WLAN based Pose Estimation for Mobile Robots

Transcription

1 Proceeings of the 17th Worl Congress The International Feeration of Automatic Control WLAN base Pose Estimation for Mobile Robots Christof Röhrig an Frank Künemun University of Applie Sciences Dortmun, Emil-Figge-Str. 4, 447 Dortmun, Germany, Abstract: Nowaays, many builings are equippe with a WLAN infrastructure, as an inexpensive communication technology. In this paper a metho to estimate position an heaing (pose) of a mobile robot using WLAN technology is escribe. The propose technique for localizing a mobile robot is base on the use of receive signal strength values of WLAN access points in range. A raio map base metho an Eucliean istance in combination with Delaunay triangulation an interpolation is propose. Measure signal strength values of an omniirectional antenna an a beam antenna are compare with the values of a raio map, in orer to estimate the pose of a mobile robot, whereby the irectionality of the beam antenna is use to estimate the heaing of the robot. The paper presents the experimental results of measurements in an office builing. 1. INTRODUCTION Navigation is a key ability of mobile robots. The task of navigation can be ivie into localization an path planning. Aim of localization is to estimate the pose (position an heaing) of a mobile robot with respect to its environment. There are three ifferent kins of localization problems in mobile robotics: position tracking, global localization an kinappe robot problem. Position tracking requires knowlege of the start position an is also known as local localization. The problem is calle global localization, if there is no priori estimate of the pose. The kinappe robot problem escribes a situation, where a localize robot is move to a ifferent place without its knowlege. It is often use to test a robot s ability to recover from localization failures. Approaches which are capable of solving the global localization problem can be moifie such that they can also solve the kinappe robot problem (Fox, ). Usually robot oometric sensors are use to solve the localization problem of wheele robots. Oometric sensors provie information about robot movements, but the provie information is noisy an accumulates errors over time. Oometrie is accurate enough for local movements but is not suitable for long term localization an global localization (Thrun, ). Aitional sensors such as laser an vision provie information about the environment of a mobile robot. Several methos have been propose to use this information to estimate the pose of a mobile robot. Unfortunately laser sensors are expensive an vision nees computational overhea of image processing. Furthermore this techniques require a map an usually a start position. If the start position is unknown, the pose have to be searche in the whole map, which is ifficult an time consuming in a large environment. A global pose estimation using WLAN technology can support such methos by fining the starting pose. Furthermore it can solve the kinappe robot problem by etecting localization failures an by proviing a new starting pose. Nowaays mobile robots often are equippe with IEEE 8.11 WLAN aapters, in orer to communicate with computers or This work was supporte in part by the research program of the University of Applie Sciences Dortmun. other mobile evices. Furthermore, many builings are alreay equippe with an IEEE 8.11 WLAN infrastructure, as a popular an inexpensive technology. Most WLAN aapters are able to measure the signal strengths of receive packets as part of their stanar operation. The signal strengths of receive packets vary noticeably by changing the position. Thus, the signal strength can be use to estimate the position of a mobile evice by cheap technology. In this paper, the problem of global localization is solve using the WLAN infrastructure in an inoor scenario. It extens the existing WLAN localization techniques in two ways. First, it escribes a technique for estimating the heaing of a mobile robot. A measure set of signal strength values of an omniirectional antenna an a beam antenna is compare with a raio map in orer to estimate position an heaing of a mobile robot. Secon, Delaunay triangulation an interpolation is propose in orer to reuce the ensity of the calibration points in the raio map an thus minimizing the manual effort to buil the map. Furthermore the propose technique can support other map base pose estimation methos by fining a global start position. The paper extens the work presente in (Röhrig an Künemun, 7) by refining the algorithm of heaing estimation an by presenting more experimental results, which show the effectiveness of the technique.. RELATED WORK Up to now there are evelope several kins of localization techniques for the use in wireless networks. Saye et al. (5) give a review of existing techniques. This techniques can be classifie by the information they use: Connectivity information, Angle of Arrival (AoA), Time of Arrival (ToA), Time Difference of Arrival (TDoA), Receive Signal Strength (RSS). Connectivity information is available in all kins of wireless networks. The accuracy of the localization epens on the range of the use technology an the ensity of the beacons. In cellular networks, Cell-ID is a simple localization metho base on /8/$. 8 IFAC /876-5-KR

2 17th IFAC Worl Congress (IFAC'8) cell sector information (Guolin Sun et al., 5). In infrastructure moe of a Wireless LAN (WLAN), the access point (AP) to which the mobile evice is currently connecte, can be etermine since mobile evices know the MAC harware aress of the AP, which they are connecte to. Bluetooth is another technology, which allows a relatively accurate localization because of its low raio range (Felmann et al., 3). Besies the eployment of Raio Frequency Ientification (RFID) in Supply Chain Management (see Michael an McCathie, 5), the RFID technology is also suitable for position estimation. RFID tags can be eploye at known positions in the environment, in orer to obtain position information when they are in range. This information can be be fuse with ata from other sensors (e.g. oometers) for the purpose of improving the accuracy of localization (Hähnel et al., 4; Koch et al., 7). AoA etermines the position with the angle of arrival from fixe anchor noes using triangulation. Nasipuri an Li () propose a metho, where a wireless sensor noe localizes itself by measuring the angle to three or more beacon signals. Each signal consists of a continuous narrow irectional beam, that rotates with a constant angular spee. Drawback of AoA base methos is the nee for a special an expensive antenna configuration e.g. antenna arrays or rotating beam antennas. ToA an TDoA estimate the range to a sener by measuring the signal propagation elay. The Cricket localization system (see Priyantha et al., 1) evelope at MIT utilizes a raio signal an a ultrasoun signal for position estimation base on trilateration. TDoA of these two signals are measure in orer to estimate the istance between two noes. This technique can be use to track the position of a mobile robot (Alriksson an Rantzer, 7). ToA as wells as TDoA require a complex wireless network infrastructure, which is usually not present in toay s WLAN installations. RSS information can be use in most wireless technologies, since mobile evices are able to monitor the RSS as part of their stanar operation. The istance between sener an receiver can be obtaine with the Log Distance Path Loss Moel escribe in (Patwari et al., 3). Unfortunately, the propagation moel is sensitive to isturbances such as reflection, iffraction an multi-path effects. The signal propagation epens on builing imensions, obstructions, partitioning materials an surrouning moving objects. Own measurements show, that this isturbances make the use of a propagation moel for accurate localization in an inoor environment almost impossible (see Röhrig an Künemun, 7). Bahl an Pamanabhan () introuce fingerprinting, which is a metho to overcome this isavantage by utilizing a raio map. Fingerprinting is ivie in two phases: In the initial calibration phase, the raio map is built by moving aroun an storing RSS values at various preefine points of the environment. In the localization phase, the mobile evice moves in the same environment an the position is estimate by comparing the current RSS values with the raio map. A metric to compare the measure RSS values with the raio map is Eucliean istance propose by Bahl an Pamanabhan (). La et al. (4) use a Bayesian algorithm an Großmann et al. (6) propose Delaunay triangulation with lines of constant signal strength. Several methos for localization in WLAN environments using RSS have been evelope, in orer to improve the accuracy of the estimation. Kumar K. et al. (6) propose a Kalman filter, a Monte-Carlo algorithm is use by Hu an Evans (4) as well a by Moster an Tews (6), Brunato an Battiti (5) apply statistical learning an Fuzzy is use by Astrain et al. (6) an by Teuber an Eissfeller (6). All of these methos utilize a raio map an estimate only the position but not the heaing of the mobile evice. The main isavantage of raio map base methos is the high manual effort to buil the map in the calibration phase. The use of Delaunay triangulation an interpolation allows a raio map with a low ensity of calibration points an reuces the time for manual generation of the map (Röhrig an Künemun, 7). 3. LOCALIZATION APPROACH The propose metho for localizing the mobile robot is base on the use of RSS values of WLAN APs in range. A raio map base metho an Eucliean istance in combination with interpolation is use, because of the escribe reasons. A measure set of RSS values of the omniirectional antenna an the beam antenna is compare with the raio map. The raio map is built in an initial calibration phase an contains measure sets of RSS values at various preefine poses (x, y,θ). In the experiments, four heaings (, 9, 18, 7 ) at every position are store. In the localization phase, RSS values of several APs are recore an compare with the raio map. One observation in both phases consists of RSS values of both antennas an all APs. Values of APs out of range are set to a minimal value c min = 1 Bm. 3.1 Eucliean istance Eucliean istance is a metric to compare the observations of the localization phase with the raio map. In WLAN localization the calibration ata are compare with the measure ata: M ( j = c AP i j s i) AP (1) i=1 where c AP i j is the RSS value of AP i at pose j in the raio map, s AP i is the measure RSS value of AP i an M is the total number of APs. The Eucliean istance j is a metric for the istance between the calibration ata c AP i j an the measure ata s AP i in signal space. After calculating j for all calibration points, there will be at least one pose with minimal j. One approach is to eclare this pose to be the estimate pose of the mobile robot. The accuracy of this metho epens besie other factors on the ensity of the unerlying gri of calibration points. 3. Estimation of the position In orer to reuce the manual effort to buil the map, the ensity of calibration points shoul be as low as possible. Thus, the interpolation of the estimate position is propose. In this case a lower ensity of calibration points is possible. Interpolation is base on Delaunay triangulation an lines of constant signal strength (isolines). For interpolation purposes of the position, the signals of the omniirectional antenna are use only. With Delaunay triangulation a network of triangles for a set of points (noes) in the plane is evelope, such that no point is insie the circumcircle of any triangle (Leach, 199). The noes are represente by the calibration points. Given a measure RSS value of one AP, triangles whose noes show RSS values higher an lower than the measure value can 1434

3 17th IFAC Worl Congress (IFAC'8) be selecte. Linear interpolation between noe values within the triangle elivers a more etaile raio map consisting of a surface of interpolate RSS values over the triangle. Moreover, it is possible to calculate an interpolate line of constant RSS (isoline) within the triangle an in the whole area of triangulation. Fig. 1 shows the isolines of AP 1, Fig. shows the isolines of AP. Fig. 3. Merge raio map for AP 1 an AP Fig. 1. Lines of constant RSS (isolines) for AP 1 Fig. 4. Visualization of estimation technique Fig. 4 shows a raio map with calibration poses (re arrows), acceptance circle (cyan) an points of intersection (magenta). The real pose of the robot is shown as blue arrow, while the green arrow represents the estimate pose. The estimate position ( ˆx, ŷ) is calculate with weighte points of intersection: N i=1 ˆx= w N ix i i=1 N i=1 w, ŷ= w iy i N i i=1 w, () i Fig.. Lines of constant RSS (isolines) for AP Given two RSS values of ifferent APs, it is possible to select triangles whose interpolation surfaces inclue the accoring isolines. If there is an intersection of both isolines, the intersection point within the triangle can be calculate. Fig. 3 shows the merge raio map for AP 1 an AP. There are two points of intersection in the raio map for this measurement. The pose is estimate using the points of intersection. Points with a large istance to the real position have to be eliminate from the calculation of the estimate position. The elimination of points of intersection is performe using an acceptance circle. This circle is built by a triangle of three points in the raio map with least Eucliean istance to the measure RSS values. It is assume that the real position is near by this triangle. The center of the circle is the balance point of the triangle. The raius of the circle is built by the largest ege of the triangle. All intersection points outsie of the acceptance circle were exclue from the calculation. where w i is the weight of intersection x i, y i an N is the total number of intersections insie the acceptance circle. Experiments have shown, that measure RSS values closer to APs are more reliable than those in larger istance (see Röhrig an Künemun, 7). Hence, the weight w i of intersection i is calculate with RSS values of the crossing isolines: w i = (s i,1 s min ) +(s i, s min ) with s min = 1 Bm (3) where s i,1 an s i, are the RSS values of the isolines at intersection i an s min is the lowest possible RSS value. Higher RSS values are measure closer to APs an lea to larger weights. 3.3 Estimation of the heaing The heaing is estimate with RSS values of the beam antenna. For every point of intersection i a heaing ˆθ i with assigne vector length ˆρ i is calculate. ˆρ i is a metric for the quality of the estimation an is use as weight. The estimation of (ˆθ i, ˆρ i ) is calculate with raio map values of the surrouning triangle: 1435

4 17th IFAC Worl Congress (IFAC'8) ˆθ i = atan sinθ j, cosθ j j=1 j j=1 j, (4) ˆρ i = sinθ j j=1 j + cosθ j j=1 j, (5) where J is the number of weighte heaings (with least Eucliean istance) at the noes of the surrouning triangle, j is the Eucliean istance between measure RSS values from the beam antenna an store RSS values in the raio map an w i is the weight of intersection i (Eqn. 3). In Fig. 4 (ˆθ i, ˆρ i ) are represente by magenta arrows. The heaing of the mobile robot ˆθ is estimate by aing the heaings of all intersections: N N ˆθ=atan sinθ j, cosθ j i=1 j=1 j i=1 j=1 j, (6) N N ˆρ= sinθ j i=1 j=1 j + cosθ j i=1 j=1 j. (7) In Fig. 4 (ˆθ, ˆρ) is represente by the green arrow. Estimation error [ ] Vector length ρ Measurement Measurement Fig. 5. ˆρ versus estimation error Fig. 5 compares ˆρ with the estimation error of the heaing of 1 measurements at the same location. Measurements 7 an 8 achieve large estimation errors of 18. This large errors correspon with very low values of ˆρ, which inicate a low estimation accuracy. It is propose to use ˆρ in a later signal processing stage to weight the estimate heaing. 4. EXPERIMENTAL SETUP The experiments are carrie out with a mobile robot Pioneer3- AT manufacture by ActivMeia (Fig. 6). The robot is equippe with an embee computer for real time robot control an an aitional PC with two WLAN cars for communication an localization. One WLAN car is connecte to an omniirectional antenna, the other car is connecte to a beam antenna. A robot server is inclue in the operating system of the embee computer. It manages the low-level tasks of robot control an operation, incluing motion an oometry. For programming purposes ActivMeia provies the toolkit ARIA (Activ- Meia Robotics Interface for Application) (see Activmeia, Fig. 6. Pioneer3-AT 6). ARIA provies an interface to control the robot s velocity, heaing, relative heaing, an provie etaile information about oometry an operating conitions from the mobile robot. The operation system on the PC is Ubuntu Linux, which offers support for wireless communication by the wireless extension (WE). WE is an application programming interface (API) allowing a user space program to configure a WLAN river an receive statistic information. The software is ivie into three parts: a localization engine, a graphical user interface (GUI) an a WLAN scanner. The localization engine an the GUI are written in the Matlab script language, the WLAN scanner is implemente in C. The WLAN scanner uses the WEioctl()- Interface for reaing RSS values from both WLAN aapters. The communication between localization engine an WLAN scanner is buil with TCP/IP sockets. Since the localization engine are built on Matlab, it is possible to run it on every computer which offers a Matlab environment an a network access. The GUI is use for monitoring information to the user an for builing the raio map. Fig. 7 shows the GUI with a map of a room in the Computer Science builing. The re arrows aroun the re ots show the four heaings of the robot in every calibration point of the raio map. The colore lines represent the RSS isolines. In orer to buil the raio map, the user moves the robot to preefine poses (re arrows) an stores the measure RSS values. In the localization phase, the blue arrow represents the real pose of the mobile robot an green an yellow arrow visualizes estimate poses. 5. EXPERIMENTAL RESULTS Experiments are performe in an office builing of the Computer Science Department. A test series was measure at the hallway shown in Fig. 4. The existing WLAN-infrastructure was use for the measurements. Fig. 8 shows a histogram of position errors achieve in this test series, which are in a range from to 4 m. In most cases, the accuracy is better than 1. m. 1436

5 17th IFAC Worl Congress (IFAC'8) Fig. 7. Matlab Localization engine an GUI The accuracy epens irectly on the position of the APs in range. For a goo an reliable estimation, three ore more APs in a short istance are require. The placement of aitional APs increases the accuracy of the estimation Euclien istance Delaunay interpolation Frequency 8 Frequency Estimation error [m] Estimation error [m] Fig. 8. Histogram of estimation error of position Fig. 9 compares the estimation accuracy of Delaunay isoline interpolation with Eucliean istance interpolation. In most cases estimations with isoline metho achieve a better accuracy than estimations with Eucliean istance only. Fig. 1 shows a histogram of the heaing error. Heaing errors are in the full range from to 18. In most cases an accuracy better than 45 can be achieve an worse estimations can be etecte by low values of ˆρ. 6. CONCLUSION AND FURTHER WORK This paper has presente a metho for estimating the pose of a mobile robot. The metho is base on a raio map an uses RSS values of WLAN APs in range. In orer to reuce Fig. 9. Comparison of Eucliean istance an Delaunay interpolation the ensity of calibration points in the raio map, Delaunay triangulation is applie to interpolate the position of the mobile robot. Furthermore the estimation of the heaing of a mobile robot with the ai of a beam antenna was presente. Since the accuracy of the estimation epens highly on the positions of the APs in the environment, the placement of the APs has to be optimize, in orer to get a goo an reliable position estimation. In future work the accuracy of the estimation shoul be improve by using a Kalman filter or a Monte Carlo particle filter. Furthermore the accuracy may be improve by fusion with position information obtaine from other sensor e.g. oometry, sonar or laser. 1437

6 17th IFAC Worl Congress (IFAC'8) Frequency Estimation error [ ] Fig. 1. Histogram of estimation error of heaing REFERENCES Activmeia. Avance Robotics Interface for Applications (ARIA) Developer s Reference Manual. ActivMeia Robotics, 6. P. Alriksson an A. Rantzer. Experimental Evaluation of a Distribute Kalman Filter Algorithm. In Proceeings of the 46th IEEE Conference on Decision an Control, pages , New Orleans, December 7. J. J. Astrain, J. Villaangos, J.R. Garitagoitia, J.R. González e Menívil, an V. Cholvi. Fuzzy Location an Tracking on Wireless Networks. In Proceeings of 4th ACM International Workshop on Mobility Management an Wireless Access, pages 84 91, Torremolinos, Spain, October 6. P. Bahl an V. N. Pamanabhan. RADAR: An In-Builing RFbase User Location an Tracking System. In Proceeings of the 19th Annual Joint Conference of the IEEE Computer an Communications Societies, volume, pages , Tel Aviv, Israel, March. M. Brunato an R. Battiti. Statistical Learning Theory for Location Fingerprinting in Wireless LANs. Computer Networks, 47(6):85 845, April 5. S. Felmann, T. Hartmann, an K. Kyamakya. An inoor Bluetooth-base Positioning System: Concept, Implementation an Experimental Evaluation. In Proceeings of the International Conference on Wireless Networks, pages , Las Vegas, USA, June 3. D. Fox. KLD-Sampling: Aaptive Particle Filters an Mobile Robot Localization. Technical Report UW-CSE-1-8-, University of Washington,. U. Großmann, C. Röhrig, S. Hakobyan, T. Domin, an M. Dalhaus. WLAN Inoor Positioning base on Eucliian Distance an Interpolation (Isobars). In Proceeings of the 8th Wireless Technologies Kongress, pages 96 35, Dortmun, Germany, 6. Guolin Sun, Jie Chen, Wei Guo, an K.J.R. Liu. Signal Processing Techniques in Network-aie Positioning: A Survey of State-of-the-art Positioning Designs. Signal Processing Magazine, (4):1 3, July 5. D. Hähnel, W. Burgar, D. Fox, K. Fishkin, an M. Philipose. Mapping an Localization with RFID Technology. In Proceeings of the International Conference on Robotics an Automation, volume 1, pages 115 1, New Orleans, USA, May 4. L. Hu an D. Evans. Localization for Mobile Sensor Networks. In Proceeings of the 1th Annual International Conference on Mobile Computing an Networking, pages 45 57, 4. J. Koch, J. Wettach, an E. Bloch. Inoor Localisation of Humans, Objects, an Mobile Robots with RFID Infrastructure. In Proceeings of the 7th International Conference on Hybri Intelligent Systems, pages 71 76, Kaiserslautern, Germany, September 7. R. Kumar K., V. Apte, an Y. A Powar. Improving the Accuracy of WLAN base Location Determination Using Kalman Filter an Multiple Observers. In Proceeings of the Wireless Communications an Networking Conference, volume 1, pages , Las Vegas, April 6. A. M. La, K. E. Bekris, A. Ruys, D. S. Wallach, an L. E. Kavraki. On the Feasibility of Using Wireless Ethernet for Inoor Localization. IEEE Transaction on Robotics an Automation, (3): , June 4. G. Leach. Improving Worst-Case Optimal Delaunay Triangulation Algorithms. In Proceeings of the 4th Canaian Conference on Computational Geometry, 199. K. Michael an L. McCathie. The Pros an Cons of RFID in Supply Chain Management. In Proceeings of the International Conference on Mobile Business, pages 63 69, Syney, Australia, July 5. F. Moster an A. Tews. Practical WiFi Localization for Autonomous Inustrial Vehicles. In Proceeings of the Australasian Conference on Robotics an Automation 6, Aucklan, New Zealan, December 6. A. Nasipuri an K. Li. A Directionality base Location Discovery Scheme for Wireless Sensor Networks. In Proceeings of the 1st ACM International Workshop on Wireless Sensor Networks an Applications, pages , Atlanta, USA, September. N. Patwari, A. O. Hero, M. Perkins, N. S. Correal, an R. O Dea. Relative Location Estimation in Wireless Sensor Networks. IEEE Transactions on Signal Processing, 51(8): , 3. N. B. Priyantha, A. K. L. Miu, H. Balakrishnan, an S. Teller. The Cricket Compass for Context-aware Mobile Applications. In Proceeings of the 7th Annual International Conference on Mobile Computing an Networking, pages 1 14, Rome, Italy, July 1. C. Röhrig an F. Künemun. Estimation of Position an Orientation of Mobile Systems in a Wireless LAN. In Proceeings of the 46th IEEE Conference on Decision an Control, pages , New Orleans, USA, December 7. A.H. Saye, A. Tarighat, an N. Khajehnouri. Network-base Wireless Location: Challenges face in Developing Techniques for Accurate Wireless Location Information. Signal Processing Magazine, (4):4 4, July 5. A. Teuber an B. Eissfeller. WLAN Inoor Positioning Base on Eucliean Distances an Fuzzy Logic. In Proceeings of the 3r Workshop on Positioning, Navigation an Communication, pages , 6. S. Thrun. Probabilistic Algorithms in Robotics. AI Magazine, 1(4):93 19,. 1438