Spinning Beacons for Precise Indoor Localization

Transcription

1 Spinning Beaons for Preise Indoor Loalization Ho-lin Chang 1 Jr-ben Tian 1 Tsung-Te Lai 1 Hao-Hua Chu 1,2 Polly Huang 2,3 Department of Computer Siene and Information Engineering 1 Graduate Institute of Networking and Multimedia 2 Department of Eletrial Engineering 3 National Taiwan University, Taipei, Taiwan {r95004, r93045, r96152, hhu}@sie.ntu.edu.tw, phuang@.ee.ntu.edu.tw ABSTRACT This work proposes the novel use of spinning beaons for preise indoor loalization. The proposed SpinLo (Spinning Indoor Loalization) system uses spinning (i.e., rotating) beaons to reate and detet preditable and highly distinguishable Doppler signals for sub-meter loalization auray. The system analyzes Doppler frequeny shifts of signals from spinning beaons, whih are then used to alulate orientation angles to a target. By obtaining orientation angles from two or more beaons, SpinLo an preisely loate stationary or slow-moving targets. After designing and implementing the system using MICA2 motes, its performane was tested in an indoor garage environment. The experimental results revealed a median error of 40~50 entimeters and a 90% error of 70~90 entimeters. Categories and Subjet Desriptors C.2.4 [Computer-Communiations Networks]: Distributed Systems General Terms Algorithms, Experimentation, Theory Keywords Sensor Networks, Spinning Beaons, Doppler Effet, Angulation, Loalization 1. INTRODUCTION Loation is often essential ontextual information for inferring high-level appliation semantis from low-level data olleted from wireless sensor networks (WSNs) [1][2]. Sensor loalization is therefore a ritial omponent in a WSN system. Although GPS is often used for determining position of a sensor node, it performs poorly and inaurately in indoor environments due to lak of diret Permission to make digital or hard opies of all or part of this work for personal or lassroom use is granted without fee provided that opies are not made or distributed for profit or ommerial advantage and that opies bear this notie and the full itation on the first page. To opy otherwise, or republish, to post on servers or to redistribute to lists, requires prior speifi permission and/or a fee. SenSys 08, November 5 7, 2008, Raleigh, North Carolina, USA. Copyright 2008 ACM /08/11...$5.00. line-of-sight to GPS satellites. Consequently, indoor positioning systems using Wi-Fi reeived signal strength (RSS) and RFIDs have gained popularity. However, despite attempts to improve their positional auray, urrent Wi-Fi RSS and RFID systems are unable to ahieve the sub-meter auray required for many indoor WSN appliations suh as personnel traking in rowded hospitals or asset traking in busy fatories. Thus, high-preision indoor loalization remains a hallenging researh problem. Current high-preision indoor loalization systems with sub-meter auray employ varying methods of overoming indoor multipath interferene. For example, the Ubisense [3] system uses ultra-wideband (UWB) to minimize multipath interferene. However, UWB-based systems [3][4][5] require speialized hardware to ahieve sampling rates and time synhronization in GHz and nanoseond ranges, respetively. Adding suh speialized hardware substantially inreases the ost of typially resoure-onstrained sensor nodes. Aousti systems [6][7] suh as Criket [8] use ultrasoni pulses suffiiently robust to overome indoor multipath interferene but are severely limited by the line-of-sight problem. Further, aousti signals have limited propagation range. The reently proposed Radio Interferometri Positioning System (RIPS) [9][10][11] is a promising and inexpensive alternative to UWB systems. The RIPS provides exellent positional auray and sensing range in outdoor environments using inexpensive and readily available sensor nodes suh as MICA2 motes. However, RIPS is unsuitable for indoor environments beause its RF interferometri ranging tehnique an be severely affeted by indoor multipath interferene. A modified RIPS system developed by Kusy et al. [12][13] used RF Doppler shifts to estimate the diretion of moving targets. Kusy reported that frequeny hange from Doppler shifts was noise-resistant to multipath interferene and thus more appropriate for indoor loalization. However, their experiments were limited to outdoor environments. Additionally, sine the tehnique detets moving targets by Doppler shifts, their tehnique falls bak to the original RIPS for loating stationary or slow-moving targets.

2 This study proposes an inexpensive yet highly preise RF-based indoor loalization system with sub-meter positional auray whih is less suseptible to indoor multipath interferene than urrent loalization systems. The proposed approah employs spinning beaons anhored in the infrastruture to produe preditable and distinguishable Doppler signals for high-preision loalization. Putting spinning motion in the infrastruture brings an additional advantage that the produed Doppler shifts an be used to loate stationary or slow-moving indoor targets. The Spin- Lo system was designed and implemented as an atual loalization system, and its performane was tested in an indoor garage environment with a eiling height of approximately 3 meters. The experimental results revealed a median positional error of 40~50 entimeters and a 90% error of 70~90 entimeters. Two important ontributions of this work are the followings: Rather than relying on the dynami movement of mobile targets to produe irregular, variable Doppler shifts as proposed by Kusy et al. [6], SpinLo reverses this setting by instead relying on the spinning motions of seletive infrastruture nodes to produe preditable, distinguishable Doppler signals for high-preision loalization. A novel Doppler-angulation method was developed to aurately estimate the angle of a target relative to a spinning beaon. An angle from a spinning beaon with known loation fixes a straight line that passes through/nearby the target. By using more than two spinning beaons to produe multiple lines, the target an be loalized from the intersetion of these lines. The rest of this paper is organized as follows. Setion 2 desribes the SpinLo approah. Setion 3 explains the SpinLo positioning system. Setion 4 disusses parametri tuning and performane tradeoff. Setion 5 details SpinLo implementation. Setion 6 desribes the experimental setup and results. Setion 7 reviews the related work. Setion 8 desribes error soures in SpinLo. Setion 9 disusses the harateristis of SpinLo. Finally, Setion 10 onludes the study and suggests diretions for future studies. 2. SPINLOC APPROACH Two phases of the SpinLo method are (1) angulation phase and (2) loalization phase. The angulation phase measures the Doppler signal from a target and a referene node to a spinning beaon. The differene in observed Doppler shifts reveals the angle between the target and the referene node to the spinning beaon 1. Suh angulation is applied repeatedly to several different spinning beaons. The loation of the referene node, the loation of the spin- ning beaon and the estimated angle are then used to determine the probable position of the target. The position of the target an be found by interseting lines drawn from estimated angles to different spinning beaons in the loalization phase. The subsetions below desribe the Doppler Effet, the proposed tehnique for deriving angular information and, finally, the loalization algorithm based on angular information. 2.1 Doppler Effet The Doppler Effet refers to the pereived variation in frequeny and wavelength given the veloity of a moving objet relative to a wave soure. Sine the speed of a radio wave is muh greater than the relative veloity between the wave soure and the observer, the frequeny shift f, also known as the Doppler shift, an be expressed as v f = f (1) where f denotes the transmitted frequeny, and v is the veloity of the wave soure relative to the observer. The relative veloity v is negative when the wave soure is moving away from the observer. As Figure 1 shows, if the wave soure is not moving diretly toward/away from the observer, the relative veloity v in Equation (1) should be replaed by the projeted veloity on the line onneting the observer and wave soure. Figure 1. Illustration of Doppler Effet. 2.2 Doppler-Angulation Doppler-angulation is the proposed method for determining the angle between a stationary target and a referene node to a spinning beaon. This angle is referred to as the orientation angle. Figure 2 illustrates the onept of the Doppler-angulation method where X is the stationary target at distane d from origin O, and the diretion from origin O to X is angle α. The position vetor of target X is: uuur OX = (d os α, d sin α ) (2) 1 The loation of a spinning beaon refers to the entre of its spinning perimeter.

3 Aording to Equation (1), the Doppler shift of the projeted veloity is: f = = = vprojet f f -ωrd sin( ωt+ ϕ - α ) 2 2 d r - 2dr os( ωt+ ϕ - α ) + -ωrdf sin( ωt+ ϕ - α ) 2 d 1 + (r/d) - 2(r/d)os( ωt+ ϕ - α ) -ωrf sin( ωt+ ϕ - α ) = (r/d) - 2(r/d)os( ωt+ ϕ - α ) (7) Figure 2. Illustration of Doppler-angulation. The S is the spinning beaon or a wave soure rotating ounterlokwise around the origin O at a onstant angular veloity ω with a rotational radius r. To maintain generality, the polar angle of S is denoted as θ(t)=ωt+φ. Note that the uuur polar angle of S is the angle inluded by OS and the positive x-axis. At time t, the position vetor of S is: uuur OS = (r os( ωt+ ϕ ), r sin( ωt+ ϕ )) The tangent line veloity vetor v(t) of the beaon S is v (t) = (-ω r sin( ωt+ ϕ ), ωr os( ωt+ ϕ )) (4) Aording to Equation (1), the observed Doppler shift is proportional to the projeted veloity on the line between the spinning beaon and the target. In Figure 2, the vetor of the target-beaon onnetion is SX uuur, so the veloity projeted on SX uuur is the inner produt of v(t) and the unit vetor of SX uuur : uuur SX v projet = v(t) uuur (5) SX Equations (2), (3) and (4) are then substituted into Equation (5) to obtain the projeted veloity: (3) If the rotational radius is minor in omparison to the distane from origin O to target X, namely r/d approximates 0, the Doppler shift an be formulated as: -ωrf sin( ωt+ ϕ- α ) lim f = lim r / d 0 r / d (r/d) -2(r/d)os( t+ - ) -ωrf = sin( ωt+ ϕ- α ) ω ϕ α The effet of this approximation error on the orientation angle estimation an be mitigated by our system. This mitigating effet is disussed further in Setion 4 and Appendix A. Aording to Equation (4), when r/d approahes 0, the Doppler shift produes a sine wave with a phase left-shift by an angle (φ-α). This sine wave is referred to as the frequeny waveform. If the Doppler shift an be measured, the angle (ωt+φ-α) an be determined. However, as Figure 3 shows, due to the diffiulty of measuring t and φ, an additional referene node with known loation is deployed to derive angle (ωt+φ-β). In Figure 3, R is the referene node, and X is the target. Polar angles of X and R are denoted as α and β, respetively. Subtrating (ωt+φ-α) from (ωt+φ-β) obtains (α β). Identifying the loation of R reveals the angle β, whih then reveals the diretion of X. (8) v projet uuur uuur uuur SX OX -OS = v(t) uuur = v(t) uuur uuur SX OX -OS = (-ω r sin( ωt+ ϕ ), ωr os( ωt+ ϕ )) (d os α, d sin α ) - (r os( ωt+ ϕ ), r sin( ωt+ ϕ )) (d os α, d sin α ) - (r os( ωt+ ϕ ), r sin( ωt+ ϕ )) = = -ωrd sin( ωt+ ϕ )osα + ωrd os( ωt+ ϕ ) sinα + ω ϕ α ω ϕ α 2 2 d r -2dr os( t+ ) os -2dr sin( t+ ) sin -ωrd sin( ωt+ ϕ- α ) d r -2dros( t+ - ) ω ϕ α (6) Figure 3. Illustration of Doppler-angulation. The referene node R is deployed to obtain the diretion of X.

4 2.3 Loalization Algorithm Figure 4 shows the loalization algorithm where X is the target and S 1 and S 2 are two spinning beaons at known loations. One the angles θ 1 and θ 2 are measured, the target an be loated at the intersetion of the two orientation lines. The orientation line extends from the spinning beaon outward in the diretion of θ 1 and θ 2. To enhane auray in a noisy environment, a robust positioning system would usually require more than two spinning beaons. If the angle measurement is noiseless, all orientation lines would ideally interset at only one point. However, perfet sensor auray is unlikely under real world onditions. Therefore, the atual intersetion may resemble that in Figure 4(b). The problem of identifying the most probable loation of the target was thoroughly explored in [14][15]. are identified, the orientation line passing through/nearby X an be obtained. (4) Loation estimation: Figure 5() shows that after repeating steps (1) ~ (3) for S 2 and S 3, two additional blue orientation lines passing through/near the position of X are obtained. The loation of X is then estimated by finding the intersetion of these three lines. Eah of these four steps is desribed in detail in the following subsetions. Figure 4. Loalization algorithm loates a target at the intersetion of multiple orientation lines. 3. SPINLOC POSITIONING SYSTEM Figure 5 illustrates an example of a deployed SpinLo positioning system. The example deployment in Figure 5(a) ontains the following sensor nodes: (1) target X, (2) referene node R and (3) three spinning beaons S 1, S 2 and S 3. Sine S 1, S 2, S 3 and R are stationary infrastruture nodes at known loations, the SpinLo positioning system must loate target X. The SpinLo loalization proess involves the following four steps: (1) Doppler shifted signal generation: While spinning, S 1 transmits RF signals at a onstant frequeny. X and R then measure Doppler shifted signals resulting from the spinning motion of the transmitter S 1. (2) Frequeny reord: Both X and R reord the Doppler shifted signals reeived from S 1 and then send their observed frequeny waveforms to the base station. (3) Orientation angle alulation: As desribed above, the base station applies the Doppler-angulation method whih takes the observed frequeny waveforms of X and R and alulates their orientation angle to S 1. As Figure 5(b) shows, sine the loations of both S 1 and R Figure 5. Deployment of the SpinLo positioning system: (a) deployment layout, (b) angulation phase and () loalization phase.

5 3.1 Doppler Shifted Signal Generation The first step is to generate Doppler shifted signals. A beaon embedded with a RF transeiver module transmits radio signals while spinning. However, a typial RF arrier frequeny in the 400 MHz ~ 2.4 GHz range is too high for analysis by a hardware-onstrained sensor node due to its slow lok and limited sampling rate. The radio interferometry method developed by Maroti et al. [9] is therefore used to overome suh limitations. Radio interferometry measures RF Doppler shift with suffiient auray using inexpensive sensor hardware suh as MICA2 motes and is used in the system as follows. First, two sensor nodes simultaneously transmit sine waves at two very similar high radio frequenies f 1 and f 2. The similarity of the two RF signals produes an interferene RF signal with a low frequeny envelope f 1 -f 2. For example, two high frequeny 900 MHz radios an transmit sine waves at slightly different frequenies to produe a low interferene frequeny envelope below 1 khz, whih is slow enough for analysis by a MICA2 mote. Radio interferometry requires simultaneous transmissions from two radios. Therefore, as Figure 6 shows, a stati assistive beaon is required. The spinning beaon transmits RF signals at its original frequeny (f 1 ) and any reeiver P will pereive radio frequeny at f 1 plus a Doppler shift ( f 1 P ). The assistive beaon simultaneously emits signals at a fixed frequeny (f 2 ). Sine the assistive beaon is stationary, the interferene frequeny, i.e., f 1 -f 2 + f 1 P, is affeted only by the Doppler shift in the signal frequeny of the spinning beaon. Fine-tuning the Doppler shifted interferene frequeny f 1 -f 2 + f 1 P to a low frequeny range under 1 khz enables detetion and analysis of the signal by a MICA2 mote. Figure 6. An example deployment of SpinLo system. An assistive beaon is added after initial deployment. 3.2 Frequeny Reord The seond step is for X and R to reord their reeived Doppler shifted frequenies from S 1. X and R reeive radio frequenies and analyze them by RSS analog-to-digital onverter (ADC) on sensor nodes. The real-time time-domain frequeny estimation method proposed by Maroti et al. [9] is used to detet the RSS peak timestamps and estimate the interferene frequeny. For example, if the RSS sampling rate is 17,800 Hz and the number of samples between adjaent RSS peaks is 30, the interferene frequeny is approximately 17,800/30 = Hz. Reeiver nodes X and R send their RSS peak timestamps to the base station where the timestamps are then analyzed to reonstrut their frequeny waveforms. 3.3 Orientation Angle Calulation The third step is to alulate the orientation angle using the frequeny waveforms of X and R. Speifially, as Figure 5(b) shows, this orientation angle is the angle between nodes R and X relative to spinning beaon S. A simple method of alulating this orientation angle is to use the Doppler-angulation method desribed in subsetion 2.2. First, by using radio interferometry, we measure the frequeny shift value pereived by X and R. Seond, this frequeny shift value is used to determine angle (α β) in Figure 3. If X and R reveal Doppler shifts f x and f R, respetively, at time t, (α β) an be alulated as follows. First, f x is substituted into Equation (8): -ωrf f x = sin( ωt+ ϕ- α ) - f x = sin( ωt+ ϕ- α ) ωrf By taking the arsine funtion on both sides of the Equation (9), α is determined: - f ωrf ω ϕ α -1 x sin ( ) = t+ - α ω ϕ - f ωrf -1 x = t+ - sin ( ) Similarly, β is omputed as follows: β ω ϕ - f ωrf -1 R = t+ - sin ( ) (9) (10) (11) The angle (α β) is obtained by simply subtrating Equation (11) from Equation (10): α β f ωrf f ωrf -1 R -1 x - = sin ( )- sin ( ) (12) Although the above solution is simple and mathematially sound, one pratial problem arises in an atual working system and environment. The radio on a MICA2 mote an only be tuned at a resolution of 65 Hz. Restated, the atual transmitted and reeived frequenies would have errors of plus/minus 65 Hz. These errors redue preision when determining the value of atual transmitted frequeny f on the spinning beaon S. On the reeiver nodes X and R, these errors also ause impreise measures of Doppler shifts f x

6 and f R. In other words, although using a finer resolution and more expensive radio hip on a beaon (infrastruture) ould improve the auray of atual transmitted frequeny f from S, the measured Doppler shifts, i.e., f x on X and f R on R, would remain impreise. Instead, a more preise measure is the overall shifted interferene frequeny, whih is the sum of the base interferene frequeny and Doppler shift. Although Equation (12) annot diretly solve (α β) with the Doppler shift at some instant t, analysis of Doppler frequeny shifts over a given time period an solve (α β) more robustly. Aording to Equation (8), f x and f R an be regarded as funtions of time t as follows: -ωrf f x(t) = sin( ωt+ ϕ- α ) -ωrf f R(t) = sin( ωt+ ϕ- β ) (13) As Equation (13) shows, f x (t) and f R (t) are both sinusoidal waves with the same periodiity but with a relative phase offset (α β). Figure 7 shows the atual measured interferene frequeny shifts f x (t) and f R (t) where y-axis is the Doppler shifted interferene frequeny. Further observation reveals that the problem of solving for (α β) is equivalent to identifying the phase shift between the measured f x (t) and f R (t). Figure 7. Example of observed Doppler shifted frequenies. A ommon problem in digital signal proessing is identifying phase shift in two disrete-time noisy sinusoid signals with the same periodiity [16]. Let X[n] and Y[n] denote two disrete-time noisy sinusoid signals with the same period N. A ommon solution is to delay Y[n] by time k and ompare the similarity between X[n] and Y[n-k], where Y[n-k] is the same as Y[n] with a right shift of time k. The phase offset an be alulated from time k suh that X[n] is most similar to Y[n-k] as follows: k 2 π ( mod 2 π ) (14) N The sum squared differenes (SSD) [16] is seleted as the distane funtion required to quantitatively express the similarity of two signals. The SSD of X[n] and Y[n-k] is defined as: n= - ( X[n]-Y[n-k] ) 2 (15) In any implementation, one an only observe the shifted frequenies for a limited time period T. Therefore, SSD is normalized by the overlap in duration of eah delay unit k: D [k] = XY T -1 2 ( X[n]-Y[n-k] ) (16) n=k T-k The square root is alulated so that D XY [k] aurately defines distane. The resulting k value is suh that D XY [k] is the minimum for k=0 to T-1. Exessive noise in f x (t) and f R (t) may yield unreasonably high D XY [k] values. Therefore, to avoid false estimations under these onditions, a threshold is applied to filter estimations from exessively noisy signals. The effet of the filter at different thresholds is evaluated in Setion 6. The alulated phase offset yields the orientation line for loation estimation. 3.4 Loation Estimation Sine three spinning beaons are used in the example deployment, Doppler-angulation alulates three orientation angles and three orresponding lines toward the target. However, as Figure 4(b) shows, varying angular errors affet varying shifts in these orientation lines away from the target. These shifts bring the three orientation lines to interset at three different points rather than at the same point. Given multiple intersetion points, a target an be loated by several different algorithms. For example, the Centroid algorithm loates the target at the Centroid of these three intersetion points. SpinLo proposed here uses the Weighted Centroid algorithm, whih weights eah intersetion point aording to an internally omputed onfidene value. This onfidene value is proportional to the auteness of the intersetion angle between two interseting orientation lines beause the auteness angle affets the sensitivity of angular error on positional error. For example, as Figure 8 shows, when the intersetion angle is lose to 0 or 180 degrees, even a minor angular error an produe a large positional error. Note that the x-axis and y-axis error sensitivities for eah intersetion point differ and are onsidered separately.

7 Figure 8. Example of a minor angular error ausing a large positional error. 4. PARAMETRIC TUNING AND PER- FORMANCE TRADEOFF Different value settings of tunable parameters determine overall system performane. The main system performane metris onsidered are positional auray and positioning lateny. Positional auray measures the differene between an estimated loation and a ground-truth loation. Positioning lateny is the time required for SpinLo to measure and estimate a position. The tunable system parameters are (1) Doppler-angulation filtering threshold, (2) data olletion time, (3) rotational veloity, (4) interferene frequeny and (5) rotational radius. Generally, tuning these parameters improves auray but at the expense of prolonged lateny, and vie versa. Restated, adjusting these parameters determines the performane tradeoff between auray and lateny. (1) Doppler-angulation filtering threshold: The delay-and-ompare proess desribed in subsetion 3.3 has the following two outputs: estimated time delay k and minimum distane D XY [k] between two input signals. The minimum distane reveals the quality of the estimation. A smaller minimum distane indiates a more reliable time delay estimation. Therefore, before using the measured angular information to loalize the target, poor angular estimation is filtered out as noise aording to the minimum distane. Sine the striter filtering threshold removes more noise from the dataset, it produes a more aurate angulation result. Although a striter filtering threshold inreases overall positioning lateny, it also prolongs the positioning lateny to aquire enough high-quality orientation angle estimates. (2) Data olletion time: A longer data olletion time inreases the number of signal samples, whih enables more preise reonstrution of frequeny waveforms and improves positional auray. However, the tradeoff is inreased positioning lateny. Notably, an overly long data olletion time has adverse effets suh as inreased arrier frequeny drift and lok drift. (3) Rotational veloity: An instantaneous radio frequeny is diffiult to detet when rotational veloity is high beause of the shorter time that a rapidly spinning beaon is positioned in a speifi diretion. Faster rotation redues the number of signal samples olleted and thus redue the preision of frequeny waveform detetion, whih inreases positional error. However, high rotational veloity improves positioning lateny beause it provides more signal samples for Doppler-angulation. Note that a very slow rotational veloity may also produe a large positional error beause frequeny waveforms of small Doppler shifts are more diffiult to distinguish. (4) Interferene frequeny: Beause of the Doppler Effet and the harateristis of radio interferometry, a higher interferene frequeny inreases the Doppler shift, whih also improves the signal-to-noise ratio (SNR) of the frequeny waveform. However, a high interferene frequeny also has ertain disadvantages. For example, if the interferene frequeny is 500 Hz and the signal sampling rate is 17,800 Hz, the number of signals sampled per period is 17,800/500 = When the interferene frequeny inreases to 1,000 Hz, the number of signals sampled per period is halved to 17,800/1,000 = Fewer signal samples degrade the preision of the frequeny detetion in radio interferometry, thus affeting the positional auray of the SpinLo system. (5) Rotational radius: A long rotational radius (r) and a high angular veloity (ω) of a spinning beaon produe large Doppler shifts beause, as Figure 2 shows, the tangent line veloity (v(t)) is the produt of r and ω. Further, large Doppler shifts produe more distinguishable frequeny waveforms, whih improve the SNR of the frequeny waveform and redue positional error. However, a long rotational radius also has disadvantages. A long rotational radius (r) not only requires a larger physial spae, it also, as Equation (8) shows, inreases the approximation error in Equation (8) as the r/d ratio inreases. Fortunately, the effet of the approximation error is mitigated by the proposed method of estimating the orientation angle. The proposed method of identifying phase shift in two Doppler shifted frequeny signals is not sensitive to r/d ratios. Varying the r/d pairs numerially simulates two non-approximated waveforms from Equation (7) when performing the orientation angle alulation, where r/d ratio pair (0, 0) stands for the approximated ase. The analytial results show that, whether or not they are approximated, the alulated orientation angles are idential. Appendix A provides a detailed example of the alulation. These parametri tradeoffs are experimentally analyzed in Setion IMPLEMENTATION The proposed SpinLo system was implemented on MI- CA2 motes manufatured by Crossbow In. These MICA2 motes ran TinyOS. One MICA2 mote was onneted to a base station via MIB520 programming board to relay pakets ontaining frequeny waveforms from infrastruture nodes and the target. These MICA2 motes were programmed to emit pure sine waves, whih produed Doppler shifts when transmitted from the rotating beaons. The

8 RIPS engine developed by Vanderbilt University [17] was modified and ported to 900 MHz MICA2 motes. At our test site loation in Taiwan, the frequeny band of GSM-900 [18] also happens to be around 900MHz, overlapping with a part of MICA2 radio hannels. To avoid interferene from the GSM-900 up/down link hannels, we seleted 18 arrier frequenies between MHz to MHz whose ranges are away from GSM-900 hannels. Additionally, as Figure 9 shows, eah spinning beaon was mounted on the arm of a rotation devie. The rotation dean attahed arm vie, onsisting of a motor for spinning and a ontrol unit with adjustable rotational veloity, was seurely mounted on a steel platform. The ost of eah ro- Industry-grade tation motor/devie is about 100 US dollars. servo motors, suh as [19] with almost one million hours MTBF (Mean Time Between Failures), an also be used to build robust rotation devies. Changing rotational veloity produed varying Doppler shifts. In eah measurement round, the base station transmits a ommand to all sensor nodes. After performing time synbeaons transmit hronization, the spinning and assistive radio signals while the referene node and the target log and send reeived signal frequenies to the base station. After reeiving the signal frequenies, the base station al- beaon and ulates the orientation angle for eah spinning estimates the position of the target. The infrastruture omponents inluded three spinning beaons, one referene node and one assistive beaon. Figure 11 shows the positions of these infrastruture nodes. The grid points on the map are 2 meters apart. To measure positional auray throughout this test environment, a tar- grid points on the get was moved among the thirty different map. Figure 10. Indoor parking garage environment with infrastruture node deployment. Figure 11. Grid map of infrastruture node positions. Figure 9. Rotation devie onsisting of motor, speed ontrol unit and steel platform for stability. 6. EXPERIMENTAL RESULTS As Figure 10 shows, the system was tested in an indoor parking garage in the basement of a university building. The parking garage had a eiling height of 3 meters, and MICA2 motes were deployed over an 8 10 m 2 area. 6.1 SpinLo Positional Errors Figure 12 shows the umulative density funtion (CDF) of the positional errors. The parametri settings were as fol- angular veloity lows: rotational radius was 50 entimeters, was 133 revolutions per minute (RPM), interferene freolletion time was 1.5 queny was 600 Hz, and data seonds. The median positional error was 39 entimeters, and the 90% error is 70 entimeters (meaning 90% of errors were 70 entimeters or less). The test results demonstrated that the SpinLo system ahieves sub-meter positional a- uray in an indoor environment.

9 Figure 13 depits the umulative density funtion (CDF) of absolute angular error with a median error of 3 degrees and 90% of errors 10 degrees or less. The dataset for the CDF was based on 300 sample position estimates over the thirty grid points (ten samples per grid point). Sine eah positioning sample is derived from three estimated angles, 900 angle estimates were sampled. Although some angular errors were seemingly large, their effets on positional auray were mitigated in the loalization phase beause these large errors ould be identified as outliers by analyzing the intersetion of multiple orientation lines measured by multiple spinning beaons. Figure 14 shows the probability distribution funtion (PDF) of angular errors from the same dataset in Figure 12 without taking absolute values. The figure shows that angular errors were equally distributed between positive and negative. Figure 14. PDF of angular errors. 6.2 Doppler-Angulation Filtering Figure 15 shows the relationship between angular errors and filtering thresholds from 5 (striter) to 15 (less strit) as well as the relationship between data redution % and filtering thresholds. Clearly, without the noise filtering, angular errors were relatively large (see No filter in the x-axis of Figure 15). However, at the stritest noise filtering threshold, the system removed almost 80% of the reeived pakets. Our system was deployed with a filtering threshold value set at 10, whih was suffiient to redue most of noisy data. Figure 16 shows the orresponding CDF of angular errors under different filtering thresholds. Figure 12. CDF of positional errors. Figure 15. Angular errors and filtered out data ratios under different filtering thresholds from none to 5. Figure 13. CDF of angular errors.

10 Figure 16. CDF of the angular errors under different filtering thresholds. The lines from right to left are in the order from No filter (top) to 5 (bottom) in the legends. Figure 18. CDF of the positional errors under different data olletion times. The lines from right to left are in the order from 0.3s (top) to 1.5s (bottom) in the legends. 6.3 Data Colletion Times Figure 17 shows the positional errors at different length of data olletion times (0.3 ~ 1.5 seonds). Two urves plot the median positional errors and 90% positional errors. The analytial results show that a longer data olletion time generally redues positional error beause the inreased number of signal samples enables more aurate reonstrution of frequeny waveforms. At the shortest data olletion time of 0.3 seond, the SpinLo system still ahieved sub-meter auray with 90% positional error of 93 entimeters. Figure 18 depits the orresponding CDF of positional errors for different data olletion times ranging from 0.3 to 1.5 seonds. 6.4 Rotational Veloities Figure 19 shows the angular errors of different rotational veloities of the spinning beaon varying from 92 to 200 RPM. The RPM range ahieved the best auray with a median error of about 3 degrees and a standard deviation of approximately 1 degree. As desribed in Setion 4, exessively fast veloities are undesirable due to the insuffiient number of signal samples for reonstruting frequeny waveforms, and exessively slow rotational veloities are also undesirable due to the less distinguishable frequeny waveforms from the smaller Doppler shifts. Figure 17. The median and 90% positional errors under different data olletion times from 0.3 to 1.5 seonds. Figure 19. Angular errors under rotational veloities varying from 92 to 200 RPM.

11 6.5 Interferene Frequeny Figure 20 shows the angular errors under different interferene frequenies (100 ~ 1,200 Hz). The interferene test results suggested that the optimal interferene frequeny range is between 200 Hz and 1,000 Hz. The proposed system used the interferene frequeny of 600 Hz, whih is in the middle of this range. Figure 20. Angular errors under different interferene frequenies from 100 Hz to 1,200 Hz. 7. ERROR SOURCES The errors were due to varying fators suh as MICA2 mote hardware limitations, the inevitable indoor multipath interferene and errors inherited from the radio interferometry tehnique. These error soures are disussed further below. (1) Time synhronization error: Time synhronization errors an be ontrolled by lok rate. Although SpinLo eliminates the need to synhronize the spinning beaon with reeiver nodes, the reeiver nodes must still be synhronized with eah other so that frequeny waveforms from different reeiver nodes an be ompared and analyzed to determine their phase shift. When the data olletion time is muh longer than the synhronization error, this soure of error an be minimized. Clok drift may ontribute to positional error beause synhronization is performed only at the beginning of eah data olletion yle. (2) Carrier frequeny drift: This error may be aused by the MICA2 motes hardware, as desribed in [9]. Although reduing the data olletion time an minimize frequeny drift, the number of signal samples olleted may be insuffiient for preise reonstrution of frequeny hange waves. However, sine the system analyzes Doppler waveforms rather than atual frequenies, arrier frequeny drift has only minor effets on system performane. (3) Indoor multipath interferene: An advantage of working in the frequeny domain is that refletion wave remains the same frequeny as the inident wave. However, this advantage applies only to a stati wave soure. When the wave soure is moving, Doppler Effet auses the line-of-sight signal and the multipath signals to be at slightly different frequenies. Therefore, the estimated frequeny is still affeted by the indoor multipath interferene. 8. RELATED WORK The most relevant work is the frequeny-differene-of-arrival (FDOA) tehnique. The FDOA is analogous to time-differene-of-arrival (TDOA) for estimating the loation of a radio emitters based on observations from several known points. Suh a tehnique was employed by Kusy et al. [13] to trak mobile targets by using infrastruture nodes as reeivers and targets as beaon emitters. When the target is moving in suh a system, eah infrastruture reeiver observes varying Doppler shifts depending on their positions relative to the moving target. The target loation and veloity are then solved by onstrained non-linear squares (CNLS) optimization given the input parameters of geometri relationships between nodes, different Doppler shift measurements and the target veloity vetor. However, due to signal noise, Kusy reported aurate results in veloity estimation but no positioning estimation. They therefore ombined the aurate estimated veloity vetor with the Extended Kalman Filter (EKF) in a traking task, whih ahieved an average loalization error of 1.5 to 3 meters in an outdoor environment. As work developed independently, Ledezi et al. [20] proposed a loalization method based on Doppler shifted RF signals from a rotating antenna. Their simulation showed potential for ahieving high positional auray below 10 entimeters. The underlying tehnique used by Kusy et al. [13] to measure Doppler shifts with low-ost hardware is the radio interferometry approah used in RIPS [9]. The RIPS approah uses two nodes transmitting RF signals at slightly different frequenies to produe an interferene frequeny envelop at a low frequeny, whih an then be measured and analyzed by a low sampling rate ADC on a MICA2 mote. The RIPS measures the relative phase offset of the reeived interferene signals to obtain q-range information, whih is the linear ombination of distanes between the two radio transmitters and the two reeivers. Wu et al. [21] developed the adaptive RIPS that enhaned the positional auray of RIPS by dynamially seleting anhor nodes as beaon senders. Many other proposed sensor network positioning systems an be broadly lassified as ranging-based and ranging-free methods. (1) Ranging-based methods. These methods ommonly require signal ommuniations between an anhor observer and a loating target. The major differenes among them are the varying alibration methods and the use of different signal soures, suh as soni, ultrasoni, infrared, amera, RF, et. For example, Aousti ENSBox [7] employs a distribution aousti sensing platform whih enables rapid deployment and self-alibration of an aousti embedded

12 networked sensing box. The system an reportedly ahieve positional auray to within 5 entimeters in a partially obstruted m 2 outdoor environment. Given that signals propagate at onstant veloity, time of-arrival (TOA) methods [22][23] estimate distane by measuring signal propagation time. Angle-of-arrival (AOA) [24] is a network-based solution that exploits the geometri properties of the arriving signal. By measuring the angle at whih the signal arrives at multiple reeivers, the system an aurately estimate loation. The TDOA [25] is another network-based system whih infers distane by measuring time differenes. Some hybrid approahes of TOA, AOA, and TDOA have also been proposed [26]. Another lass of tehniques measures the RSS. These tehniques exploit the deaying model of eletromagneti fields to translate RSS into a orresponding distane [23][27][28]. The frequeny bands used for transmission may also vary. For example, the well-known RADAR system [29] uses RF, and LADAR and SONAR use visible light and audible sound bands, respetively. Other systems suh as LADAR and SONAR analyze signals refleted from an objet to estimate its loation. A reent innovation, Criket [8], employs a hybrid approah using both RF and ultrasoni bands. However, the propagation harateristis are irregular in atual outdoor environments [30]. Loalization systems using RSS information have similar limitations and usually ahieve only meter-level auray. (2) Range-free methods. These methods use other alternatives to range estimation between anhor nodes to loalize targets. For example, APIT [31] estimates the loation of targets based on the onnetivity information to anhor node with known loation. The greater the number of deployed anhor nodes, the greater the auray of the tehnique. Restated, auray highly depends on the density of deployed anhor nodes. One lass of tehniques detets sequenes of artifiially generated events from an event sheduler. For example, Spotlight [32] and Lighthouse [33] orrelate the event detetion time of a sensor node with known spatiotemporal relationships. The detetion events are then mapped to estimate position. However, generating and disseminating these events in a large-sale area is relatively diffiult, partiularly given the alibration requirements. 9. DISCUSSION The following aspets of SpinLo are disussed: (1) target node salability, (2) beaon node salability, and (3) 3D positioning. On the target node salability, the base station proessing time per target loalization is 0.47 seond. The urrent algorithm for omputing the target loalization an be further optimized and parallelized to redue proessing time, hene improves salability in the number of target nodes. On the beaon node salability, sine nearby beaons an interfere with eah other s transmission, SpinLo runs a time synhronization protool on nearby beaons and shedules their transmissions in different time slots to avoid signal interferene. This means that inreasing the beaon density improves positional auray at the ost of prolonged positional lateny. More omplex sheduling algorithms an be used to improve beaon node salability. Although the urrent SpinLo system loates targets only in 2D, it is possible to extend SpinLo to 3D by adding additional beaons that rotate on the vertial plane. Analyzing produed Doppler shifted signals on the vertial plane yields the height of a target. 10. CONCLUSIONS AND FUTURE WORK The proposed SpinLo system presents a novel indoor loalization method that overomes indoor multipath interferene and ahieves sub-meter positional auray. The experimental results in an indoor garage environment ahieved a median positional error of 39 entimeters and a 90% positional error of 70 entimeters. By using spinning beaons to produe preditable and distinguishable Doppler Effets, the SpinLo system ahieves sub-meter loalization auray. Additionally, SpinLo is highly ost effetive sine its deployment requires only inexpensive hardware available on MICA2 motes. Our future work will explore more harateristis and further refinements of the SpinLo system. (1) We are interested in onduting a systemati evaluation to explore the potential relationship between the target s loation (e.g., its proximity to a wall, its relative distanes to nearby beaons, et.) and its positional error. (2) Although the urrent system is effetive for traking stationary or slow-moving targets, it may not work well for traking fast-moving targets. Hene, we are interested in developing a method that an ompensate for the inreased Doppler shifts from fast-moving targets. (3) A typial offie deployment is likely to have obstrutions, suh as furniture. We are interested in evaluating the effet of different types of obstrutions on the SpinLo positional auray. (4) It is important to design and mask the exterior of these spinning beaons suh that their appearane beomes less obtrusive in an offie setting. For example, it may be possible to mask a spinning beaon on a eiling fan. REFERENCES [1] A. Harter, A. Hopper, P. Steggles, A. Ward, and P. Webster, The anatomy of a ontext-aware appliation, in Pro. of 5th ACM International Conferene on Mobile Computing and Networking (Mobiom), August [2] J. Hightower and G. Borriello, Loation systems for ubiquitous omputing, IEEE Computer, vol.34, no.8, pp.57-66, August [3] Ubisense. [4] M. Tuhler, V. Shwarz, and A. Huber, Loation auray of an UWB loalization system in a multi-path

13 environment, in Pro. of IEEE International Conferene on Ultra-Wideband (ICUWB), September [5] S. Gezii, Z. Tian, G. B. Giannakis, H. Kobayashi, A. F. Molish, H. V. Poor, and Z. Sahinoglu, Loalization via ultra-wideband radios: a look at positioning aspets for future sensor networks, IEEE Signal Proessing Magazine, July [6] J. Sott and B. Dragovi, Audio loation: aurate low-ost loation sensing, in Pro. of 3rd International Conferene on Pervasive Computing, May [7] L. Girod, M. Luka, V. Trifa, and D. Estrin, The design and implementation of a self alibrating aousti sensing platform, in Pro. of 3rd ACM International Conferene on Embedded Networked Sensor Systems (SenSys), Otober [8] N. B. Priyantha, A. Chakraborty, and H. Balakrishnan. The Criket loation-support system, in Pro. of 6th ACM International Conferene on Mobile Computing and Networking (MobiCom), August [9] M. Maroti, B. Kusy, G. Balogh, P. Volgyesi, A. Nadas, K. Molnar, S. Dora, and A. Ledezi, Radio interferometri geoloation, in Pro. of 3rd ACM International Conferene on Embedded Networked Sensor Systems (SenSys), November [10] B. Kusy, M. Maroti, G. Balogh, P. Volgyesi, J. Sallai, A. Nadas, A. Ledezi, and L. Meertens, Node density independent loalization, in Pro. of 5th International Symposium on Information Proessing in Sensor Networks (IPSN/SPOTS), April [11] B. Kusy, G. Balogh, A. Ledezi, J. Sallai, and M. Maroti, intrak: High preision traking of mobile sensor nodes, in Pro. of 4th European Workshop on Wireless Sensor Networks (EWSN), January [12] B. Kusy, J. Sallai, G. Balogh, A. Ledezi, V. Protopopesu, J. Tolliver, F. DeNap, and M. Parang, Radio interferometri traking of mobile wireless nodes, in Pro. of 5th International Conferene on Mobile systems, appliations and servies (MobiSys), June [13] B. Kusy, A. Ledezi, and X. Koutsoukos, Traking mobile nodes using RF Doppler shifts, in Pro. of 5th ACM International Conferene on Embedded Networked Sensor Systems (SenSys), November [14] C. D. MGillem and T. S. Rappaport, A beaon navigation method for autonomous vehiles, IEEE Transations on Vehiular Tehnology, vol.38, no.3, pp , August [15] A. Nasipuri and R. el Najjar, Experimental evaluation of an angle based indoor loalization system, in Pro. of 5th International Symposium on Modeling and Optimization in Mobile, Ad Ho and Wireless Networks (WiOpt), April [16] F. Viola and W.F. Walker, A omparison of the performane of time-delay estimators in medial ultrasound, IEEE Transations on Ultrasonis, Ferroeletris and Frequeny Control, vol.50, no.4, pp , April [17] rib/vu/apps/ripsonehop/ [18] 3rd Generation Partnership Projet, 3GPP TS 05.05, [19] Tamagawa Seiki Co., Ltd. or/resolver.html [20] A. Ledezi, P. Volgyesi, J. Sallai, and R. Thibodeaux, A novel RF ranging method, in Pro. of Sixth Workshop on Intelligent Solutions in Embedded Systems (WISES08), July [21] H. Wu, H. Chang, C. You, H. Chu, and P. Huang, Modeling and optimizing positional auray based on hyperboli geometry for the adaptive radio interferometri positioning system, in Pro. of 3rd International Workshop on Loation- and Context-Awareness (LoCA), September [22] F. Izquierdo, M. Ciurana, F. Barelo, J. Paradells, and E. Zola, Performane evaluation of a TOA-based trilateration method to loate terminals in WLAN, in Pro. of 1st IEEE International Symposium on Wireless Pervasive Computing, January [23] N. Patwari, A. O. Hero III, M. Perkins, N. S. Correal, R. J. O'Dea, Relative loation estimation in wireless sensor networks, IEEE Transations on Signal Proessing, vol. 51, no. 8, pp , August [24] N. Dragos, and B. Nath, Ad ho positioning system (APS) using AoA, in Pro. of 22nd IEEE International Conferene on Computer Communiations (InfoCom), April [25] A. Savvides, C. C. Han, and M. B. Srivastava, "Dynami fine-grained loalization in ad-ho networks of sensors," in Pro. of 7th ACM International Conferene on Mobile Computing and Networking (Mobi- Com), July [26] L. Cong and W. Zhuang, Hybrid TDOA/AOA mobile user loation for wideband CDMA ellular systems, IEEE Transations on Wireless Communiations, vol. 1, no. 3, pp , July [27] D. Niulesu, Positioning in ad ho sensor networks, IEEE Networks, vol. 18, no. 4, pp , July [28] K. Lorinz and M. Welsh, Motetrak: a robust, deentralized approah to RF-based loation traking, in Pro. of 1st International Workshop on Loation- and Context-Awareness (LoCA), May 2005.

14 [29] P. Bahl and V. Padmanabhan, RADAR: an in-building RF-based user loation and traking system, in Pro. of 19th IEEE International Conferene on Computer Communiations (InfoCom), Marh [30] G. Zhou, T. He, and J. A. Stankovi, Impat of radio irregularity on wireless sensor networks, in Pro. of 2nd ACM International Conferene on Mobile Systems, Appliations, and Servies (MobiSys), June [31] T. He, C. Huang, B. M. Blum, J. A. Stankovi, and T. Abdelzaher, Range-free loalization shemes in large-sale sensor networks, in Pro. of 9th ACM International Conferene on Mobile Computing and Networking (MobiCom), September [32] R. Stoleru, T. He, J. A. Stankovi, and D. Luebke, A high-auray, low-ost loalization system for wireless sensor networks, in Pro. of 3rd ACM International Conferene on Embedded Networked Sensor Systems (SenSys), November [33] K. R omer, The lighthouse loation system for smart dust, in Pro. of 1st ACM International Conferene on Mobile Systems, Appliations, and Servies (MobiSys), May sin( t - 150) f (t)= sin( t - 50) f (t)= (0.8) -2 (0.8) os( t - 150) 2 1+(0.2) -2 (0.2) os( t - 50) (17) The two waveforms are onstruted by substituting the following parameters into Equation (7) : o 360 r f 1(t) : ωrf/ = 25, ω =, = 0.8, ϕ o α1= ms d o 360 r f 2(t) : ωrf/ = 25, ω =, = 0.2, ϕ α2 = ms d o (18) The gray waveform in Figure 21 is the first shifted frequeny waveform orresponding to Doppler shift f 1 (t), and the blak waveform orresponds to f 2 (t). Note that the y-axis is the shifted frequeny, whih is the sum of the base frequeny 600 Hz and the Doppler shift. Sine r/d of the f 1 (t) is 0.8, whih is far from zero, the waveform is quite different from a perfet sine wave. APPENDIX A. Figure 22. Distane between two signals in Figure 21 for different time delays. Figure 21. Non-approximated shifted frequeny waveforms. Figure 21 shows the simulation results for varying r/d ratios mentioned in Setion 4. Two atual Doppler shifted frequeny waveforms are onstruted as follows: The delay-and-ompare method mentioned in subsetion 3.3 is used to estimate the phase shift of the two waveforms. Aording to the simulation result in Figure 22, the minimum distane ours when delay = 2,222 milliseonds. The orresponding minimum distane is 5.99 Hz. Sine the period of both waveforms is 8,000 milliseonds, the estimated relative phase offset is (2,222/8,000)*360 = 100 degrees. The result equals the exat phase offset of the two approximated sinusoids. Further simulations with two r/d ratios varying from 0 to 1 reveal estimated phase offsets idential to the phase offsets used to generate the waveforms. This suggests that the approximation in the derivation of orientation angle alulation does not affet the loation estimation error beause the phase shift estimation method is not sensitive to the ratio of r/d.