Laser Measurement System based maneuvering Target tracking formulated by Adaptive Competitive Neural Networks

Transcription

1 ADAPIVE 00 : he Second Internatonal Conference on Adaptve and Self-Adaptve Systems and Applcatons Laser Measurement System based maneuverng arget trackng formulated by Adaptve Compettve Neural Networks Lokukaluge P. Perera Centre for Marne echnology and Engneerng echncal Unversty of Lsbon, Insttuto Superor écnco, Lsbon, Portugal prasad.perera@mar.st.utl.pt Carlos Guedes Soares Centre for Marne echnology and Engneerng echncal Unversty of Lsbon, Insttuto Superor écnco, Lsbon, Portugal guedess@mar.st.utl.pt Abstract o mprove safety and securty ssues, maneuverng target detecton and trackng are mportant facltes for navgaton systems. herefore, conventonal navgaton systems are equpped wth adar-based systems for the same purpose. However, adar systems suffer some practcal problems that are assocated wth the targets n close quarter navgaton. Furthermore, adar sngles attenuate wth dstance, weather (e. ran) and sea condtons, where the target tackng performances are degraded. herefore, a Laser Measurement System (LMS) s proposed n ths study to overcome the problems faced by the conventonal adar systems at close quarter navgaton as well as bad weather and envronmental condtons. Furthermore, capabltes of a LMS to measure accurate dstance n close proxmty as well as to observe the shape and sze of the target are llustrated. In ths study, each target s approxmated by a cluster of data ponts rather than a sngle pont target that s the man contrbuton n ths paper. he adaptve Neural Network approach s proposed as a method of trackng maneuverng targets that are represented by clusters of data ponts. Successful smulaton and expermental results of target detecton and trackng that are tested on a expermental platform, SICK LMS, are also presented n ths paper. Keywords- Laser Measurement System, Compettve Neural Networks, arget trackng, Data Ponts rackng I. INODUCION Maneuverng target detecton and trackng capabltes are mportant facltes for a navgaton system to mprove safety, securty and survvablty durng ts voyage. he conventonal navgaton systems are equpped wth adarbased systems to facltate maneuverng targets and obstacles detecton and trackng. However, adar-based systems are suffered by practcal problems especally wth detecton and trackng of targets n close quarter navgaton. Furthermore, adar sngles attenuate wth the dstance, weather (e. ran) and sea condtons, where the target tackng performances are degraded []. herefore, under the dstance, weather and envronmental condtons, the frequent calbratons for adar systems are requred to mprove ts accuracy []. Furthermore, adar-based systems are lmted n evaluaton of accurate range, bearng, shape and sze of targets n long dstance as well as close quarter navgaton. he unsuccessful target detecton and trackng n close quarter navgaton could affect on naccuracy of the dstance Fgure. LMS Expermental Setup measurements wth respect to the targets and obstacles n the envronment. herefore, the errors n dstance measurements can eventually affect on naccurate collson rsk evaluatons and wrong navgatonal decsons. hs study proposes, a Laser Measurement System (LMS) that s ntegrated wth an adaptve Neural Network algorthm for maneuverng target detecton and trackng n close quarter navgaton. Hence, these facltes can be formulated n navgaton systems for accurate collson rsk evaluatons and better maneuverng decsons. he proposed LMS expermental platform n ths study s presented n Fgure. As presented n the fgure, the expermental setup conssts of a Laptop computer, where the adaptve Neural Network algorthm s mplemented, SICK LMS, whch s the target detecton sensor, and a movng target (e. movng robot). Further detals on ths system are presented on Secton V of ths paper. he work presented n ths study s a part of ongong effort to formulate an Intellgent Collson Avodance System (ICAS) n ocean navgaton as descrbed n [3] and [4]. herefore, two-dmensonal target trackng formaton wth respect to ocean navgatonal condtons s consdered n ths study. he organzaton of ths paper as follows: An overvew of recent developments n a LMS s presented n Secton II. he proposed adaptve Neural Network approach s Copyrght (c) IAIA, 00 ISBN:

2 ADAPIVE 00 : he Second Internatonal Conference on Adaptve and Self-Adaptve Systems and Applcatons presented n Secton III. he computatonal smulatons are presented n Secton IV. he expermental results generated by the LMS platform are presented n Secton V. Fnally, a bref concluson s presented n Secton VI of ths paper. II. ECEN DEVELOPMENS IN LMSS AND AGE ACKING A. Laser Measurement System (LMS) People and vessel/vehcles detecton and trackng are popular expermental applcatons n recent LMSs. he people trackng system, based on Laser range data and usng a mult-hypothess Leg-racker, n-cooperated wth the Kalman flter wth a constant velocty based model, s proposed by [5]. Almost all research and commercal applcatons n the LMS area are D due to the lmtatons n current LMS sensor technology. However, 3D LMS applcatons are also been proposed n some studes by crculatng the sensor n a thrd dmensonal axs. he D Laser-based obstacle moton trackng and predctng n a dynamc unconstraned envronment usng the Kalman flter [6], and the Partcle Flters and Probablstc Data Assocaton [7] are presented n respectve studes. he mplementaton of a LMS for detecton and classfcaton of 3D movng objects s llustrated by [8] and [9]. Furthermore, an expermental evaluaton of a LMS for tackng people by a moble platform s presented by [0] and classfcatons of people by a moble robot s presented by []. he ntegraton of a LMS and a vson system can use for successful obstacle trackng and classfcaton due to ts capabltes of capturng the comprehensve detals of the targets as well as the envronment. he combned Laser and vson based approach for smultaneous detecton and trackng of multple pedestrans based on the Bayesan method s proposed by []. Furthermore, several ndustral applcatons of LMSs can be observed n recent lterature: As a navgatonal ad for a truck-traler combnaton vehcle system [3], a collsons warnng system for a transt bus [4], a safe drvng ad system for a car drvng n polluted envronment [5], an obstacle avodance system for a car navgaton system [6] and an obstacle detecton system for an off-road vehcle [7] are presented n respectve studes. However, the most model based LMS and other sensor based target-tackng algorthms could not facltate the dmensonal based target trackng and a target s approxmated to a sngle data pont. herefore, n ths study, ths concept s further elaborated to formulate a target as a cluster of data ponts durng ts trackng process, whch s the man contrbutons n ths study. B. Detecton and rackng of Movng Objects Detecton and rackng of Movng Objects (DMO) s one of the man research areas that was developed towards maneuverng target trackng. he man functonaltes of the DMO can be dvded nto three sectons [8]: Scan unt, arget Classfcaton unt and arget rackng & Behavor Predcton unt. he man objectve of the Scan unt s to formulate geometrcal clusters, where a cluster defned as a set of measured data ponts that could belong to a same object or multple objects, of data ponts, lnes and arcs wth respect to targets and obstacles n the envronment. However, ths could be generated by the sensors (e. adar and LMSs) n the target trackng system. he segmentaton of data clusters by a geometrcal method s proposed by [8]. Inscrbed angle varatons and recursve lne-fttng methods for lnes and arc/crcles detecton by LMS data are proposed by [9]. However, specal consderatons for the jonts and break ponts should be consdered durng ts segmentaton of data clusters, n ths method. herefore, the proposed adaptve Neural Network approach can overcome the falures can occur n the segmentaton process of data clusters due to varyng geometrcal constrans. he man objectve n the arget Classfcaton unt s to formulate the Segment-Objects correspondence. he correspondence manly classfed nto geometrcal fgures lke crcles or polygons. However, four classfcaton methods are proposed by recent studes [0] for ths prupose: Features to Features, Ponts to Features, Ponts to Ponts and Combnatons. Furthermore, the dentfcatons of geometrcal fgures and features are successful done by the Neural Network approach n recent studes []. Even though the proposed adaptve Neural Network approach s lmted for detecton and trackng of clusters of data ponts, ths method can further develop for dentfcaton of geometrcal fgures and features of the targets. Fnally, the arget rackng and Behavor Predcton unt s proposed to estmate target current states and to predct future navgaton trajectores. he EKF based system states estmaton and maneuverng trajectory predcton for ocean vessel navgaton s proposed n []. However, ths area s beyond the scope of ths paper. III. ADAPIVE NEUAL NEWOK BASED DEECION & ACKING A. LMS Scan & Data Collecton he LMS expermental platform s presented n Fgure. he LMS sensor generates respectve range r, and bearng &, values n polar coordnates as presented n the fgure. he accumulated data clusters of range and bearng values that represent complete envronmental condtons, ncludng the statonary and movng targets, at the k-th tme nstant n polar coordnates can be wrtten as: r ϑ [ r r... r ] [ ϑ ϑ... ϑ ] hen the range and bearng values n polar coordnates are converted nto Cartesan poston coordnates. he -th poston data pont n Cartesan coordnates [ x y ] can be formulated as: () Copyrght (c) IAIA, 00 ISBN:

3 ADAPIVE 00 : he Second Internatonal Conference on Adaptve and Self-Adaptve Systems and Applcatons x r cos y r sn ( ϑ ) ( ϑ ) () herefore, the -th poston data pont of the data cluster at the k-th tme nstant can only have two coordnates of [ x y ] that are measured by the sensor. However, these poston data ponts should be normalzed wth respect to the maxmum range of the sensor. he normalzaton requrements are further dscussed n sub-secton C of ths man secton. he normalzaton of the poston coordnates can be wrtten as: x y x y max max where max s the maxmum range of the LMS sensor. B. Artfcal Neural Networks he theoretcal foundaton of artfcal neurons s derved from bologcal concepts and theores n the bran and nervous system. An artfcal neuron has several nputs that correspond to the synapses of a bologcal neuron. An artfcal neuron has one output that s correspondng to the axon of a bologcal neuron. Each nput of a neuron s correspondng to a certan weght value that nfluences the correspondng sgnal over the neuron output. hs concept can formulate nto a transfer functon n an artfcal neuron. he transfer functon calculates sum of the net nput wth respect to the assgned weght values and compares that wth a certan threshold level to generate the neuron output [3]. he connecton of several neurons n a combnaton of seres and/or parallel formatons can recognze as a Neural Network. C. Compettve Neural Network he Compettve Neural Network (CNN) [4] ntegrated wth an adaptve learnng algorthm of the Instar ule s proposed n ths study for detecton and trackng of maneuverng targets. he CNN s traned to track movng data clusters by competng ts neurons, where a target s approxmated for a cluster of data ponts.. he structure of the CNN s presented n Fgure. As presented n the fgure, the CNN conssts of four unts: Scan unt (Data Ponts), Prototype vectors unt (W), Competton unt (C), and Feedback-loop (Instar ule). he nput to the CNN conssts of a accumulated poston vector p, x3. he prototypes vectors, W, Sx3, are stored as rows vectors n secton W, that are target trackng neurons of the CNN. he net nput n,, s the nput to the Competton unt, C, and a, S, s the output from the Competton unt, C, at the k-th tme nstant. Fnally, the feedback loop, assocated wth the Instar ule that s proposed to adjust the prototype neurons to contnue trackng of maneuverng targets. (3) ) Compettve Layers he Competton unt, C, conssts of a transfer functon that s used to generate competton among neurons. Hence, the proposed transfer functon can be wrtten as: a compet compet ( n ) ( W p ) where the compettve (compet) transfer functon can be further elaborated as: compet ( ) (4) for neron wth max n n (5) 0 all other nerons and the accumulated poston vector p [p p p ] s assocated wth the -th poston vector, p [x y z ] that represents the poston of x, y and z coordnates of a data cluster as descrbed prevously. However, only normalzed x and y poston coordnates are calculated from equaton (3). For a far competton among neurons, each poston vector, p, 3, should have a unt magntude condton. Hence, the poston value of z can be derved consderng a unt magntude condton as proposed prevously and can be formulated as: p x + y + z (6) Hence the coordnate z can be calculated consderng equaton (6) that gves a unt magntude condton for each data pont n the data culster. he coordnate z can be calculated as: z x y (7) Fgure. Compettve Neural Network One should note that ths mplementaton can nterpret as a transformaton of D space poston coordnates n the sensor range nto 3D space poston coordnates wth a unt magntude condton. herefore, ntally x and y coordnates are normalzed consderng the sensor maxmum range (see equaton (3)) that s an essental requrement of the neural competton. Copyrght (c) IAIA, 00 ISBN:

4 ADAPIVE 00 : he Second Internatonal Conference on Adaptve and Self-Adaptve Systems and Applcatons Furthermore, the net nput, n, can calculate from the scalar product between two vectors W and p as presented n equaton (4). hs scalar product between two poston vectors related to the dstance between poston vector, p, and each prototype vectors w j, where W [w w w S ]. A unt magntude condton for each prototype vector, w j, should also be consdered for far competton among neurons. Hence the j-th prototype vector, w j, magntude condton can be wrtten as: w j (8) In the Competton unt, C, (see Fgure ), the dstance between poston vector, p to each prototype vector w j s calculated. hs concept can further be elaborated as: w w n W p p M w S w w w S p p M p cos θ cos θ M cos θ S he -th net nput of scalar product between two vectors, n, s equal to cos( ), where s the angle between a poston vector, p, and a prototype vector, w j. However, the scalar product between two vectors, n, s the nput to the compettve transfer functon. herefore, the neuron, whose prototype vector s n the drecton closest to the respectve poston vector, p, s assgned output of and others are assgned 0 by the transfer functon as formulated n equaton (5). hs concept can further elaborate as a stuaton where the closet neuron gets excted by a data cluster and the excted neuron takes over all data ponts n the respectve data cluster. However, after wnnng the data cluster, the prototype vector of the respectve neuron should be mproved (should move further closer to the data cluster). hs contnues process conssts of two dfferent teraton loops. he frst teraton loop formulates a contnuous mechansm, where the wnng neuron contnuously gets closer to ts respectve data cluster. he second teraton loop formulates another contnuous mechansm, where the dynamc data clusters that are observed by the sensor at dfferent tme nstants are ntroduced nto the CNN. herefore, a capable learnng rule should be formulated to facltate proper update of the wnnng neurons wth respect to dfferent data cluster condtons. (9) ) Compettve Learnng Intally, the values of prototype vectors, W, n the CNN, are assumed to be unknown. herefore, the learnng rule s expected to calculate approprate values for the prototype vectors. hs concept s categorzed as unsupervsed learnng. When a compettve layer exctes a neuron that s closest to the data cluster, then the learnng rule wll use to modfy the approprate prototype vectors n the CNN to move close to the data cluster n ths process. he Instar ule s proposed n ths study as an unsupervsed learnng mechansm to modfy the approprate prototype vectors n the CNN. 3) Instar ule he Instar ule that s derved from the Hebb ule s llustrated n [4] s brefly dscussed n ths secton. he unsupervsed Hebb ule to update prototype vectors can be wrtten as: W W (k ) + αa p (0) where α s a learnng rate. However, a constant learnng rate could be a dsadvantage n the learnng process of a neural network, where t could affect on the error convergence rate. Even though the begnnng of a learnng process a hgher learnng rate s an advantage to the neural network, wth the error reducton t could be a dsadvantage. Hence, to mprove the Hebb ule, a weght decayng term that s proportonal to a and W(k-) s ntroduced. Equaton (0) wth a weght decayng term can be wrtten as: W ( k) W (k ) + αa p γa W (k ) () where ( s the decay rate. Furthermore, assumng ( ", equaton () can be wrtten as: ( p W (k ) ) W ( k) W (k ) + αa () Equaton () s called as the Instar ule that s proposed as an unsupervsed learnng rule for the CNN n ths study. IV. COMPUAIONAL IMPLEMENAION AND SIMULAIONS he computatonal smulaton of a mult-target trackng stuaton s presented n Fgure 3. he smulaton conssts of two movng targets that are presented by two clusters of data ponts. Furthermore, two prototype vectors are also assgned n ths smulaton to tack both data clusters. Copyrght (c) IAIA, 00 ISBN:

5 ADAPIVE 00 : he Second Internatonal Conference on Adaptve and Self-Adaptve Systems and Applcatons V. EXPEIMENAL PLAFOM AND SIMULAION ESULS Fgure 3. Computatonal Smulatons : Mult-arget rackng he target trackng algorthm, smulated n ths study, conssts of two man loops: the LMS target scannng loop and the CNN target trackng loop. he man objectve of the LMS target scannng loop s to scan the envronment and to observe the statonary and movng targets as an accumulated data cluster. hen ths nformaton wll transfer nto the CNN target tackng loop. he man objectve of the CNN target tackng loop s to adapt the CNN to track target maneuvers by updatng ts respectve prototype vectors. hs should be done by the proposed learnng rule. he two prototype vectors of the CNN are called as NN (Neural Network) racks and. As presented n the fgure, the NN acks and are presented by and respectvely. he ntal prototype vector values of the NN racks and can be any arbtrary values as presented n the ntal postons of the NN acks and (see Fgure 3). Fnally, both NN racks are adapted ts prototype vectors to track movng targets that represented by two clusters of data ponts. he trackng trajectores of the neurons are presented by rajectory NN racks and n the fgure. Furthermore, t s observed that NN racks and are fnally converged nto approxmate mean values of the respectve data clusters at each tme nstant. herefore, the mean poston values can be consdered as measurement postons of the targets at each tme nstant and that can be used for further analyss of target state estmatons and trajectory predctons []. A. Laser Measurement System he expermental platform s presented n Fgure. As presented n the fgure, the hardware secton manly conssts of SICK Laser Measurement System (LMS). he SICK LMS s an actve poston measurement unt that operates by measurng the tme of flght of Laser lght pulses, where Laser beam pulses are emtted by the sensor and reflected due to the objects n the envronment [5]. However, the LMS s desgned to scan D space and to collect range and bearng data of the targets that are located n the envronment. he SICK LMS model of LMS, that s desgned for marne envronment s used n ths study. hs sensor s capable of scannng bearng angle of 80 0 wth accuracy feld vews wth 75 Hz scannng frequency. he operatng range of 8 m wth the mnmum lnear and angular resoluton of mm and are ntally programmed nto the sensor. he SICK LMS data communcaton s facltated by S- 3 wth the speed of 9.6 kbd. Furthermore, the expermental platform conssts of a Laptop computer wth Wndows operatng system and a power supply unt to power the LMS sensor. he Laptop computer s equpped wth the S-3 connecton to communcate wth the LMS sensor. B. Software Archtecture he software archtecture that s used n ths study manly conssts of LABVIEW eal-tme platform. Further MALAB toolbox of neural networks s also ntegrated nto the LABVIEW eal-tme platform for mplementaton of the CNN. C. Expermental esults he expermental result of a statonary and movng target trackng stuaton n eal-tme envronment s presented n Fgure 4. he measurements are noted n mlle-meters (mm) of SI unts n the fgure. he two targets, a statonary target and a movng target, are consdered n ths experment. he statonary target s located n the mddle of the fgure and the movng target s crculatng around the statonary target as presented n the fgure. he movng target s presented by a movng cluster of data ponts. he CNN conssts of two NN racks to track both targets as presented n the fgure. Furthermore, the statonary target s montored by the NN rack and movng target s montored by the NN rack are also presented n the fgure. However, the CNN trackng regon s lmted by upper, lower, left and rght boundary values 9000 (mm), 0 (mm), - 00 (mm) and 00 (mm), respectvely. As presented n the fgure, NN rack s followng each pont n the data cluster of the movng target alone ts maneuverng trajectory. As a concluson, the expermental results have shown that the NN rack and rack are successfully trackng both statonary and movng targets as observed n the smulatons. Copyrght (c) IAIA, 00 ISBN:

6 ADAPIVE 00 : he Second Internatonal Conference on Adaptve and Self-Adaptve Systems and Applcatons compromse could affect on the stablty of the prototype vectors. he second, the statonary and movng target trackng under complex envronmental condtons: several neurons can track dfferent parts of the same target and one neuron can track several targets n close range navgaton. hs s another challenge that s faced n ths CNN approach. However, ths stuaton can be solved by selectng proper number of neurons wth respect to the targets n the envronment.. Furthermore, the Laser-based CNN approach can further develop for dentfcaton and classfcaton of maneuverng targets where the Neural Network approach s extensve mplemented on statstcal pattern recognton [6]. Furthermore, ntegraton of mage based (e. Infra-red) facltes could mprove the target detecton and trackng process [7]. Hence, the ntegraton of llustrated features (e. dentfcaton and classfcaton) nto target detecton and trackng are proposed as future work n ths study. ACKNOWLEDGMEN Fgure 4. Expermental esults: Statonary and Movng arget rackng However, the data ponts beyond the lmts of upper, lower, left and rght boundary values are gnored n ths analyss. hese data ponts are located beyond the smulaton lmts, whch represent other statonary and movng objects n the expermental envronment. VI. CONCLUSION he dmenson based target detecton and tackng are man contrbutons n ths study, where the most of target trackng methods are smulated for a data pont or an approxmated small data cluster based targets. Furthermore, one of the popular machne learnng applcatons of an adaptve Neural Network assocated wth an unsupervsed learnng algorthm, the CNN, s mplemented and successful smulaton and expermental results are obtaned n ths study. Even though the Neural Network applcatons are extensvely used for recognton of statonary data patterns, movng data clusters can also be detected and tracked by the proposed method. Even though, the proposed CNN behave as an effectve adaptve network for trackng targets, t also been affected by some nherted problems. he frst, the selecton of a learnng rate should be compromsed wth the target trackng speed. However, ths hs work has been made wthn the project Methodology for shps maneuverablty tests wth self-propelled models, whch s beng funded by the Portuguese Foundaton for Scence and echnology (Fundação para a Cênca e ecnologa) under contract PDC/A/7433 /006. he research work of the frst author has been supported by a Doctoral Fellowshp of the Portuguese Foundaton for Scence and echnology (Fundação para a Cênca e ecnologa) under contract SFH/BD/4670/008. EFEENCES [] S. E. Gangrande and A. V. yzhkov, Calbraton of dualpolarzaton radar n the presence of partal beam blockage, Journal of Atmospherc and Oceanc echnology, vol., pp , 004. []. Naranjo, adar revsted, Ocean Navgator Onlne, [retreved: August, 00]. [3] L. P. Perera, J. P. Carvalho, and C. Guedes Soares, Decson makng system for the collson avodance of marne vessel navgaton based on COLEG rules and regulatons, n Proceedngs of 3th Congress of Internatonal Martme Assocaton of Medterranean, Istanbul, urkey, 009, pp. 8. [4] L. P. Perera, J. P. Carvalho, and C. Guedes Soares, Smooth transton between fuzzy regons to overcome falures n fuzzy membershp functons of decsons n collson avodance of ocean navgaton, n Proceedngs of 5th Mn-EUO Conference on Uncertanty and obustness n Plannng and Decson Makng, Combra, Portugal, 00, pp -8. [5] K. O. Arras, S. Grzonka, M. Luber, and W. Burgard, Effcent people trackng n laser range data usng a mult-hypothess legtracker wth adaptve occluson probabltes, n Proceedngs of the 008 IEEE Internatonal Conference on obotcs and Automatons, CA, USA, 008, pp [6] M. Berker,. Hall, S. Kolsk, K. Macek,. Segwart, and B. Jensen, d laser-based probablstc moton trackng n urban-lke envronments, Journal of the Brazlan Socety of Mechancal Scences and Engneerng, vol. 3, no., pp , 009. Copyrght (c) IAIA, 00 ISBN:

7 ADAPIVE 00 : he Second Internatonal Conference on Adaptve and Self-Adaptve Systems and Applcatons [7] A. Almeda, J. Almeda, and. Araujo, eal-tme trackng of multple movng objects usng partcle flters and probablstc data assocaton, Automatka, vol. 46, no. -, pp , 005. [8] A. Lourenco, P. Fretas, M. I. bero, and J. S. Marques, Detecton and classfcaton of 3d movng objects, n Proceedngs of the 0 th Medterranean Conference on Control and Automaton, Lsboa, Portugal,, 00. [9] G. Gallagher, S. Srnvasa, and J. Andrew, GAMO : A generalzed approach to trackng movable objects, n IEEE Internatonal Conference on obotcs and Automatons, 009, pp [0] M. Lnstrom and J. O. Eklundh, Detecton and trackng movng objects from a moble platform usng a laser range scanner, n Proceedngs of the 00 IEEE/SJ Internatonal Conference on Intellgent obots and Systems, Hawa, USA, 00, pp [] M. Luber, K. O. Arras, C. Plagemann, and W. Burgard, Classfyng dynamc objects : An unsupervsed learnng approach, n Proceedngs of the obotcs: Scence and Systems IV, Zurch, Swtzerland, 008, pp [] X. Song, J. Ch, H. Zhao, and H. Zha, A bayesan approach : Fuson of laser and vson for multple pedestrans trackng, Internatonal Journal of Advanced Computer Engneerng, vol. 3, no., pp. 9, 008. [3]. Stahn, G. Heserch, and A. Stopp, Laser scanner-based navgaton for commercal vehcles, n Proceedngs of the 007 IEEE Intellgent Vehcles Symposum, Istanbul, urkey, 007, pp [4]. A. Maclachlan and C. Mertz, rackng of movng objects from a movng vehcle usng a scannng laser rangefnder, n Proceedng of the IEEE Intellgent ransportaton Systems Conference, oronto, Canada, 006, pp [5] H. Hrose, K. Katabra, H. Zhao, and. Shbasak, A study for safe drvng usng a laser scanner ntegraton of the sensor on motonless objects and movng objects, n Proceedngs of the Asan Assocaton on emote Sensng, Colombo, Sr Lanka, 008. [6]. C. Ng, J. I. Ibanez-guzman, J. Shen, and Z. Gong, Vehcle followng wth obstacle avodance capabltes n natural envronments, n Proceedngs of Internatonal Conference on obotcs and Automaton, 004, pp [7] C. S. Dma, N. Vandapel, and M. Hebert, Sensor and classfer fuson for outdoor obstacle detecton: An applcaton of data fuson to autonomous off-road navgaton, n Proceedngs of the 3nd Appled Imagery Pattern ecognton Workshop, 003, pp [8] A. Mendes, L. C. Bento, and U. Nunes, Mult-target detecton and trackng wth a lasers canner, n 004 IEEE Intellgent Vehcles Symposum, Parma, Italy, 004, pp [9] J. Xaver, M. Pacheco, D. Castro, A. uano, and U. Nunes, Fast lne arc/crcle and leg detecton from laser scan data n a player drver, n obotcs and Automaton, 005, Proceedngs of the 005 IEEE Internatonal Conference on, 005, pp [0] C. Wang and C. horpe, Smultaneous localzaton wth detecton and trackng of movng objects, n IEEE nt. Conf. on obotcs and Automaton, Washngton DC, 00, pp [] Z. Je-yu, A novel recurrent neura network for face recognton, Journal of Software, vol., no. 8, pp. 8 39, 00. [] L. P. Perera and C. Guedes Soares, Ocean vessel trajectory estmaton and predcton based on extended kalman flter, n Proc. nd Internatonal Conferrence on Adaptve and Self-adaptve Systems and Applcatons, Lsbon, Portugal, 00, (In Prnt). [3] M. Crstea, A. Dnu, J. Khor, and M. McCormck, Eds., Neural and Fuzzy Logc Control of Drves and Power Systems, st ed. MA, USA: Elsever Scence, 00. [4] M.. Hagan, H. B. Demuth, and M. H. Beale, Eds., Neural Network Desgn. Boston: PWS Publshng, 996. [5] echncal Descrpton, LMS00///9 laser measurement systems., [retreved: August, 00]. [6] C. M. Bshop, Neural Networks for Pattern ecognton. Oxford, UK: Clarendon Press, 995. [7] X. Jpng, I. U. Haq, C. Je, D. Lhua and L. Zawen, Movng target detecton and trackng n FLI mage sequences based on thermal target modelng, n Proceedngs of the Internatonal Conference on Measurng echnology and Mechatroncs Automaton, 00, pp Copyrght (c) IAIA, 00 ISBN: