內(nèi)容新手區(qū)-slam基礎(chǔ)robot master filter_第1頁
內(nèi)容新手區(qū)-slam基礎(chǔ)robot master filter_第2頁
內(nèi)容新手區(qū)-slam基礎(chǔ)robot master filter_第3頁
內(nèi)容新手區(qū)-slam基礎(chǔ)robot master filter_第4頁
內(nèi)容新手區(qū)-slam基礎(chǔ)robot master filter_第5頁
已閱讀5頁,還剩75頁未讀 繼續(xù)免費閱讀

下載本文檔

版權(quán)說明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請進行舉報或認領(lǐng)

文檔簡介

1、The Fundamental of Mobile RobotMachine Perception and Interaction Group (MPIG) Send an email with specified subject Robot stuff to An automatic reply will come. All you need would be included! All the codes panying this course are open-sourced at Github. The project homepage: The git repository:Lear

2、n more about git: (software) The best GUI-based git client on Windows:Particle filterA Tutorial on Particle Filtering and Smoothing: Fifteen years later The particle filter was proposed for full nonlinear and non-Gaussian estimation, and has e a serious alternative for real-time applications classic

3、ally approached by model-based Kalman filter techniques. The more nonlinear the model is, or the more non-Gaussian the noise is, the more potential the PF has, especially in applications where computational power is rather cheap and the sampling rate is slow. Generally, the basic theory of the gener

4、al PF includes the sequential importance sampling (SIS) algorithm, the choice of importance density and resampling.A General Convergence Result for Particle Filtering.pdfA Particle Filter for Bid Estimation in Ad Auctions with.pdfA Theoretical Framework for Sequential Importance Sampling with Resamp

5、ling. Statistics for Engineering and Information Science. Springer.pdfA Tutorial on Bayesian Estimation and Tracking techniques applicable to nonlinear and non-gaussian processes.pdfA Tutorial on Particle Filtering and Smoothing fifteen years later.pdfA tutorial on particle filters for online nonlin

6、earnon-Gaussian Bayesian tracking.pdfan overview of exsiting Methods and Recent Advances in SMC.pdfAn Overview of Sequential Bayesian Filtering in Ocean Acoustics.pdfAn overview of sequential Monte Carlo methods for parameter estimation in general state-space models.pdfAn Overview on Bayesian Tracki

7、ng for Video analytics.pdfBayesian Filtering-From Kalman Filters to particle filter.pdfBayesian Tracking for Video Analytics.pdfImprovement Strategies for Monte Carlo Particle Filters.pdfNew probabilistic inference algorithms that harness the strengths of variational and monte carlo methods.pdfNonli

8、near Filters Beyond the Kalman filter.pdfONLINE SEQUENTIAL MONTE CARLO EM ALGORITHM.pdfParticle Filter Theory and Practice with positioning applications.pdfparticle filteringIEEE signal process. Magazine.pdfParticle Filters for Positioning, Navigation and tracking.pdfParticle Filters in Robotics.pdf

9、Particle_Filters_in_Decision_Making_Problems_under_Uncertainty.pdfRecent Developments in Auxiliary Particle Filtering.pdfPF in geophysical systems.pdfSequential Decision Making in General state space models.pdfSequential Monte Carlo Methodsa Survey.pdfstochastic_filtering_ a brief tutorial.pdfUnscen

10、ted Filtering and Nonlinear Estimation.pdfParticle Distribution Optimization (PDO): A Fundamental Requirement of Particle Filters.pdf Parametric filtering the fixed functional form of posteriorMeanCovarianceUnimodalMulti-hypothesis Kalman filter Approximates the nonlinear transformationEKFEKFPFFrom

11、EKF to PFGeneral Nonlinear modelsNonlinear motion modelNonlinear sensor modelPredictionUpdateInnovation / ResidualInnovation varianceKalman gainExtended Kalman filterPredictionUpdateParticle filterfor i = 1 : mendfor i = 1 : mendPFInitializationPredictionWeightResampling21222324252627282930313233343

12、53637Generate arbitrary random numberPF initializationPredictionUpdateParticle filterfor i = 1 : mendfor i = 1 : mendPFfunction r = arbitraryRand(nRow, nCol, minR, maxR) if minR maxR error(minR should be smaller than maxR!); end r = rand(nRow, nCol) * (maxR - minR) + minR;endm = 1000;xLim = -100, 10

13、0;yLim = 30, 190;p = arbitraryRand(1, m, xLim(1), xLim(2);. arbitraryRand(1, m, yLim(1), yLim(2);w = rand(1, m);%figure(1)scatter(p(1,:), p(2,:), 31, w, fill);axis(-130, 130, 0, 220);axis equal;P59Compute weightsPredictionUpdateParticle filterfor i = 1 : mendfor i = 1 : mendPFfunction w = compWeight

14、(zPred, zTrue, Sigma, hasAngle, idxAngle)Always bound the resulting angles using function piTopi.General solution to 0=2pi problemfunction w = compWeight(zPred, zTrue, Sigma, hasAngle, idxAngle) Y = abs(zPred - zTrue); if hasAngle Y(idxAngle) = piTopi( Y(idxAngle) ); end w = det(2 * pi * Sigma)(-0.5

15、) * exp(-0.5 * Y/Sigma * Y) + 1e-99;Sigma = 0.1;zTrue = degtorad(179);zPredA = degtorad(178); % error: 1zPredB = degtorad(-179); % error: 2zPredC = degtorad(-178); % error: 3zPredD = degtorad(185); % error: 6zPredE = degtorad(0); % error: 179% The bigger the error, the smaller the weight.wA = compWe

16、ight(zPredA, zTrue, Sigma, true)P60Sigma = 0.1;zTrue = degtorad(179);zPredA = degtorad(178); % error: 1zPredB = degtorad(-179); % error: 2zPredC = degtorad(-178); % error: 3zPredD = degtorad(185); % error: 6zPredE = degtorad(0); % error: 179% The bigger the error, the smaller the weight.wA = compWei

17、ght(zPredA, zTrue, Sigma, true)P60wA = 1.2596wB = 1.2539wC = 1.2444wD = 1.1943wE = 8.0687e-22ResamplingPredictionUpdateParticle filterfor i = 1 : mendfor i = 1 : mendPF The most difficult component in PFA huge number of literatures focus on the topic of sequence importance resampling.Four flavor are

18、 implemented in function resampleParticle newInd = resampleParticle(weight, hugh);newInd = resampleParticle(weight, hugh);n = 500;m = 5;particle = zeros(2, n);% weight = randn(1,n).2;weight = exp(randn(1,n); %figure(name, Test resampling, units,normalized,. outerposition,0 0 1 1);subplot(m, 1, 1)ste

19、m(weight, markersize, 3, MarkerFaceColor,red,MarkerEdgeColor,red);ylabel(Weight);xlim(1, n);P62Mean angles PredictionUpdateParticle filterfor i = 1 : mendfor i = 1 : mendPFWhen calculating theaverage or mean of an angleone has to take into account how angles wrap around so that any angle in degrees

20、plus any integer multiple of 360 degrees is a measure of the same angle.If one wanted an average direction of the wind over two readings where the first reading was of 350 degrees and the second was of 10 degrees then the average of the numbers is 180 degrees, whereas if you can note that 350 degree

21、s is equivalent to -10 degrees and so you have two readings at 10 degrees either side of zero degrees leading to a more fitting mean angle of zero degrees.Assume all angles are on the unit circle and convert them to complex numbers expressed in real and imaginary form.Compute the mean of the complex

22、 numbers.Convert the complex mean to polar coordinates whereupon the phase of the complex mean is the required angular mean.function outAng = meanAngle(inAng, dim, sens) in = exp( inAng * 1i ); % According to Eulers formulamid = mean( in, dim );outAng = atan2( imag(mid), real(mid) );outAng( abs(mid)

23、 sens ) = nan; P63angs = degtorad(-177, 179);disp(radtodeg(meanAngle(angs); angs = degtorad(350, 10);disp(radtodeg(meanAngle(angs); angs = degtorad(10, 20, 30);disp(radtodeg(meanAngle(angs); angs = degtorad(90, 180, 270, 360);disp(radtodeg(meanAngle(angs);P63 -179 -1.6148e-15 20.0000 -90Implement PF

24、 with bearing-only sensorMotion modelnonlinearMeasurementmodelnonlinearP61Just play with the codes!Note that the performance is quite sensitive to the number of particles and the process noise. Generally, the process noise should be larger than the motion noise employed in generating the ground truth to ensure the sampled particle set more probably covers the true state. Tune the arguments and see the results.PF Demo with Bearing & range Motion modelnonlinearMeasurementmodelnonlinearParticle filters are an implementation of recursive Bayesian filteringThey represe

溫馨提示

  • 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
  • 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
  • 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內(nèi)容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
  • 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
  • 5. 人人文庫網(wǎng)僅提供信息存儲空間,僅對用戶上傳內(nèi)容的表現(xiàn)方式做保護處理,對用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對任何下載內(nèi)容負責。
  • 6. 下載文件中如有侵權(quán)或不適當內(nèi)容,請與我們聯(lián)系,我們立即糾正。
  • 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。

評論

0/150

提交評論