
Currently, typical target tracking methods are: mean shift, ordinary particle filtering, etc., wherein:
although the mean shift method is simple and easy to implement and small in calculated amount, the mean shift method has local convergence characteristics and is easy to fall into a local optimal solution to generate obvious tracking hysteresis effect, and the target with a large dynamic range moving quickly is easy to lose track;
the tracking performance of the common particle filtering method depends on the number of used particles to a large extent, and after a plurality of iterations in the resampling process, all the particles tend to the same particle, so that the diversity of the particles is lost, the problem of particle shortage is serious, the estimation on the state of a target has large deviation, even the tracking is failed, meanwhile, the weights of most of the particles become very small, a large amount of operation time is wasted on the particles with small weights, and the calculation efficiency is reduced.
Disclosure of Invention
The invention aims to provide a particle filter target tracking method based on reinforced particle swarm optimization aiming at the defects in the prior art, so as to solve the problem of particle shortage of a particle filter algorithm in target tracking, thereby increasing the target tracking precision and improving the calculation efficiency.
The key technology for realizing the invention is as follows: the particle in the particle filter tracking algorithm is optimized by using the enhanced particle swarm optimization algorithm ELPSO, namely in a particle filter frame, the particle distribution in a state space is optimized by enhancing the particle swarm optimization algorithm ELPSO, so that the particle distribution tends to a high-likelihood region, the expression capability of the particles to a target state is improved, the diversity of the particles is increased, and the problem of particle shortage is solved. The method comprises the following implementation steps:
(1) initialization:
(1a) reading in an image I at time k-1k-1According to the initial state X of the object0Initializing the particle set at time k-1Wherein,the state interval of the ith particle at the time of k-1 is shown, i is the particle number and takes the values of 1,2, ·, N, N represents the total number of the particles, k represents the time, and the initial time k is 1;
(1b) initializing a target tracking window: b isk-1＝(rk-1,ck-1)TWherein r isk-1And ck-1Respectively representing the length and width values of a target tracking window at the k-1 moment, wherein T represents vector transposition operation;
(1c) initializing a local binary feature LBP statistical histogram M of a target as a feature template;
(2) target state prediction:
(2a) reading in an image I at time kkBy counting the particle sets at time k-1To obtain a predicted particle set at time kWhereinThe state prediction value of the ith particle at the moment k;
(2b) predicting a set of particles from time kAnd a target tracking window Bk-1Determining a candidate target set at time kWhereinIs the ith candidate target at the time k and is expressed byIs a center, Bk-1A rectangular area defined by the length and the width;
(3) extracting a set of candidate objectsFeature set corresponding to the candidate targetWherein ViRepresenting the ith candidate objectA corresponding feature;
(4) candidate target feature set is obtainedDistance set from feature template MAnd according to the distance setComputing a set of weights for candidate targetsWherein d isiRepresents the distance, ω, between the ith candidate target feature and the feature template MiRepresenting the weight of the ith candidate target;
(5) particle swarm optimization:
(5a) initializing a set of particles to be optimizedIndividual optimum value pid＝RiGroup optimum value pgd＝RmaxAnd the iteration time T is 1, and the maximum iteration time T is setmaxWhereinCorresponding set of velocities isThe set of weights isRmaxIs composed ofThe particle with the largest weight value corresponds to the weight value omegamax；
(5b) Judging whether the iteration number T satisfies T ≦ TmaxIf yes, carrying out the Tth iteration from the step (5c), otherwise, directly turning to the step (6);
(5c) according to individual optimum value pidAnd population optimum pgdSet of particles to be optimizedEach of the particles RiOptimized to obtain optimized particles QiAnd composing the optimized particle set
(5d) Updating optimized particle setsEach of which optimizes the particle QiCorresponding weight value omegaiAnd updating the individual optimum value pid；
(5e) Updating population optimal value pgd：
(5f) By enhancing the search, the population is given the optimal value pgdUpdating is carried out;
(5h) judging whether the iteration number T satisfies T > TmaxIf yes, finishing the particle swarm optimization, and if not, finishing the particle swarm optimizationLet T be T +1, let the set of particles to be optimizedWhich corresponds to a set of weights asAnd returning to the step (5 b);
(6) weight set for candidate targetNormalization is carried out, and the result after normalization is obtainedOptimizing particle sets for time kResampling is carried out to obtain a particle set at the k momentWhereinThe state interval of the ith particle at the k moment is shown;
(7) particle set according to time kEstimating a target state X at time kkDetermining a Target estimation range Target at the moment k;
(8) and (4) checking whether the information at the next moment arrives, if so, making k equal to k +1, and returning to the step (2), otherwise, ending the target tracking process.
According to the invention, the reinforced particle swarm optimization algorithm ELPSO is utilized to optimize the particles in the particle filtering process, so that the description capacity of the particles on the target state is enhanced, the diversity of the particles is increased, and the problem of particle shortage is solved; meanwhile, the description capability of the particles on the target state is enhanced, so that the target can be accurately described by fewer particles in tracking, and the calculation efficiency is improved.
Drawings
FIG. 1 is an overall flow diagram of the present invention;
FIG. 2 is a graph of the results of an experiment using the present invention and prior methods to track a pedestrian video sequence;
FIG. 3 is a graph of experimental results of tracking an infrared pedestrian video sequence using the present invention and prior methods;
fig. 4 is a graph of the results of an experiment using the present invention and prior art methods to track a toy video sequence.
Detailed Description
Referring to fig. 1, the implementation of the present invention includes the following steps:
step 1, initialization.
1.1) setting the initial time k to 1, according to the initial state X of the target0Initializing the particle set at time k-1
1.1.1) initial State X according to the target0Generating the ith particle at time k-1Wherein i is the serial number of the particle, the values are 1,2, N, N represents the total number of the particles,obey mean value of X0Gaussian distribution with variance of Ψ, X0Psi is the process noise variance for the initial state of the target;
1.1.2) forming a particle set by using the N particles obtained in the step 1.1.1)
1.2) initializing a target tracking window: b isk-1＝(rk-1,ck-1)TWherein r isk-1And ck-1Respectively representing the length and width values of a target tracking window at the k-1 moment, wherein T represents vector transposition;
1.3) initializing local binary feature LBP statistical histogram M of the target as a feature template.
And 2, predicting the target state.
2.1) reading in the image I at time kkBy counting the number of particles in the k-1 time imageObtaining a predicted particle set at time k
WhereinRepresenting the ith predicted particle at time k by counting the ith particle at time k-1The transfer is obtained by the following transfer formula:
wherein, wiIs state noise, obeys an average ofA gaussian distribution with variance Ψ;
2.2) prediction of the set of particles from the k timeAnd a target tracking window Bk-1Determining a candidate target set at time k
WhereinIndicating the ith predicted particle at time kThe corresponding candidate target is determined according to the following formula:
wherein,andrespectively representing the ith predicted particle at time kAbscissa and ordinate of (a), rk-1And ck-1Respectively represent target tracking windows B at the k-1 momentk-1Length and width values of (a).
And 3, extracting a candidate target feature set.
Common target features include gray scale features, color features, texture features, shape features, and the like.
The existing methods for extracting the target features comprise a color histogram method, a color moment method, a color correlation graph method and the like aiming at the color features; the extraction method aiming at the texture features comprises a local binary feature LBP statistical histogram method, a gradient method, a gray level co-occurrence matrix method, an autoregressive texture model method, a wavelet transformation method and the like; the shape feature extraction method includes a boundary feature method, a fourier shape descriptor method, a geometric parameter method, and the like.
The example uses, but is not limited to, a local binary feature LBP statistical histogram method in the existing method as a target feature, and includes the following steps:
3.1) computing a set of candidate objectsEach candidate target ofLocal binary characteristic LBP value of each pixel:
wherein, gcRepresents the center point pcGray value of gnRepresents the center point pcP represents p pixel points in the neighborhood around the central pixel point, n is the serial number of the p pixel points, s (g)n-gc) Is a function, represented in the form:
3.2) calculating a candidate target set according to the result obtained in the step 3.1)In each candidate targetLocal binary feature LBP statistical histogram Vi：
Wherein m represents the coding length of the local binary feature LBP, and m is 2pI (x, y) ═ I) is a function expressed in the form:
3.3) using the characteristics of the N candidate targets obtained in the step 3.2) to form a candidate target characteristic set
And 4, calculating a weight value set.
4.1) computing a candidate target feature setDistance set from feature template M
Wherein d isiRepresenting the distance between the ith candidate target feature and the feature template,ρiis the babbitt coefficient between the ith candidate target feature and the feature template,
4.2) set according to distanceComputing a set of candidate target weights
Wherein ω isiRepresenting the ith candidate objectThe calculation formula of the weight value of (2) is:
wherein d isiAnd (3) representing the distance between the ith candidate target feature and the feature template, wherein R is the feature observation noise variance.
And 5, optimizing the particle swarm.
5.1) initializing the set of particles to be optimizedIndividual optimum value pid＝RiGroup optimum value pgd＝RmaxSetting the maximum iteration number T as 1maxWhereinCorresponding set of velocities isThe set of weights isRmaxIs composed ofThe weight of the particle with the maximum medium weight is omegamax；
5.2) judging whether the iteration number T satisfies T < ═ TmaxIf yes, carrying out the Tth iteration from the step 5.3), otherwise, directly turning to the step 6;
5.3) according to the individual optimum value pidAnd population optimum pgdSet of particles to be optimizedEach of the particles RiOptimized to obtain optimized particles QiAnd composing the optimized particle set
5.3.1) updating the set of particles to be optimizedMesoparticle RiVelocity v ofi：
vi＝w·vi+c1·r1·(pid-Ri)+c2·r2·(pgd-Ri)，<14>
Wherein w is an inertia weight, c1For a learning factor related to the individual optimum, c2For a learning factor related to the global optimum, r1And r2Is two different random numbers between (0, 1);
5.3.2) velocity v obtained in step 5.3.1)iUpdating a set of particles to be optimizedParticle R in (1)iTo obtain optimized particles Qi：
Qi＝Ri+vi；<15>
5.3.3) forming an optimized particle set by using the N optimized particles obtained in the step 5.3.2)
Its corresponding weight value set is
5.4) updating the optimized particle setEach of which optimizes the particle QiCorresponding weight value omegaiAnd updating the individual optimum value pid：
5.4.1) calculation of optimized particles QiTemporary weight value omeganew：
Wherein d isnewRepresents the optimized particle QiThe distance between the corresponding candidate target feature and the feature template;
5.4.2) obtaining the temporary weight omega according to the step 5.4.1)newUpdating optimized particle QiCorresponding weight value omegaiAnd updating the individual optimum value pid：
If ω isnew＞ωiThen update the optimized particle QiCorresponding weight value omegai＝ωnewUpdating the individual optimum value pid＝QiOtherwise ω isiConstant value, pidThe value is unchanged;
5.5) updating the population optimal value pgd：
If particle Q is optimizediCorresponding weight value omegaiSatisfy omegai＞ωmaxThen update the population optimal value pgd＝QiIts corresponding weight value omegamax＝ωiElse pgdThe value is not changed, and the value corresponds to the weight omegamaxThe change is not changed;
5.6) optimal value p for the population by enhancing the searchgdUpdating:
5.6.1) optimal value p for the populationgdPerforming the first search to obtain a search result pg1：
pg1＝pgd+(Xmax-Xmin)·Gaussin(0,h)，<18>
Wherein, XmaxRepresenting the upper bound, X, of the decision vectorminRepresenting the lower bound of the decision vector, Gaussin (0, h) is a gaussian distribution with mean zero and standard deviation h, according to the rule h-1/TmaxPerforming linear decreasing;
5.6.2) population optimal valuepgdAnd (3) carrying out first updating:
if p isg1Weight ω of corresponding candidate targetg1Satisfy omegag1＞ωmaxThen update the population optimal value pgd＝pg1Which corresponds to the weight omegamax＝ωg1Else pgdThe value is not changed, and the value corresponds to the weight omegamaxThe change is not changed;
5.6.3) optimal value p for the populationgdPerforming the second search to obtain a second search result pg2：
pg2＝pgd+(Xmax-Xmin)·Cauchy(0,s)，<19>
Wherein Cauchy (0, s) is a Cauchy distribution with a peak parameter of zero and a scale parameter of s, s being s-1/T according to the lawmaxPerforming linear decreasing;
5.6.4) optimal value p for the populationgdAnd (5) carrying out second updating:
if p isg2Weight ω of corresponding candidate targetg2Satisfy omegag2＞ωmaxThen the population optimum value pgd＝pg2Which corresponds to the weight omegamax＝ωg2Else pgdThe value is not changed, and the value corresponds to the weight omegamaxThe change is not changed;
5.7) let T ═ T +1, let the set of particles to be optimizedWhich corresponds to a set of weights asAnd returns to step 5.2).
And 6, resampling.
The existing resampling method comprises polynomial resampling, system resampling, residual resampling and the like, but the example uses, but is not limited to, the system resampling method in the existing method to resample the optimized particle set, and the steps are as follows:
6.1) weight set for candidate targetsNormalization is carried out to obtain a normalized weight set
Wherein, omega'iIs omegaiThe result of the corresponding normalization is then,
6.2) obtaining the normalized weight set according to the step 6.1)Optimizing particle sets for time kResampling is carried out to obtain a particle set at the k momentWhereinIndicating the state interval of the ith particle at time k.
And 7, updating the target state.
7.1) particle set according to time kEstimating a target state X at time kk：
Wherein N represents the total number of particles,representing the state estimation value of the ith particle at the k moment;
7.2) target State X according to time kkAnd a target tracking window BkDetermining a Target estimation range Target at the time k:
wherein x iskAnd ykRespectively representing target states X at time kkAbscissa and ordinate of (a), rkAnd ckRespectively representing the length and width values of the target tracking window at the time k.
And 8, checking whether the information at the next moment arrives, if so, making k equal to k +1, and returning to the step 2, otherwise, ending the target tracking process.
The effect of the invention can be further illustrated by the following simulation experiment:
1. and (5) simulating conditions.
Simulation environment: the computer adopts Intel Pentium D CPU 2.8Ghz, 1GB memory, and the software adopts Matlab7.4 simulation experiment platform.
2. The simulation method comprises the following steps: the method of the invention, the existing mean shift method and the common particle filtering method.
3. Simulating the content and result.
Simulation 1: with the three methods, a pedestrian video sequence is tracked, and the result is shown in fig. 2, in which:
FIG. 2(a) is a diagram of the results of tracking frames 27, 39, 86 of a pedestrian video sequence using the present invention;
FIG. 2(b) is a diagram showing the result of tracking the 27 th, 39 th and 86 th frames of a pedestrian video sequence by using a mean shift method;
fig. 2(c) is a diagram of the result of tracking the 27 th, 39 th and 86 th frames of the pedestrian video sequence by using the ordinary particle filtering method.
As can be seen from fig. 2, the tracking effect of the present invention is better than that of methods 2 and 3 in the case of partial occlusion.
Simulation 2: the infrared pedestrian video sequence is tracked by the three methods, and the result is shown in fig. 3, wherein:
FIG. 3(a) is a diagram of the results of tracking frames 45, 57, 69, 81 of an infrared pedestrian video sequence using the present invention;
FIG. 3(b) is a graph showing the result of tracking 45 th, 57 th, 69 th and 81 th frames of an infrared pedestrian video sequence by a mean shift method;
fig. 3(c) is a diagram of the result of tracking 45 th, 57 th, 69 th and 81 th frames of the infrared pedestrian video sequence by using a common particle filtering method.
As can be seen from FIG. 3, when the target encounters a similar background, the tracking result of the invention is obviously better than that of the method 3, and the method 2 has a false following.
Simulation 3: with the three methods, a toy video sequence is tracked, and the result is shown in fig. 4, in which:
FIG. 4(a) is a diagram showing the results of tracking frames 21, 26, 30, 32 of a toy video sequence using the present invention;
FIG. 4(b) is a diagram showing the result of tracking the 21 st, 26 th, 30 th and 32 th frames of a toy video sequence by using a mean shift method;
fig. 4(c) is a diagram showing the result of tracking the 21 st, 26 th, 30 th and 32 th frames of the toy video sequence by using the ordinary particle filtering method.
As can be seen from fig. 4, when the target enters a complex background, the tracking effect of the present invention is significantly better than that of methods 2 and 3.
The monte carlo experiments are performed 100 times for the scenes in fig. 2 to 4, and the average tracking error Err and the average running time per frame RT are counted, and the results are shown in table 1:
TABLE 1
As can be seen from the statistics in table 1: in the aspect of tracking error, the tracking results of the three groups of video sequences are respectively reduced by 21%, 76% and 19% compared with a mean shift method, and are respectively reduced by 42%, 64% and 85% compared with a common particle filtering method; in terms of running time, because the description capacity of the particles to the target is increased, more accurate tracking can be realized by using fewer particles, the tracking speed is respectively reduced by 36%, 24% and 25% compared with the common particle filtering method, and the mean shift method is used for the least time.
In conclusion, although the mean shift method is used for the least time, the target cannot be accurately tracked under complex conditions, and the method is superior to the common particle filter method in terms of tracking precision and running time.
Claims (9)
1. A particle filter target tracking method based on reinforced particle swarm optimization comprises the following steps:
(1) initialization:
(1a) reading in an image I at time k-1k-1According to the initial state X of the object0Initializing the particle set at time k-1Wherein,the state interval of the ith particle at the time of k-1 is shown, i is the particle number and takes the values of 1,2, ·, N, N represents the total number of the particles, k represents the time, and the initial time k is 1;
(1b) initializing a target tracking window: b isk-1＝(rk-1,ck-1)TWherein r isk-1And ck-1Respectively representing the length and width values of a target tracking window at the k-1 moment, wherein T represents vector transposition operation;
(1c) initializing a local binary feature LBP statistical histogram M of a target as a feature template;
(2) target state prediction:
(2a) reading in an image I at time kkBy counting the particle sets at time k-1To obtain a predicted particle set at time kWhereinThe state prediction value of the ith particle at the moment k;
(2b) predicting a set of particles from time kAnd a target tracking window Bk-1Determining a candidate target set at time kWhereinIs the ith candidate target at the time k and is expressed byIs a center, Bk-1A rectangular area defined by the length and the width;
(3) extracting a set of candidate objectsFeature set corresponding to the candidate targetWherein ViRepresenting the ith candidate objectA corresponding feature;
(4) candidate target feature set is obtainedDistance set from feature template MAnd according to the distance setComputing a set of weights for candidate targetsWherein d isiRepresents the distance, ω, between the ith candidate target feature and the feature template MiRepresenting the weight of the ith candidate target;
(5) particle swarm optimization:
(5a) initializing a set of particles to be optimizedIndividual optimum value pid＝RiGroup optimum value pgd＝RmaxAnd the iteration time T is 1, and the maximum iteration time T is setmaxWhereinCorresponding set of velocities isThe set of weights isRmaxIs composed ofThe particle with the largest weight value corresponds to the weight value omegamax；
(5b) Judging whether the iteration number T satisfies T ≦ TmaxIf yes, carrying out the Tth iteration from the step (5c), otherwise, directly turning to the step (6);
(5c) according to individual optimum value pidAnd population optimum pgdSet of particles to be optimizedEach of the particles RiOptimized to obtain optimized particles QiAnd constitute an optimized particle set
(5d) Updating optimized particle setsEach of which optimizes the particle QiCorresponding weight value omegaiAnd updating the individual optimum value pid；
(5e) Updating population optimal value pgd：
(5f) By enhancing the search, the population is given the optimal value pgdUpdating:
(5f1) for the population optimum value pgdPerforming the first search to obtain a search result pg1：
pg1＝pgd+(Xmax-Xmin)·Gaussin(0,h)，
Wherein, XmaxRepresenting the upper bound, X, of the decision vectorminRepresentsThe lower bound of the decision vector, Gaussin (0, h), is a gaussian distribution with mean zero and standard deviation h, h-1/T according to the lawmaxPerforming linear decreasing;
(5f2) for the population optimum value pgdAnd (3) carrying out first updating:
if p isg1Weight ω of corresponding candidate targetg1Satisfy omegag1＞ωmaxThen update the population optimal value pgd＝pg1Which corresponds to the weight omegamax＝ωg1Else pgdThe value is not changed, and the value corresponds to the weight omegamaxThe change is not changed;
(5f3) for the population optimum value pgdPerforming the second search to obtain a second search result pg2：
pg2＝pgd+(Xmax-Xmin)·Cauchy(0,s)，
Wherein Cauchy (0, s) is a Cauchy distribution with a peak parameter of zero and a scale parameter of s, s being s-1/T according to the lawmaxPerforming linear decreasing;
(5f4) for the population optimum value pgdAnd (5) carrying out second updating:
if p isg2Weight ω of corresponding candidate targetg2Satisfy omegag2＞ωmaxThen the population optimum value pgd＝pg2Which corresponds to the weight omegamax＝ωg2Else pgdThe value is not changed, and the value corresponds to the weight omegamaxThe change is not changed;
(5h) judging whether the iteration number T satisfies T > TmaxIf yes, finishing particle swarm optimization, otherwise, making T equal to T +1 and making the particle set to be optimizedWhich corresponds to a set of weights asAnd returning to the step (5 b);
(6) weight set for candidate targetNormalization is carried out, and the result after normalization is obtainedOptimizing particle sets for time kResampling is carried out to obtain a particle set at the k momentWhereinThe state interval of the ith particle at the k moment is shown;
(7) particle set according to time kEstimating a target state X at time kkDetermining a Target estimation range Target at the moment k;
(8) and (4) checking whether the information at the next moment arrives, if so, making k equal to k +1, and returning to the step (2), otherwise, ending the target tracking process.
2. The method of claim 1, wherein the particle set at time k-1 is initialized in step (1a)Is determined by the following steps:
(1a1) according to the initial state X of the target0Generating the ith particle at time k-1Wherein i is the serial number of the particle, the values are 1,2, N, N represents the total number of the particles,obey mean value of X0Gaussian distribution with variance of Ψ, X0Psi is the process noise variance for the initial state of the target;
(1a2) using the N particles obtained in step (1a1) to form a particle set
3. The method of claim 1, wherein the prediction of particle sets at time k in step (2a) is performedIs represented as follows:
whereinRepresenting the ith predicted particle at time k by counting the particles at time k-1The transfer is obtained by the following transfer formula:
wherein, wiIs state noise, obeys an average ofThe variance is a gaussian distribution of Ψ.
4. According to claim1, wherein in step (2b) the candidate target set at time kIs represented as follows:
whereinIndicating the ith predicted particle at time kThe corresponding candidate target is determined according to the following formula:
wherein,andrespectively representing the ith predicted particle at time kAbscissa and ordinate of (a), rk-1And ck-1Respectively representing the length and width values of the target tracking window at the time k-1.
5. The method of claim 1, wherein the set of candidate objects is extracted in step (3)Feature set corresponding to the candidate targetIs determined by the following steps:
(3a) computing a set of candidate objectsEach candidate target ofLocal binary characteristic LBP value of each pixel:
wherein, gcRepresents the center point pcGray value of gnRepresents the center point pcP represents p pixel points in the neighborhood around the central pixel point, n is the serial number of the p pixel points, s (g)n-gc) Is a function, represented in the form:
(3b) calculating a candidate target set according to the result obtained in the step (3a)In each candidate targetLocal binary feature LBP statistical histogram Vi：
Wherein m represents the coding length of the local binary feature LBP, and m is 2pI (x, y) ═ I) is a function expressed in the form:
(3c) using the features of the N candidate targets obtained in step (3b) to form a candidate target feature set
6. The method of claim 1, wherein the candidate set of target features is found in step (4)Distance set from feature template MAnd according to the distance setComputing a set of weights for candidate targetsIs determined by the following steps:
(4a) computing a set of candidate target featuresDistance set from feature template M
Wherein d isiIs shown asThe distance between the i candidate target features and the feature template,ρiis the babbitt coefficient between the ith candidate target feature and the feature template,
(4b) according to distance setComputing a set of candidate target weights
Wherein ω isiRepresenting the ith candidate objectThe calculation formula of the weight value of (2) is:
wherein d isiAnd (3) representing the distance between the ith candidate target feature and the feature template, wherein R is the feature observation noise variance.
7. The method of claim 1, wherein step (5c) is performed according to the individual optimal value pidAnd population optimum pgdSet of particles to be optimizedEach of the particles RiOptimized to obtain optimized particles QiAnd constitute an optimized particle setOptimizing particle setsIs determined by the following steps:
(5c1) updating a set of particles to be optimizedMesoparticle RiVelocity v ofi：
vi＝w·vi+c1·r1·(pid-Ri)+c2·r2·(pgd-Ri)，
Wherein w is an inertia weight, c1For a learning factor related to the individual optimum, c2For a learning factor related to the global optimum, r1And r2Is two different random numbers between (0, 1);
(5c2) using the velocity v obtained in step (5c1)iUpdating a set of particles to be optimizedOf (ii) particles RiTo obtain optimized particles Qi：
Qi＝Ri+vi；
(5c3) Using the N optimized particles obtained in the step (5c2) to form an optimized particle set
Its corresponding weight value set is
8. The method of claim 1, wherein the updating in step (5d) is performedOptimizing particle setsEach of which optimizes the particle QiCorresponding weight value omegaiAnd updating the individual optimum value pid(ii) a Is determined by the following steps:
(5d1) calculating an optimized particle QiTemporary weight value omeganew：
Wherein d isnewRepresents the optimized particle QiThe distance between the corresponding candidate target feature and the feature template;
(5d2) obtaining the temporary weight ω according to the step (5d1)newUpdating optimized particle QiCorresponding weight value omegaiAnd updating the individual optimum value pid：
If ω isnew＞ωiThen update the optimized particle QiCorresponding weight value omegai＝ωnewUpdating the individual optimum value pid＝QiOtherwise ω isiConstant value, pidThe value is unchanged.
9. The method of claim 1, wherein the population optimal value p is updated in step (5e)gdDetermined by the following method:
if particle Q is optimizediCorresponding weight value omegaiSatisfy omegai＞ωmaxThen update the population optimal value pgd＝QiIts corresponding weight value omegamax＝ωiElse pgdThe value is not changed, and the value corresponds to the weight omegamaxAnd is not changed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610230000.XA CN105913457B (en) | 2016-04-14 | 2016-04-14 | Based on the particle filter method for tracking target for reinforcing particle group optimizing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610230000.XA CN105913457B (en) | 2016-04-14 | 2016-04-14 | Based on the particle filter method for tracking target for reinforcing particle group optimizing |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105913457A CN105913457A (en) | 2016-08-31 |
| CN105913457B true CN105913457B (en) | 2018-11-16 |
Family
ID=56746861
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610230000.XA Active CN105913457B (en) | 2016-04-14 | 2016-04-14 | Based on the particle filter method for tracking target for reinforcing particle group optimizing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105913457B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106408590B (en) * | 2016-10-21 | 2019-03-08 | 西安电子科技大学 | Particle filter method for tracking target based on regression analysis |
| CN106558064A (en) * | 2016-12-07 | 2017-04-05 | 江南大学 | A kind of method for tracking target based on cellular quantum-behaved particle swarm optimization |
| CN107220993B (en) * | 2017-04-25 | 2019-11-01 | 西北工业大学 | Particle filter maneuvering target method for tracing based on Mean shift cluster |
| CN107301652B (en) * | 2017-05-04 | 2020-07-14 | 西北工业大学 | Robust target tracking method based on local sparse representation and particle swarm optimization |
| CN107248166B (en) * | 2017-06-29 | 2020-05-15 | 武汉工程大学 | Target state estimation method in dynamic environment |
| CN108776972B (en) * | 2018-05-04 | 2020-06-12 | 北京邮电大学 | Object tracking method and device |
| CN112039496B (en) * | 2020-09-10 | 2024-04-19 | 南京航空航天大学 | Rapid Gaussian particle filter data fusion method based on artificial fish swarm optimization |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103985141A (en) * | 2014-05-28 | 2014-08-13 | 西安电子科技大学 | Target tracking method based on HSV color covariance characteristics |
| CN104036525A (en) * | 2014-06-18 | 2014-09-10 | 上海交通大学 | Self-adaptive particle swarm optimization (PSO)-based particle filter tracking method |
-
2016
- 2016-04-14 CN CN201610230000.XA patent/CN105913457B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103985141A (en) * | 2014-05-28 | 2014-08-13 | 西安电子科技大学 | Target tracking method based on HSV color covariance characteristics |
| CN104036525A (en) * | 2014-06-18 | 2014-09-10 | 上海交通大学 | Self-adaptive particle swarm optimization (PSO)-based particle filter tracking method |
Non-Patent Citations (4)
| Title |
|---|
| 基于多策略协同作用的粒子群优化算法;李俊 等;《计算机应用》;20160310;第36卷(第3期);第681-686页 * |
| 基于局部区域LBP特征的人脸识别研究;王聃 等;《黑龙江工程学院学报（自然科学版）》;20120930;第26卷(第3期);第46-48页，第59页 * |
| 改进的粒子群优化粒子滤波的人脸跟踪;严石 等;《淮阴工学院学报》;20160229;第25卷(第1期);第1节，第3节 * |
| 自适应权重粒子群优化的粒子滤波算法;刘峰 等;《计算机仿真》;20131130;第30卷(第11期);第4.2节 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105913457A (en) | 2016-08-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105913457B (en) | Based on the particle filter method for tracking target for reinforcing particle group optimizing | |
| CN109636829B (en) | Multi-target tracking method based on semantic information and scene information | |
| CN107180426B (en) | Migratable multi-model integration-based computer-aided lung nodule classification device | |
| CN105405151B (en) | Anti-shelter target tracking based on particle filter and weighting Surf | |
| CN112489083B (en) | Image feature point tracking matching method based on ORB-SLAM algorithm | |
| CN103530893B (en) | Based on the foreground detection method of background subtraction and movable information under camera shake scene | |
| CN115311329B (en) | Video multi-target tracking method based on double-link constraint | |
| CN109859238A (en) | One kind being based on the optimal associated online multi-object tracking method of multiple features | |
| CN105335986A (en) | Characteristic matching and MeanShift algorithm-based target tracking method | |
| CN111626200A (en) | Multi-scale target detection network and traffic identification detection method based on Libra R-CNN | |
| CN102142085B (en) | Robust tracking method for moving flame target in forest region monitoring video | |
| CN108320293A (en) | A kind of combination improves the quick point cloud boundary extractive technique of particle cluster algorithm | |
| CN109242019B (en) | Rapid detection and tracking method for optical small target on water surface | |
| CN113744311A (en) | Twin neural network moving target tracking method based on full-connection attention module | |
| CN113344256B (en) | System and method for predicting multiple degrees of freedom motion characteristics of shield attitude and evaluating control performance | |
| CN113516713B (en) | Unmanned aerial vehicle self-adaptive target tracking method based on pseudo twin network | |
| CN105427340B (en) | Based on the video target tracking method for improving case particle filter | |
| CN114092517A (en) | Multi-target tracking method based on traditional and deep learning algorithm | |
| CN113920159A (en) | Infrared aerial small target tracking method based on full convolution twin network | |
| CN107871315A (en) | A kind of video image motion detection method and device | |
| CN103985139B (en) | Particle filter target tracking method based on color model and prediction vector cluster model information fusion | |
| CN112200831B (en) | Dynamic template-based dense connection twin neural network target tracking method | |
| CN112270696B (en) | Discriminative target tracking method based on fusion meta classifier | |
| CN113806857A (en) | High-speed train energy-saving braking method based on variational graph self-encoder | |
| CN117634486A (en) | Directional 3D instance segmentation method based on text information |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |