Cloud Based Localization for Mobile Robot in Outdoors

Transcription

1 Cloud Based Localizaion for Mobile Robo in Oudoors Xiaorui Zhu, Member, IEEE, Chunxin Qiu, Yulong Tao and Qi Jin Absrac Cloud Roboics is he applicaion of he cloud compuing concep o he robo. I uilizes modern cloud compuing infrasrucure o disribue compuing resources and daases. A cloud based localizaion echnique is proposed in his paper o allow he robo o idenify is locaion relaive o a road nework map in he cloud. The updae of he road nework map and he exracion of he robo-errain inclinaion model (RTI model) are running in he cloud. A paricle filer localizaion is achieved on he mobile robo based on he local RTI model sen from he cloud. Experimens were carried ou for validaion of he proposed cloud based localizaion echnique. Preliminary resuls show ha his mehod could be poenially applicable o long-erm auonomous. Index Terms Cloud roboics, localizaion, Mobile robo. L I. INTRODUCTION ocalizaion is an imporan problem for auonomous mobile robo. Tradiional roboic echnologies, however, have been limied by he inheren physical consrains especially for large-scale exploraions since all he compuaions have o be conduced in he onboard compuers/microchips of he robo ha have limied compuing capabiliies. Hence his paper proposes a cloud-based archiecure o achieve long-erm auonomous localizaion of mobile robos in oudoor environmens aking advanage of he powerful compuaion, sorage and oher shared resources of he cloud. In he las decade, many researchers have sared o focus on he robo localizaion in oudoor environmens. is a popular ool for localizaion in oudoor environmens[-2]. However, in some urban circumsances, he signals would be blocked resuling in degrading of he posiion esimaes. Laser range scanner is an opion for auonomous oudoors navigaion where he poin clouds were generaed for represenaion of he surroundings[3-9]. However, he con of he poin clouds represenaion was higher compuaional load [3]. The combinaion of poin reducion and kd-ree was proposed o reduce he compuaional load of poin clouds represenaion[4]. Some researchers used he occupancy grid map o divide he whole poin clouds ino a grid of cells wih he occupancy evidence inferred from sensors[5]. This research is currenly suppored by he NSF of China under NSFC Gran No Xiaorui Zhu is he corresponding auhor wih Harbin Insiue of Technology Shenzhen Graduae School, Shenzhen, Guangdong 58055, China ( xiaoruizhu@hisz.edu.cn). C. Qiu, Y. Tao and Q. Jin are wih he Harbin Insiue of Technology Shenzhen Graduae School, Shenzhen, Guangdong 58055, China (qiu_chun_xin@63.com). Since he represenaion of he enire space mus be sored in memory, even 2D evidence grids are large and expensive o copy. In order o reduce he processing and sorage requiremens, an ocree daa srucure was furher developed o finish a underwaer unnel exploraion proec using a simplified occupancy grid map[6]. On he oher hand, several researchers proposed o sor he raw poin clouds daa in he way of he sandard elevaion map (DEM)[7-8] and he muli-level surface map (MLS)[9]. The applicaion of he vision sysems for oudoor localizaion has also received increasing aenion[0]. Differen robus core algorihms such as he Scale Invarian Feaure Transform (SIFT), he Posiion Invarian Robus Feaure (PIRF) and ec. have been developed o adap o he complex oudoor environmens[-2]. Bu unpredicable long compuaion periods sill made he above echniques fail for many real-ime applicaions[3]. Ousourcing map based localizaion is anoher kind of he mehods o esimae he robo posiion. Mandel e al. proposed a novel approach o ake advanage of he road nework s srucure and is heigh profile for posiion esimaion when was los [4]. Our research group also proposed a new localizaion mehod wih less occupied memory where he opographical map was uilized as he prior available errain map for localizaion[5]. However, in he case of he large-scale exploraion and long-erm auonomous, all he above effors are no enough. Recenly few researchers have ried o face he challenges on he long-erm auonomous robo in oudoor environmens where he mobile robo is expeced o run auonomously over a long period of ime and adap o he real dynamic scenarios. Zhao e al. proposed a simulaneous localizaion and mapping (SLAM) mehod o simulaneously deec and rack he moving obecs using a laser scanner in a dynamic environmen[6]. The auhors found ou ha he mehod was very ime-consuming o rack many saic or moving obecs. Badino e al. and Neuber e al. described a novel concep of appearance change predicion o learn how he environmen changes over ime beforehand, and hen ake advanage of he learned knowledge o predic is appearance under differen environmenal condiions[7-8]. The key limiaion of his mehod was he requiremen of a large sorage space o sore differen environmenal condiions as many as possible and he map informaion of he navigaion area. On he oher hand, if he real environmen changed, i would be hard o updae he map. Cloud roboics provides a very promising soluion o overcome

2 such problems[9]. Cloud roboics is applying he cloud compuing concep o robos in order o augmen he robos capabiliies by off-loading compuaion and shares huge daa or new skills via he inerne. The relaed works on cloud roboics are sill rare so far. Arumugam e al. buil a cloud compuing infrasrucure DAvinCi o improve he implemenaion speed of simulaneous localizaion and mapping (SLAM) [20]. Kehoe e al. developed an archiecure for a cloud roboics sysem o recognize and grasp he common household obecs[2]. Wang e al. inroduced a generic infrasrucure of cloud roboic sysem o enable several poor-equipped robos o rerieve locaion daa from a dynamically updaed map which was buil by a well-equipped robo[22]. In his paper, we proposed a cloud based localizaion echnique using ousourcing road nework maps o achieve long-erm/large-scale auonomous navigaion of he mobile robo in oudoor environmens. The proposed echnique in his paper is aimed o solve wo problems. One is ha would receive insufficien saellie signals among he buildings and oher consrucions during he long-erm oudoor auonomous navigaion. The oher problem is ha he long-erm/large-scale auonomous navigaion would require grealy increasing compuaional payloads. Hence he road nework maps are firs exraced by Google Earh, OpenSreeMap or oher commercially available resources such ha he new roads could be added ino he map daabase in he cloud. The Terrain Inclinaion Aided Localizaion algorihms recenly proposed by our research group will hen be applied o achieve online localizaion only aking advanage of he road nework maps sored in he cloud [5]. In his paper, he cloud can no only provide sorage space o sore he large amoun of map daa, bu also provide a new way o obain he laes map informaion. The srucure of his paper follows. Background and Relaed work are discussed in Secion. The cloud roboics archiecure and he deailed algorihms are proposed in Secion2. Experimenal resuls and discussion are presened in Secion 3. Conclusions are described in Secion4. II. CLOUD ROBOTICS ARCHITECTURE AND ALGORITHMS The proposed cloud roboics archiecure has wo phases: offline and online, Figure. A. Offline phases The offline phase is o exrac he new road neworks and add hem o he road nework map sored in he cloud before each ask is execued, Figure. In order o obain he new road neworks, a se of poins were labeled along he new roads, Figure.2. Then he geodeic coordinae (laiude, longiude, aliude) of hese poins locaed on he road neworks ogeher wih he new road names can be exraced from he Google Earh ha could provide he Offline phase: Online phase: Google Earh Iniial Posiion Exraced daa: Laiude Longiude Aliude New road informaion Updaed he whole road nework daabase IMU Encoder RTI model Paricle Filer Cloud Figure.Cloud-based localizaion Archiecure. Remoe Robo Posiion Esimaion laes informaion of he roads. Therefore he road nework map in he cloud is updaed and ready for online ask execuion. B. Online phases The online phase is o achieve he localizaion ask based on he laes road nework map. According o Figure, he cloud based archiecure consiss of wo secions: he mobile robo and he cloud. The cloud par includes he updaed road nework map and he RTI model proposed by our research group [23]. RTI model is used o describe he relaionships beween he robo aiude and he robo posiion and can be exraced from he road nework map. When he robo moves on a road, he iniial posiion esimaed by he will be sen o he cloud. If signal is no on he road, he esimaed iniial posiion will be pulled o a neares road poin by searching he road nework map. This road poin would be reaed as he new esimaion of he iniial posiion. All he road segmens wihin he neighborhood of he iniial esimaion wih he radius of are searched. The robo-errain inclinaion (RTI) model for he corresponding road segmens is hen Figure.2. Google Earh and he poins ses on he pre-planned pah.

3 B A z z B A o B A y y B A B L A x O x Figure 3.Robo pah is segmened ino a series of line segmens. Figure 4.Geomeric exracion of RTI model. compued in he cloud and sen back o he robo. On he robo par, a 3-D inerial sensor is insalled o measure he aiude and velociies of he robo. Then a paricle filer algorihm is used o incorporae he inerial sensor daa o deermine he 3-D posiion of he robo based on he RTI model sen from he cloud. RTI Model in he Cloud: Suppose here is one porion of he road neworks B A B A, Figure 3. The posiions of he poin se on he road are firs ransformed from he geodeic coordinae (laiude, longiude, aliude) o he Caresian coordinae (x, y, z). I is hen proeced ono he x-o-y plane as B A B A, Figure 3. The proeced road is segmened ino a series of line segmens wih a fixed inerval Lk where L is he lengh of he robo. The B A represens h he line segmen. The poins B and A are he proecions of he road poins B and A, respecively. The z value of hese road poins can be obained by a weighed average inerpolaion mehod from he road nework map h [24]. When he robo moves above he line segmen along he pre-planned pah, he poins B and A represen he midpoin beween wo ground conac poins of he fron wheels and he lef-rear wheel, respecively. The BA represens he direcion of he robo moion, Figure 4. The heading angle,,is defined as he angle beween he B A and he x axis. The heading angle is exclusively deermined by he pah. The angle is defined as he one beween he robo direcion BA and he x-o-y plane. So he angles can be obained from he following equaions, ( za zb) sin ( ) () BA B A ( x x ) ( y y ) ( z z ) (2) B A B A B A where he coordinae informaion of he poins includes B = ( xb, yb, zb) and A = ( xa, ya, z A).Therefore, a number of angles (, ) can be exraced from he serial line segmens B A. Then he robo posiion ( x, y, z ) a each line segmens corresponds wih each group (, ).By linear inerpolaion of he above discree T k k k k k relaionship, RTI _ Model ( x, y, z ) can be obained, k, 2,, N. The number N can be adused for he accuracy requiremen. Communicaion beween he Cloud and he Robo: Assumpion in his paper is ha he robo and he cloud share wih he same nework. The cloud service creaes he lisener socke ha is waiing for remoe cliens o connec. The clien issues he connec() socke funcion o sar he TCP handshake. This socke conains many parameers of he clien, such as IP address, por number and so on. If hese parameers are he same as hose in he lisener socke, hen he cloud server issues he accep() socke funcion o accep he connecion reques. Thus he communicaion beween he cloud and he robos can be esablished. Paricle Filer Algorihm on he Robo: See he deailed algorihm in Table.The sysem sae, X, represens he hree-dimensional posiion of he robo in he inerial frame (x, y, z) a he ime. The superscrip [m] denoes he paricle m, T is he sampling period, and v is he linear velociy in he direcion of robo movemen.

4 Table Algorihm: Paricle Filer based localizaion, v, z ) : ( X 2: [] [2] [ M ] X,,..., 3: for m= o M do, z,, dis RTI _ Model and RTI _ Model are he RTI model downloaded from he cloud ha is reaed as he [ ] [ ] [ ] measuremen model. p( x m, m, m y z ) represens he posiion of he paricle m and p( x, y, z ) represens he road poin on he robo pah closes o his paricle ha is gained [ ] from he map. ˆ m dis is he disance beween he paricle m [ m] and he poin p( x, y, z ) [4]. w is he weighing facor of 4: [ m] [ m] cos cos [ m] [ m] [ m] sin cos [ m] [ m] z z sin v T, Q, X X x x v T y y v T //moion model 5: ˆ[ m] ˆ [ m] [ m] [ m] [ m] RTI _ Model ( x, y, z ) [ m] [ m] [ m] [ m] [ m] [ m] zˆ ˆ ˆ RTI _ Model ( x, y, z ) ˆ [ m] ˆ [ m] [ m] [ m] [ m] dis dis dis( p( x, y, z) p( x, y, z )) // measuremen model 6: [ ] T m m m w 2 2 Q exp z zˆ Q z zˆ 2 // weigh calculaion 7: [ m] [ ] Add and w m o X 8: endfor 9: for m= o M do 0: Draw i wih probabiliy w [] : [] Add i o X 2: endfor 3: reurn X,,..., [] [2] [ M ] he paricle m for resampling of he paricle filer. III. EXPERIMENTS AND DISCUSSION A. Mehods and Procedures The experimens were conduced on he plaform, a Summi XL mobile robo. The NAV440 from Crossbow Technology was used as he inerial measuremen uni (IMU) ha mouned on he op surface of he robo in order o measure he roll, pich, yaw angles. The measuremen accuracy was 0.5 degree in he roll and pich direcions while degree in he yaw direcion. The line velociy of he robo was provided by he encoders. An oudoor environmen wih he area of 200m x 400m in he Shenzhen Universiy Town, Figure.2, was used for performance evaluaion. The sampling period for all experimens is 0.s. B. Resuls and Discussion The road neworks of he seleced area were exraced from he Google Earh and sen o he cloud service before he localizaion ask sared. Comparing wih he road nework daabase saved previously (Figure 5 (a)-(b)), i was found ou ha Road # and #2 on he curren road neworks were new, Figure 5 (b). Therefore,he laes road nework informaion was updaed in he cloud. When he Summi robo sared o move on he road, he iniial posiion esimaion by, G, was ransmied o he cloud service. So he shor-disance road poin, E, was searched in he cloud because his esimaed posiion was no on any of he road. The roads wihin he neighborhood of he poin E wih he radius 200m were obained, Figure 5 (b). Then a local RTI model saring from his new iniial poin along he pre-planned pah was compued in he cloud service and sen back o he mobile r Road # G E d Road #2 (a) Road neworks previously exising (b)updaed road neworks from he Google Earh Figure 5.Road neworks of he experimenal area.

5 laiude G R E Saring poin (Ref) d Saring poin () Esimaed by Proposed Mehod Reference Road Neworks E2 R R E3 Desinaion poin () Desinaion poin (Ref) R longiude Figure 6. The esimaion of he robo posiion by he proposed echnique. robo. Finally, he localizaion can be achieved on he robo by applying he paricle filer algorihm in Secion 2. The esimaed posiion a he end of each 200m ravelling disance of he robo (E2 and E3 in Figure 6) was sen o he cloud again, and he above procedures were repeaed. Figure 6 depics he posiion esimaion of he robo by he proposed echnique (solid line) compared wih he alone (dashed line) and he reference posiions (circle signs) wih he robo speed.0m/s. According o Figure 6, he posiion esimaion by he proposed mehod was much closer o he ground ruh values inside he Area (from he reference poin R R2) and Area 3 ( R3 R4). On he Area 2(from he reference poin R2 R3 ), he performance of he proposed mehod was quie similar o he one esimaed by alone. I was found ou he Area 2 was a wide playground and signal was already effecive while Area was surrounded by wo buildings and signals were worse. The same phenomena were also observed in Figure 7. Figure 7 shows he comparison of he posiion esimaion errors using he proposed echnique (circle sign) and he alone (circle sign). The esimaion error a he ravelling disance of 380m was up o (9m, 9m) using alone according o Figure 7 because was parially blocked. This resul has apparenly poined o somewhere off-road. A he same ime, he esimaion error using he proposed mehod has reached o he value (0.3m, 3m). This esimaed posiion was sill on he road, which was coinciden wih he acual fac. Hence i is concluded he proposed cloud based echnique can achieve online localizaion for large-scale road neworks. In he near fuure, more long-erm experimens will be carried ou, and more complex scenarios will be considered. IV. CONCLUSIONS This paper inroduces a cloud-based ousourcing localizaion echnique for a mobile robo on oudoor road neworks. Preliminary experimenal resuls validae he proposed echnique and illusrae ha he proposed echnique has capabiliy o achieve online localizaion aking advanage of ousourcing road nework maps and he relaive algorihms in he cloud. This mehod will be applied o more large-scale/long-erm circumsances. REFERENCES [] P. Bonnifai, P. Bouron, P. Crubille, and D. Meizel, "Daa fusion of four ABS sensors and for an enhanced localizaion of car-like vehicles," presened a IEEE Inernaional Conference on Roboics and Auomaion(ICRA), 2-26 May 200, Seoul, Souh Korea, 200. [2] D. Bouve and G. Garcia, "Civil-engineering ariculaed vehicle localizaion: soluions o deal wih masking phases," presened a

6 error in x direcion(m) error in y direcion(m) IEEE Inernaional Conference on Roboics and Auomaion(ICRA), April 2000, San Francisco, CA, USA, [3] C. Brenneke, O. Wulf, and B. Wagner, "Using 3D Laser Range Daa for SLAM in Oudoor Environmens," presened a 2003 IEEE/RSJ Inernaional Conference on Inelligen Robos and Sysems, Oc , Las Vegas, NV, Unied Saes, [4] A. Nücher, H. Surmann, K. Lingemann, J. Herzbergand S. Thrun."6D SLAM wih an Applicaion in Auonomous Mine Mapping," presened a Inernaional Conference on Roboics and Auomaion(ICRA), April 26-May, New Orleans, USA, [5] M. C. Marin and H. Moravec," Robo evidence grids. Technical Repor CMU-RI-TR-96-06," Roboics Insiue, Carnegie Mellon Universiy, Pisburgh, PA. [6] N. Fairfield, G. Kanor and D. Weergreen, " Real-Time SLAM wih Ocree Evidence Grids for Exploraion in Underwaer Tunnels," Journal of Field Roboics, vol. 23, pp [7] E. Hygounenc, I.-K. Jung, P. Soueres, and S. Lacroix. " The auonomous blimp proec of laas-cnrs: Achievemens in fligh conrol and errain mapping". In'l Journal of Roboics Research, Vol.23, No.4-5, pp.473 5, [8] P. Pfaff, R. Triebel, and W. Burgard. " An efficien exension o elevaion maps for oudoor errain mapping and loop closing ". In'l Journal of Roboics Research, Vol. 26, No. 2, pp , [9] R. Triebel, P. Pfaff, W. Burgard."Muli-Level Surface Maps for Oudoor Terrain Mapping and Loop Closing. " presened a Inernaional Conference on Inelligen Robos and Sysems (IROS),9-5 Oc. 2006, Beiing, China,2006. [0] P. Blaer and P. Allen, Topological Mobile Robo Localizaion Using Fas Vision Techniques, presened a Inernaional Conference on Roboics and Auomaion(ICRA), -5 May 2002, Washingon DC, USA, [] T. D. Barfoo, Online visual moion esimaion using fasslam wih sif feaures, presened a Inernaional Conference on Inelligen Robos and Sysems (IROS), 2-6 Aug, Edmonon, Canada, [2] A. Kawewong, N.Tongprasi, S.Tangruamsub, O.Hasegawa,"Online and Incremenal Appearance-based SLAM in Highly Dynamic Environmens," In'l Journal of Roboics Research, Vol.30, No., pp.33 33, 200. [3] N. Tongprasi, A. Kawewong, O. Hasegawa," PIRF-Nav 2: Speeded-up Online and Incremenal Appearance-based SLAM in An Indoor Environmen," presened a IEEE Workshop on Applicaions of Compuer Vision, Washingon, DC, USA. [4] C. Mandel, T. Laue, " Paricle filer-based posiion esimaion in road neworks using digial elevaion models," presened a IEEE/RSJ Inernaional Conference on Inelligen Robos and Sysems, Taipei, Taiwan.200 [5] X. Zhu, C.Qiu, M. A. Minor, "Terrain Inclinaion Aided Three-Dimensional Localizaion and Mapping for an Oudoor Mobile Robo," Inernaional Journal of advanced Roboic Sysems, vol.0,76. [6] H. Zhao e al., "SLAM in a Dynamic Large Oudoor Environmen using a Laser Scanner," presened ainernaional Conference on Roboics and Auomaion (ICRA), 9-23 May. 2008, Pasadena,USA, [7] H. Badino, D. Huber, T. Kanade, "Real-Time Topomeric Localizaion," presened a Inernaional Conference on Roboics and Auomaion (ICRA), June. 203, Berlin, Germany, 203. [8] N. Sunderhauf, P. Neuber, P. Prozel, "Predicing he change-a Sep Towards Life-Long Operaion in Everyday Environmens," presened a 203Roboics: Science and Sysems (RSS) Roboics Challenges and Vision Workshop,4-8 May. 202, Minnesoa,USA, 202. [9] J. Kuffner, "Cloud-EnabledRobos," In IEEE-RAS Inernaional Conference on Humanoid Robos, 6-8 December, Nashville,USA,200. [20] R. Arumugam e al., "DAvinCi: A Cloud Compuing Framework for Service Robos," presened a Inernaional Conference on Roboics and Auomaion (ICRA), 3-8 May, Anchorage,USA, 200. [2] B. Kehoe e al., "Cloud-Based Robo Grasping wih he Google Obec Recogniion Engine," presened a Inernaional Conference on Roboics and Auomaion (ICRA), 6-0 May, Karlsruhe, Germany, 203. [22] L.Wang, M. Liu and M. Meng, "Towards Cloud Roboic Sysem: A Case Sudy of Online Co-localizaion for Fair resource Compeence," presened a Inernaional Conference on Roboics and Biomimeics, -4 December, Guangzhou, China, 202. [23] C. Qiu, X. Zhu and L. Liu, Robo Terrain Inclinaion Model Exraced from Laser Scanner Daa for Oudoor Environmen. presened ainernaional Conference on Mecharonics and Auomaion, 5-8 Augus,Chengdu, China, 202. [24] H. Chen, X. Li and W. Ding, "Twelve Kinds of Gridding Mehods of Surfer 8.0 in Isoline Drawing," Chinese ournal of engineering geophysics, vol. 4, pp , TIL based on Paricle Filer 0 9 TIL based on Paricle Filer disance(m) disance(m) Figure 7.The posiion esimaion errors of he robo using he proposed echnique.