您好,欢迎来到筏尚旅游网。
搜索
您的当前位置:首页结合初值选取的最大似然估计被动定位算法(英文)

结合初值选取的最大似然估计被动定位算法(英文)

来源:筏尚旅游网
第4O卷第3期 2013年3月 光电工程 Opto—Electronic Engineering Vo1.4O.No.3 March,2013 文章编号:1003—5OLX(2013)03—0007—07 Maximum Likelihood Estimation Algorithm Combining with Initial Value Selection for Passive Localization CHEN Jinguang 1,2 1 MA Lili (1.School ofCompu ̄rScience,Xi'an Polytechnic University,XJ'an 710048,China; 2.SchoolofElectronicEngineering,XidianUniversity,Xi'an 710071,China) Abstract:The problems of conventional Maximum Likelihood Estimation(MLE)algorithm are addressed for passive localization and an improved MLE passive localization algorithm is presented.Firstly,we estimate an initial target position using least square method.Moreover,in order to adapt the measurement error,the square root of the diference between the estimated position and the sensor’s position is used as the approximate covariance matrix for measurement error.Then,a weighted least square formula is employed to estimate a new position.Finally,we regard the estimation value as the necessary initial value,and employ the conventional MLE to calculate the final results.The improved algorithm have some advantages,i.e.,it does not need to set 811 initial target position,its localization results dO not diverge easily,its computational complexity is lOW,and it has the same level in accuracy as that of conventional MLE. Experimental results showthatthe proposedalgorithmis effective. Key words:passive localization;target localization;maximum likelihood estimation;data fusion CLC number:TP391 Document code:A do I:10.3969 ̄.issn.1003-501X.2013.03.002 结合初值选取的最大似然估计被动定位算法 陈金广 ,马丽丽 (1.西安工程大学计算机科学学院,西安710048, 2.西安电子科技大学电子工程学院,西安710071) 摘要:针对传统最大似然被动定位算法(MLE)在定位过程中需要设定初始位置的问题,提出了一种改进的MLE 定位算法。首先采用最小二乘法计算目标初始位置。此外,为了适应量测误差,将其与传感器位置之间的差方根 作为传感器测量误差的近似加权矩阵,再使用加权最小二乘公式,估计新的目标位置。最后,将该估计值作为初 始值,使用传统MLE算法获得最终定位结果。改进算法无需设定初始目标位置,运算过程不易发散,时间复杂 度不高,取得的定位精度和传统MLE算法相同。仿真结果表明了改进算法的有效性。 关键词:被动定位;目标定位;最大似然估计;数据融合 中图分类号:TP391 文献标志码:A 0 IntrOductiOn This work addresses the problem of bearing—only localization of a single target using the synchronous measurements from multiple sensors in 3-dimensional space.Typical sensors are sonar and passive radar,which only accept signals radiating from targets.and do not detect the opposing targets actively.Because passive sensors re anot easy to be found in the battlefield,they attracted much interest of the researchers in the past[1-3]. 收稿日期:2012—11—06l收到修改稿日期:2013—01-05 基金项目:国家自然科学基金项目(612011l8)#陕西省教育厅科研计划项目(12JKO529) 作者简介:陈金广(1977一),男(汉族),河南南阳人。副教授,博士,主要研究工作是信息融合、目标跟踪。E-mail:xacjg@163.com。 http://www.gdgc.ac。cn 8 光电工程 2013年3月 Most researchers took care of passive localization algorithms in the past and numerous solutions have been proposed【lJ.There are O ̄hogonal Vector Estimator(OVE),Pseudo—Linear Estimator(PLE)and Maximum Likelihood Estimator(MLE),etc.In【4],a closed-form 3-dimensional passive local ̄ation algorithm has been developed based on the method of instrumental variables,which is called Weighted Instrumental Variables(wry), and the error performance is better than OVE and PLE,but is worse than MLE.In fact,if there are some constraints in he tpassive localization system,it can be helpful to promote he tperformance.Bishop et a1.used the relationships between the measured parameters and their corresponding errors to derive a geometric cons ̄aint equation,and used he toptimization method to solve the optimal problem and to get a modiied fmeasurement,and then the localization results were enhanced,which is called Geometrically Constrained Lest Square algoritm h(GCLS)I5-7]However,it may be possible that most cross points do not overlap because the results are the local optimal or approximation optima1.Not only the sensor’s performance but also the radial range influences the target localization error,so these wo tfactors are considered to improve he tperformance in article[8].In fact,these methods based on GCLS have‘notaher problem,i.e.,heavy computational complexity.Hmam to0k advantage of the geometric consraitnts introduced by the uniform rotating motion of the antenna main beam as it sweeps across a number of separate receivers,and promoted the error performance .In literature【1 0],the convergence of iterative methods for bearings-only tracking was studied,and a geometric and niuifed framework was developed. Beside the above methods of passive localization,there are many works on passive rackitng[11-14]In he passitve .rtacking,one deal wih ta motion target at a sequence time series,not a stationary target.In the passive localization algoritms menthioned above,MLE get the best error performance.Meanwhile,its time eficifency is better than GCLS in a great dea1.However,the MLE algorihm tneed he ttarget’s initial position,if the initial value has a large error,it will divergence in he tprocess of iteration l’ .In this paper,we present an algorithm based Oll the MLE。 First,we get a good initial artget position,and hen the tconventional MLE is employed.As a result,we obtain a good performance in estimation error,and the results do not diverge.Experimental results show hatt the proposed algorithm owns these advantages. Problem Formulation In the passive localization of 3一dimensional space,passive sensors achieve angles of the target’s position,i.e., azimuth and elevation angles,which can be represented in Fig.1.From Fig.1,the position of sensors is represented as =[ (后), (七), (Ji})r,k=1,2,3,and P=(p ,Py,P:)denotes the position of the target.The relationship between P and is P:rk+sk where sk represent the distance from sensor to trget.The ameasurements of the sensors are( , ),k=1,2,3.Subscript k appeared above is the index of sensors.In general,we often study the case that there are three passive sensors,because he tother cases can be derived easily from he tstudy.From he tknowledge of geometry nd aritgonometry,we have =arc咖 一=(1) (2) sin IIs 『l[cos ̄ ̄cos#k cosek sin sinSk] (3) Because the measurements are always imprecise,and Gaussian disritbution are used to describe the measurement error.The relationship between ideal angles( , )and noise-corrupted nglaes( , )is given = 十 (4) : 十n http://www。gdgc.ac.ca (5) 第40卷第3期 陈金广,等:结合初值选取的最大似然估计被动定位算法 9 where and . are the measurement noises and they are both object to Gaussian distribution with variance 00 2 and The aim ofthe passive localization is to evaluate the target position more precisely by using sensors’positionmeasrementus nd ahe tvariance of every sensor. , Fig.1 Illustration for passive localization According to the optimal maximum likelihood localization,the position of a target is obtained from the maximization of the joint probability densiy tfunction of the angle measurements( , ).Let e(p)=[ 一 (p),…, 一 (p), 一破(p),…, 一 (p)】 (6) K=diag([0 ̄,…,0-2 , ,…0-2 】) And define the object function as arg m ine (p e(p) .P he Tproblem ofpassive localization becomes a typical optimal problem. (7) (8)  ’2 The Proposed Algorithm In the conventional maximum likelihood method,Gauss-Newton method Can be used to get tl1e solution of equation(8),and he titerative formula is +。= 一(. 置 ) ~e(b ), i=0,1,… (9) where Ji is the 2Ⅳ×3 Jacobian of e(p、wih restpect to P.Subsequently,we derive the Jacobian matrix . According to the defniition of Jacobian matrix (p)/a(p),we derive the items in which contained measurements ofsensor 1,and he stoltiuons ofother items are easy to get in me same way. For the azimuth nglae ofsensor 1,we have 一 = c 一 c = lI l 1、 y 印 II I 1Ⅲ = ! :一旦鱼 :一 ̄3arctanSA .:0 : 印: : (I2) Forthe elevation angle ofsensor 1,wehave !亟= ! :一旦!l :! 垒! !! ! ! 2 Ij http://www.gdgc.ac。cn (13) 10 光电工程 2013年3月 !亟二亟! :一旦鱼 :—sin ̄i(p)—sinOj(p) 印 { IS I1 (14) 1 二垒! 2:一旦 :二! !:垒! 2 : (15) When equations(1O)一(15)are substituted into the Jocabian matrix ̄e(p)/Op and evaluated at P:bi,we canget sin( (p))IIs II;) ; Ji= sin(ON(p)) cos( ( ))IIsI1 ̄; 0 ; cos(G(p)) 瞄0 ~cost(v) (16) in ̄(p)sinOl(p)IIs sin ̄l(p)cos01(P)lIs lr ssin (p)cos (p)II s l广 sin#N(P)sin0N(P)IIsll~一COS (p)IIsI巴 where A =1=[l 1si n:i ON 二一 jc os:G。]} ,bl=[ l sin0三 ̄’:,一 (c os G ]s Ⅳ, )]l} .4 =ATb。 (18) Equation(17)callbewriten as According to the method of instrumental variables,(1 8)call be modiifed to 4 =GTbl (19) canbe rewritten where G1 is the instrumental Variable matirx.Assume a matrix W is nonsingular,formula f 1 9) as W~A = W (2O) And the target estimation iS =( W 4) W (21) For convenience,we construct G1 using the approach in【151.First,we calculate the angle estimation via target position from(18),i.e., :arctan ,尼:1,…,Ⅳ (22) P 一 K Because the accurate value of Gl is unknown in advance,an approximate value is calculated by 4 in formula(17)with = .In order to adapt the measurement error,let W=diag([d ̄,…, 】),where :l1 一Sxy(七)ll is the range rfom sensor to target.In fact,assume three sensors are not on one line,three lines of azimuth angles from the sensors to the target direction will construct a triangle.The estimation of target point will be within the triangle and near to the real trgeta position.This can lead to a good initial position,and make http:llwww。gdgc.ac cn 第40卷第3期 陈金广,等:结合初值选取的最大似然估计被动定位算法 ll MLE algorithm convergence.In 3-dimentional space,z coordinate is calculated by 专喜 ㈣+l 七1)ltna ̄) 2-dimensional space is easy to obtain from he tabove derivation,then it is omitted. (23) Until now,we get the modified MLE in 3-dimensional space,and the corresponding algorithm in 3 Experimental Results In 3-dimensional space,assume the positions of sensors are(0,0,0),(200,200,0),and(1 00,0,0),and the position of the target is(250,250,1 00).The target was observed for 1 00 times,and diferent algorithms are used to estimate the position of he targett.Then the estimation RMS is given as (24) where(bi, :, )denotes the estimation position of het traget at i ht time,and M is hte whole observation times. Consider for the diference of sensor’s measurement performance,we give 6 diferent groups for the measurement standard error,which are shown in Table 1.We give the measurement stndarad error of three azimuth angles and three elevation angles,and divide them into 6 groups.Group 1 represents there are six large values which equal to each other Group 6 represents there are six small values which equal to each other.Group 2 to 5 show diferent amount errors are combined in diferent ways.Under these diferent combinations,we calculate the RMS and give them in Table 2.The initial position of target is well selected to maintain the conventional MLE algoritm converhgence.From the results,it is shown hatt he ltatest algorithm GCLS is a little better than PLE and WIV,but it is worse than the proposed algorithm.In fact,our algorithm obtains the lowest RMS,and displays its best performance in estimation error. When we use conventional MLE,an initial argett position has to be given at ifrst.In the simulation,assume thatthe estimatedinitialvaluefortargetposition(Xo,Yo,z0)subjectstomulti-variableGaussiandistribution,i.e., (px0 Pyo'P ) ~Ⅳ(( ,P ,P:)T diag([o' ̄, 2, 】)) and(Px,P ,Pz)is the real position of the artget.For convenience,let (25) where o-x,Oy,and o-z are the standard error of he tinitial target position between he hypotthesis and he treal, = = = .The conventional MLE and our method re arun by using diferent initial target positions,and the results are shown in Table 3.These data show that our method cannot diverge in he tsimulation,and he conventtional MLE become diverge more easily as he sttandard error of initial target position being lrger.Of acourse,the easron is that our method does not need hese tinaccurate values but a more accurate estimation initial value in he tifrst part of the algorithm. We calculate the average computational time and list them in Table 4.From the results,GCLS has the heaviest computational burden,and the proposed algorihm tis smaller than hatt of GCLS,although it is lrger athan hat tof he tother’s(i.e.,PLE and WIV). In order to study the localization accuracy under diferent bearing measurement error levels,we use three passive sensors in two—dimension space to estimate the position of one target.The target’s real position is (100,100),and the sensors’positions are(0,0),(1OO,30),and(250,0),respectively.The range of variance of measurement noise is set from 0.5 to 5.We repeat running diferent localization algoritmsh for 1 000 times,and compute RMS ofdiferent algorithms at diferent bearing measurement error.The results are shown in Fig 2.As we see from the Figure,no matter what the bearing measrementu error is,the conclusion is the same,i.e.,the proposed algorithm has the lowest RMS,the PLE has he thighest RMS,and he otther’s performance is between the proposed algorithm and PLE.In hits experiment,considering he tresults of the conventional MLE are he tsame http://www。gdgc.ac.cn l2 光电工程 2013年3月 as those ofthe proposed algorithm,SO they are omiRed. Table 1 Angle standard error setting of sensors(unit:degree) g0— z!一§一 0∞吾H Table 2 Passive lcoalization RMS under diferent angle standard error groups Table 3 Ratio ofiterative divergence using diferent initial target position(%) Table 4 Computational time S 6 2 0.5 1.5 2.5 3.5 4.5 Bearing measurement error/degree Fig.2 RaMS of localization of diferem passive localization algorithms 4 Conelusion Localization and tracking via passive sensors are attractive in the field of information fusion.We proposed one modiifed MLE algorithm to deal wiht the divergence in calculation process of iteration.The idea is very direct, it is that we choose a good initial target position,and it will make the running process convergence more easily. Because of the outperformance of MLE in estimation error,the proposed is useful in the real application.We only http:I/www gdgc.ac。cn 第40卷第3期 陈金广,等:结合初值选取的最大似然估计被动定位算法 1 3 derive the case in 3-dimensional space,and the case in 2-dimensional space is easily obtained from it.If the number ofpassive sensors is larger than three,we can reduce to multiple groups and each group owns trhee sensors,and it can be handled by this decomposed style. References: [1]Nardone SC,LindgrenAG,GongKF.Fundamentalproperties andperformance ofconventionalbearings—onlytargetmotion analysis[J].IEEE Transactions on Automatic Control(S0018-9286),1984,29(9):775—787. [2]Don J T.Statistical theory of passive location systems[J].IEEE Transacitons on Aerospace and Electronic Systems(S0018-9251),1984,20(2):183—198. 【3】 Gavish M,Weiss A J.Performance ofanalysis ofbearing-only target location algorithms【J].IEEE rTansactions on Aerospace and Electronic Systems(SO018-9251),1992,28(3):817-827. [4】Dogangay K,Ibal G Instrumental variable estimator for 3D bearings—only emitter localization【c]//Proceedings of the 2 International Conference on Intelligent Sensors,Sensor Networks and Information Processing,Melbourne,Australia, 2005:63—68. 【5】 Bishop AN,Anderson B D O,Fidan B,et a1.Bearing—only localization using geometircally constrained optimization【J】.IEEE rTansactions on Aerospace and Electronic Systems(SO018—9251),2009,45(1):308—320. 【6】 Bishop A N,Fidan B,Anderson B D 0,et a1.Optimality analYsis ofsensor-target geometries in passive lcoaliaztion:Part 1-Beaitng only localization[C】//Proceedings of the 3 International Conference on Intelligent Sensors,Sensor Networks,and Information Processing,Melbourne,Australia,2007:7-12. [7】 Bishop A N,Fidan B,Anderson B D 0,et a1.Optimaliyt analysis of sensor-target geometires in passive localization:Part 2-Time-of-arrival based localization【c]//Proceedings of the 3州International Conference on Intelligent Sensors,Sensor Networks,and Information Processing,Melbourne,Australia,2007:13—18. 【8]陈金广,李洁,高新波.改进的基于几何约束的加权被动定位算法【J].光电工程,2010,37(2):16—21. CHEN Jinguang,LI Jie,GAO Xinbo.Modiifed Weighted Passive Localization Algorithm Based on Geometric Constraint[J】. Opto-Eieetronic Engineering,2010,37(2):16—21. [9 9】Hmam H.Scan-based emitter passive localization[J】.IEEE Transactions on Aerospace and Electronic Systems(S0018-9251),2007,43(1):36—54. 【10】Cadre J P Le,Jauffret C.On the convergence of iterative methods for beatings-only tracking[J】.IEEE Transactions on Aerospace and Electronic Systems(SO018—9251),1999,35(3):801—818. 【1 1]Nardone S C,Graham M L.A closed-ofrm solution to bearings-only target motion analysis[J].IEEE Journal of Oceanic Engineering(S0364・9059),1997,22(1):168~178. [12】Dogancay K.Bias compensation for the bearings・only pseudolinear target track estimator【J].IEEE rTansactions on Signal Processing(S1053,587X),2006,54(1):59—68. [13】Dogancay K.Relationship between geometirc transltaions and TLS estimation bias in beatings-only target localization[J】. IEEE rTansacitons on Signal Processing(S1053-587x),2008,56(3):1005—1017. [14】 Zhan R H, Wan J W Iterated unscented Kalman iflter ofr passive target tracking[J】.IEEE rTansactions on Aerospace and Electronic Systems(SO018-9251),2007,43(3):l155—1163. [15]Doganqay K。Passive emiter localiaztion using weihgted instrumental Beraing RMS(degrees)vairables[J】.Signal Processing(S0165—1684),2004,84(3):487-497. http://www。gdgc。ac.cn 

因篇幅问题不能全部显示,请点此查看更多更全内容

Copyright © 2019- efsc.cn 版权所有 赣ICP备2024042792号-1

违法及侵权请联系:TEL:199 1889 7713 E-MAIL:2724546146@qq.com

本站由北京市万商天勤律师事务所王兴未律师提供法律服务