CN102542577A - Particle state evaluation method - Google Patents
Particle state evaluation method Download PDFInfo
- Publication number
- CN102542577A CN102542577A CN2011104356312A CN201110435631A CN102542577A CN 102542577 A CN102542577 A CN 102542577A CN 2011104356312 A CN2011104356312 A CN 2011104356312A CN 201110435631 A CN201110435631 A CN 201110435631A CN 102542577 A CN102542577 A CN 102542577A
- Authority
- CN
- China
- Prior art keywords
- particle
- state
- filter
- filtering
- target
- 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.)
- Pending
Links
Images
Landscapes
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a particle state evaluation method, which is suggested aiming at solving the problem of unstable filtering or even diffusing of filtering caused by large evaluation deviation or even wrong evaluation of initial prior state of particles due to insufficient prior information in a traditional particle filtering initial phase. The particle state evaluation method obtains state information by evaluating state of target observation data, applies the obtained state information to an importance sampling process in the particle filtering, revises particle weight numbers, further can solve the problem of unstable filtering or even diffusing caused by wrong evaluation of initial prior state of particles due to insufficient prior information in a traditional particle filtering initial phase, and improves convergence speed and accuracy of target evaluation.
Description
Technical field
The invention belongs to the target following technical field, be specifically related to a kind of particle state method of estimation.
Background technology
In the maneuvering target tracking field, particle filter has obtained deep research, if do not consider other factors, target following is exactly simple Filtering Estimation problem.Particle filter is owing to be applicable to any NLS that can represent with state-space model, and the NLS that can't represent of legacy card Kalman Filtering, and precision can be approached optimal estimation, so be widely applied and target tracking domain.But itself also exists some problems.Particle filter need could obtain relatively good filtering effect under certain assumed condition.At first, state equation x
K+1=f
k(x
k, u
k) (wherein x represents state, and k represents constantly, and u represents process noise) and measurement equation z
k=h
k(x
k, v
k) (z represents observer state, and v represents observation noise) known; Secondly can be from the process noise u that distributes
kWith sample in the prior distribution; Moreover be used for the likelihood function p of point estimation
k(z
k| x
k) known.
Traditional particle state method of estimation synoptic diagram is as shown in Figure 1, comprises the steps:
Initialization is from priori probability density p (X
0) middle randomly drawing sample point
Sampling primary collection
(upper right corner variable i is represented particle, and population adds up to N);
Importance sampling, the particle
of from suggestion distribution importance density function, sampling
The importance weighting is according to formula
Calculate the weighting of particle importance
And carry out normalization
State estimation
is by particle weights and state estimation dbjective state value;
Resampling step: in iterative process; The effective sample number can reduce promptly to occur the particle degradation phenomena; When particle degenerated to certain degree, population was carried out valuation with regard to not being used for to state, solved degenerate problem so increase resampling step increases number of particles.According to weights
; Duplicate high weight respectively; Abandon the particle of low weights; Obtain N new particle, normalization then.This step is one can select step.
Prior distribution can receive condition effect such as the randomness of particle filter environment, the maneuverability of tracking target and the deviation of recording geometry and be difficult to obtain information accurately.In inaccurate observation space, adopt the particle importance of weight weighting to estimate actual value, so obviously can not obtain to estimate accurately.Same posterior probability is difficult to obtain more, so consider from importance function, to sample to obtain.Under many circumstances, be the comparison difficulty from the sampling of the important density function of optimum, a kind of easier and method that be easy to realize selects probability density function as important density function exactly.That is:
obviously; Bringing following formula into important density calculation formula can get:
owing to fail to utilize up-to-date metrical information in the important density function in the following formula, want big many with sample its variance of weight coefficient that the variance that obtains weight coefficient obtains as important density compared with optimum density function
of priori density function.Simpler and be easy to realize because of its form, in particle filter algorithm, be widely used.But thisly substitute for tracker and be not suitable for, be difficult to realize with extensive sampling constantly because the motion state parameter of multidimensional is difficult to measure and follows the tracks of detection system.
Summary of the invention
The objective of the invention is to estimate to cause filtering unstable even disperse the problem of existence, proposed a kind of particle state method of estimation for to overcome traditional particle filter starting stage because prior imformation is not enough to the initial priori state estimation of particle deviation big even wrong.
Technical scheme of the present invention is: a kind of particle state method of estimation; Be specially: the state to the target observation data is estimated; Obtain status information; The status information that obtains is used in the importance sampling process of particle filter the particle weights are revised, obtains the population of estimated state.
Further, the state of target observation data is estimated specifically to use the Kalman wave filter.
Beneficial effect of the present invention: method of the present invention is through estimating the state of target observation data; Obtain status information; The status information that obtains is used in the importance sampling process of particle filter; The particle weights are revised, and then can have been overcome the problem of the initial priori state of particle being carried out wrong estimation and then causing the filtering instability even disperse, improved the accuracy of convergent speed and target estimation owing to filtering starting stage prior imformation is not enough.
Figure of description
Fig. 1 is existing particle state method of estimation principle schematic.
Fig. 2 is a particle state method of estimation process flow diagram of the present invention.
Fig. 3 test in one adopt that existing particle state method of estimation estimates inaccurate formation disperse the form synoptic diagram.
Fig. 4 tests and adopts method of the present invention to follow the tracks of the simulation result synoptic diagram in one.
Fig. 5 test in two adopt that existing particle state method of estimation estimates inaccurate formation disperse the form synoptic diagram.
Fig. 6 tests and adopts method of the present invention to follow the tracks of the simulation result synoptic diagram in two.
Embodiment
Below in conjunction with accompanying drawing and concrete embodiment the present invention is done further elaboration.
Particle state method of estimation of the present invention; Be specially: the state to the target observation data is estimated, obtains status information, the status information that obtains is used in the importance sampling process of particle filter; The particle weights are revised, obtained the population of estimated state.
Here, the state of target observation data is estimated specifically can use the Kalman wave filter.Here by the Kalman wave filter; Being based on the theoretical method of optimal estimation comes the initial priori state of particle is estimated; Employing to the sampling of the particle state in the particle filter, makes it to be applicable to actual filtering application environment based on the method for parameter estimation of Kalman filtering.Because filtering starting stage prior imformation is not enough the initial priori state of particle is carried out the wrong drawback of estimating and then causing the filtering instability even disperse with regard to having overcome like this.Can find out, also can adopt the wave filter of other form that the state of target observation data is estimated here.
Here, set up initial particle population according to existing information on the one hand, also open a Kalman wave filter simultaneously.Detection of a target positional information then; Carry out the importance weight calculating of particle filter and the iteration estimation of Kalman wave filter with this information; And the valuation of adopting Kalman filtering at initial period carries out analysis and assessment to the particle filter importance weight then as final result, if the priori situation that importance weight can not the reasonable reaction tracker; Particle all differs greatly with observed value, at this time continues to adopt the estimated value of Kalman wave filter that primary is carried out update all.If the priori situation of ability reasonable reaction tracker adopts the filtering of particle filter independence in order to reduce calculated amount separately with regard to not needing the Kalman wave filter.Through corresponding iteration several times; When after adopting the correction of Kalman filter data, obtaining importance weight comparison operators resultant motion target detection system randomness; Then particle filter can be realized comparatively tenacious tracking, and just can carry out normal particle filter and carry out the moving target parameter estimation this moment.
For the ease of to understanding of the present invention, the method for present embodiment is done further detailed description, specifically not like Fig. 2 institute:
S1. an elementary particle wave filter is opened in particle filter (PF) initialization, the particle of taking a sample, importance sampling, weights normalization, resampling and state estimation, the independent filter forecasting of accomplishing the tracking target state;
S2. when particle filter is estimated dbjective state, open a Kalman wave filter (KF), simultaneously this target is carried out tracking filter and estimate;
S3. in the filtering starting stage; The state estimation and the state covariance structure importance density function that utilize the Kalman filter filtering to obtain; Particle filter utilizes this importance density function to carry out importance sampling, calculates particle weights and normalization, accomplishes the correction to the particle parameter;
The correction here be appreciated that for: adopt state estimation and the state covariance structure importance density function of Kalman filtering to carry out particle weights that importance sampling obtains and particle filter and only utilize to observe and carry out the particle weights that importance sampling obtains and compare; That is be to recomputate particle weights and normalization through the Kalman wave filter, rather than the importance density function that uses particle filter directly to utilize observation information to obtain carries out importance sampling.
S4. the particle filter weights are assessed, when importance weight met moving target detection system randomness, then particle filter can be realized comparatively tenacious tracking, cancelled the Kalman wave filter, used particle filter independently tracked;
Here, described importance weight meets moving target detection system randomness and can adopt following process to judge, as effective sample size N
EffDuring＜0.9 * N, then importance weight meets moving target detection system randomness, wherein,
Be the particle weights, N is the population sum.
S5. the resampling particle gets into next filtering constantly.
The design philosophy of the inventive method be since particle filter in the particle filter incipient stage; Prior imformation is difficult to accurately obtain; In tracker, also can not realize simultaneously shortcoming that the initial detection information of system is sampled on a large scale, utilize the characteristics of Kalman wave filter linearity optimum that the initial value of particle filter is revised.
The essence of the inventive method is based on the theoretical method of optimal estimation and comes the initial priori state of particle is estimated; Employing is based on the method for parameter estimation of Kalman filter filtering; To the sampling of the particle state in the particle filter, make it to be applicable to actual filtering application environment.After the particle filter system gets into tenacious tracking; Then no longer utilize the Kalman wave filter to revise, because filtering starting stage prior imformation is not enough the initial priori state of particle is carried out the wrong problem of estimating and then causing the filtering instability even disperse with regard to having overcome like this.
For the performance of assessment the inventive method, carried out the emulation experiment under the different condition.Hypothetical target moves on two dimensional surface, and the observation of x direction and y direction is independent mutually, and the observation noise mean square deviation is identical, and the flight path information of observation is carried out filtering.
Emulation experiment one: in 0.1 second sampling period, target x direction of principal axis initial velocity is 515m/s, and y direction of principal axis initial velocity is 0m/s, and the x axle acceleration is 0m/s
2, the y axle acceleration is 9.8m/s
2, the standard deviation of measurement is 50m.
Adopt the elementary particle filtering method; Estimate inaccurate the time that when the target original state adopt the value of first three time detection to carry out initial estimation, its filtering image is that two dimensional motion target simulator result is as shown in Figure 3; Wherein, The longitudinal axis is represented the Y axial coordinate of object run, and transverse axis is represented the X axial coordinate of object run, and unit is a rice.The target initial position is (0,12100), begins to carry out uniformly accelerated motion then, and this emulation is in order to simulate a level throw motion, and target receives downward gravitation, downwards operation.Wherein, solid line has been represented the true flight path of object run, and x type curve is to have added error in measurement systematic observation value afterwards, and dotted line is that wave filter is to measuring observed value acquisition filter value afterwards.Filter curve has just departed from true flight path soon in observation, shows that this filtering is unsuccessful, and dispersing has appearred in target following.
Can find that from Fig. 3 the starting stage particle filter only can be estimated the initial situation of motion state according to the some mark of surveying, in this case, estimated value has the comparison large deviation, and therefore along with the increase of deviation, particle state more and more departs from the real goal flight path.So just caused and followed the tracks of the result who loses, be difficult to satisfy actual tracker demand.
Equally under these conditions, adopt method of the present invention, start the filtering that Kalman wave filter and particle filter carry out track at first simultaneously.The particle filter original state adopts Kalman wave filter result calculated to produce, and after particle filter tracking is stable, do not re-use the information of Kalman filtering, and the tracking of employing itself is carried out.
The result is as shown in Figure 4 in the two dimensional motion target simulator, and wherein, the longitudinal axis is represented the Y axial coordinate of object run, and transverse axis is represented the X axial coordinate of object run, and unit is a rice.The target initial position is (0,12100), begins to carry out uniformly accelerated motion then, and this emulation is in order to simulate a level throw motion, and target receives downward gravitation, downwards operation.Wherein, solid line has been represented the true flight path of object run, and x type curve is to have added error in measurement systematic observation value afterwards, and dotted line is that wave filter is to measuring observed value acquisition filter value afterwards.As can be seen from the figure the filtering flight path just begins to approach true flight path more than the observation flight path observing soon, shows that the successful realization of this filtering follows the tracks of target, and noise has obtained inhibition.
As can beappreciated from fig. 4; Though the incipient stage since the Kalman wave filter can not be very fast the original state that just detects target, when dbjective state that ought detect subsequently and True Data were more approaching, particle filter just can adopt this numerical value as particle initial estimation state; Begin to generate distribution of particle then; Thereby in tracking subsequently, slowly converge on the true flight path, and dispersing of track can not occur.
Emulation experiment two: in 0.1 second sampling period, target x direction of principal axis initial velocity is 515m/s, and y direction of principal axis initial velocity is 0m/s, and the x axle acceleration is-16m/s
2, the y axle acceleration is 9.8m/s
2, the standard deviation of measurement is 50m.
With respect to emulation experiment one, emulation experiment two has all had acceleration change at both direction.Still adopt traditional particle filter method and method of the present invention to carry out emulation respectively.
Fig. 5 is the simulation result that adopts the elementary particle filtering method, and wherein, the longitudinal axis is represented the Y axial coordinate of object run, and transverse axis is represented the X axial coordinate of object run, and unit is a rice.The target initial position is (0,0), begins to carry out uniformly accelerated motion then, and this emulation is in order to simulate a turning motion, and target receives the acceleration of a turning, and this acceleration is all important on X axle and Y axle, causes the turning motion of one 360 degree of target.Wherein, solid line has been represented the true flight path of object run, and x type curve is to have added error in measurement systematic observation value afterwards, and dotted line is that wave filter is to measuring observed value acquisition filter value afterwards.The filtering flight path just begins to fall behind and true flight path in observation soon, and in the real process, big like this error can cause predicting the fall short position, causes to follow the tracks of and disperses.
Under the above-mentioned moving condition of target, because the original state estimation is inaccurate, along with the continuous prolongation of target travel time, the accumulation of error of tracking is bigger.Estimation range is surveyed in the final goal monitoring that will fly out, thereby is unfavorable in the actual monitored process predictive filtering to target.
Fig. 6 is the simulation result that adopts method of the present invention, and wherein, the longitudinal axis is represented the Y axial coordinate of object run, and transverse axis is represented the X axial coordinate of object run, and unit is a rice.The target initial position is (0,0), begins to carry out uniformly accelerated motion then, and this emulation is in order to simulate a turning motion, and target receives the acceleration of a turning, and this acceleration is all important on X axle and Y axle, causes the turning motion of one 360 degree of target.Wherein, solid line has been represented the true flight path of object run, and x type curve is to have added error in measurement systematic observation value afterwards, and dotted line is that wave filter is to measuring observed value acquisition filter value afterwards.Can find out that from last figure filter curve just begins to approach true flight path more than x type curve observing soon, show that the successful realization of this filtering follows the tracks of target, noise has obtained inhibition.
As can beappreciated from fig. 6, method of the present invention can realize the accurate tracking to target, can not occur dispersing.
Those of ordinary skill in the art will appreciate that embodiment described here is in order to help reader understanding's principle of the present invention, should to be understood that protection scope of the present invention is not limited to such special statement and embodiment.Those of ordinary skill in the art can make various other various concrete distortion and combinations that do not break away from essence of the present invention according to these teachings disclosed by the invention, and these distortion and combination are still in protection scope of the present invention.
Claims (4)
1. particle state method of estimation; Be specially: the state to the target observation data is estimated, obtains status information, the status information that obtains is used in the importance sampling process of particle filter; The particle weights are revised, obtained the population of estimated state.
2. particle state method of estimation according to claim 1 is characterized in that, described state to the target observation data estimates specifically to use the Kalman wave filter.
3. particle state method of estimation according to claim 2 is characterized in that, specifically comprises the steps:
S1. an elementary particle wave filter is opened in particle filter initialization, and the particle of taking a sample, importance sampling, weights normalization, resampling and state estimation are independently accomplished the filter forecasting of tracking target state;
S2. when particle filter is estimated dbjective state, open a Kalman wave filter, simultaneously this target is carried out tracking filter and estimate;
S3. in the filtering starting stage; The state estimation and the state covariance structure importance density function that utilize the Kalman filter filtering to obtain; Particle filter utilizes this importance density function to carry out importance sampling, calculates particle weights and normalization, accomplishes the correction to the particle parameter;
S4. the particle filter weights are assessed, when importance weight met moving target detection system randomness, then particle filter can be realized comparatively tenacious tracking, cancelled the Kalman wave filter, used particle filter independently tracked;
S5. the resampling particle gets into next filtering constantly.
4. particle state method of estimation according to claim 3 is characterized in that, the described importance weight of step S4 meets moving target detection system randomness and adopts following process to judge, as effective sample size N
EffDuring＜0.9 * N, then importance weight meets moving target detection system randomness, wherein,
Be the particle weights, N is the population sum.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2011104356312A CN102542577A (en) | 2011-12-22 | 2011-12-22 | Particle state evaluation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2011104356312A CN102542577A (en) | 2011-12-22 | 2011-12-22 | Particle state evaluation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102542577A true CN102542577A (en) | 2012-07-04 |
Family
ID=46349393
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2011104356312A Pending CN102542577A (en) | 2011-12-22 | 2011-12-22 | Particle state evaluation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102542577A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106338651A (en) * | 2016-08-31 | 2017-01-18 | 长沙理工大学 | Particle filter analysis method applied to lower frequency oscillation mode identification of power system |
| CN112762928A (en) * | 2020-12-23 | 2021-05-07 | 重庆邮电大学 | ODOM and DM landmark combined mobile robot containing laser SLAM and navigation method |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101644758A (en) * | 2009-02-24 | 2010-02-10 | 中国科学院声学研究所 | Target localization and tracking system and method |
-
2011
- 2011-12-22 CN CN2011104356312A patent/CN102542577A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101644758A (en) * | 2009-02-24 | 2010-02-10 | 中国科学院声学研究所 | Target localization and tracking system and method |
Non-Patent Citations (4)
| Title |
|---|
| 《计算机学报》 20080215 王法胜等 一种用于解决非线性滤波问题的新型粒子滤波算法 第347页右栏，第349页，第350页左栏第1段 1－3 第31卷, 第2期 * |
| 李涛: "一种使用非等权值粒子的确定性粒子滤波算法", 《国防科技大学学报》 * |
| 王法胜等: "一种用于解决非线性滤波问题的新型粒子滤波算法", 《计算机学报》 * |
| 雷明: "扩展卡尔曼粒子滤波算法的一种修正方法", 《西安交通大学学报》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106338651A (en) * | 2016-08-31 | 2017-01-18 | 长沙理工大学 | Particle filter analysis method applied to lower frequency oscillation mode identification of power system |
| CN106338651B (en) * | 2016-08-31 | 2018-09-14 | 长沙理工大学 | Particle filter analysis method applied to low-frequency oscillation of electric power system pattern-recognition |
| CN112762928A (en) * | 2020-12-23 | 2021-05-07 | 重庆邮电大学 | ODOM and DM landmark combined mobile robot containing laser SLAM and navigation method |
| CN112762928B (en) * | 2020-12-23 | 2022-07-15 | 重庆邮电大学 | ODOM and DM landmark combined mobile robot containing laser SLAM and navigation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Konatowski et al. | Comparison of estimation accuracy of EKF, UKF and PF filters | |
| CN103776453B (en) | A kind of multi-model scale underwater vehicle combined navigation filtering method | |
| CN106407677B (en) | A kind of multi-object tracking method in the case of missing measurement | |
| CN105760811B (en) | Global map closed loop matching process and device | |
| CN107084714B (en) | A kind of multi-robot Cooperation object localization method based on RoboCup3D | |
| CN101975575B (en) | Multi-target tracking method for passive sensor based on particle filtering | |
| CN104730537B (en) | Infrared/laser radar data fusion target tracking method based on multi-scale model | |
| CN105205313A (en) | Fuzzy Gaussian sum particle filtering method and device as well as target tracking method and device | |
| CN101980044B (en) | Method for tracking multiple targets under unknown measurement noise distribution | |
| CN105372659A (en) | Road traffic monitoring multi-target detection tracking method and tracking system | |
| CN103973263B (en) | Approximation filter method | |
| CN108319570B (en) | Asynchronous multi-sensor space-time deviation joint estimation and compensation method and device | |
| CN110738275B (en) | UT-PHD-based multi-sensor sequential fusion tracking method | |
| CN101701826A (en) | Target tracking method of passive multi-sensor based on layered particle filtering | |
| CN103152826A (en) | Moving target tracking method based on NLOS (non line of sight) state inspection compensation | |
| Straka et al. | Performance evaluation of local state estimation methods in bearings-only tracking problems | |
| CN107064865A (en) | The passive co-located method of polar coordinates Dynamic Programming clustered based on depth | |
| CN112146648A (en) | Multi-target tracking method based on multi-sensor data fusion | |
| CN102542577A (en) | Particle state evaluation method | |
| CN104463841A (en) | Attenuation coefficient self-adaptation filtering method and filtering system | |
| Delahaye et al. | TAS and wind estimation from radar data | |
| CN105549003A (en) | Automobile radar target tracking method | |
| CN104320108A (en) | AHCIF based centralized measurement value weighted fusion method | |
| CN104268597A (en) | AHCIF-based centralized measured value dimension-expansion fusion method | |
| Zhang et al. | The research for a kind of information fusion model based on BP neural network with multi position sources and big data selection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C12 | Rejection of a patent application after its publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20120704 |