Accuracy of a Commercial UWB 3D Location/Tracking System and its Impact on LT Application Scenarios

Transcription

1 2 IEEE International Conference on Ultra-Wideband Paper # Accuracy of a Commercial UWB 3D Location/Tracking System and its Impact on LT Application Scenarios Volker Schwarz, Alex Huber, and Michael Tüchler, Members, IEEE Abstract A newly available commercial Ultra-Wideband 3D location and tracking system has been characterized in terms of measurement accuracy and disturbance of the line-ofsight path between the located tag and the receiving sensors. While the standard deviation of a large number of position readings is 1 cm or better, bias on the order of meters often occurs that we address to reflections at metal walls. A human torso or hand blocking the line of sight prevents location. The results measured in our class room environment indicate that the majority of presumed applications for UWB location and tracking are not yet feasible with available LT technology. Index Terms Delay estimation, Direction of arrival estimation, Position measurement, Ultra-Wideband U I. MOTIVATION AND INTRODUCTION LTRA-WIDEBAND Impulse Radio (IR) technology promises feasibility of Low Data Rate (LDR) applications with precise Location and Tracking (LT) capability. Several market scenarios for UWB-LT with or without data communication have been published. An extensive list of such scenarios is available at the IEEE a WPAN Task Group [1] or at the EU FP6 Integrated Project "PULSERS" [2]. When collecting and categorizing these applications, basically two main scenarios can be found: Asset tracking or machine guidance in industrial environments and person tracking in office or outdoor environments. Specific UWB 3D location technology is currently getting commercially available. The prices, form factors and installation procedures are not yet suited for the mass market but they are intended to demonstrate the feasibility of UWB LT technology. One of these commercially available UWB 3D LT systems has been installed in a class room environment at the "Center of Microelectronics Aargau" (zma). The system operates with Position Modulation of UWB pulses in a frequency band from 7 GHz, a Pulse Repetition Frequency Manuscript received March 31, 2. V. Schwarz was with the center of microelectronics aargau, University of Applied Sciences Aargau, 21 Windisch AG, Switzerland. He is now with Huber+Suhner AG, 833 Pfaeffikon ZH, Switzerland ( [email protected]). A. Huber is with the center of microelectronics aargau, University of Applied Sciences Aargau, 21 Windisch AG, Switzerland ( [email protected]). M. Tuechler is with the center of microelectronics aargau, University of Applied Sciences Aargau, 21 Windisch AG, Switzerland (phone: ; fax: ; [email protected]). of approximately 12 MHz and is certified by the US FCC. The promised location accuracy of this system is 1 cm without mentioning further conditions or measurement requirements. The maximum update rate is announced to be 39 Hz per location cell, but the update rate for a single tag is at most 1 Hz. We have measured the received signal strength and the retrieved coordinates of a tag located at positions distributed over our measurement environment. We wanted to characterize the system and get a more profound statement on the achievable location accuracy. During the measurements we made some observations from which we think that they are generic for UWB location technology. Therefore, we will judge the feasibility of published applications for UWB LT on the basis of our observations and propose countermeasures for improved performance. II. SETUP A. LT System The location and tracking system comprises four sensors and a Windows PC for processing and management. The four sensors form a "location cell" for an unlimited number of tags. In our set, five tags have been delivered. Here, only one tag has been used, the others have been removed from the system to reduce the probability of misinterpretations and UWB radio noise. The sensor network uses Angle-of- Arrival (AoA) technique by means of phased array antennas at each sensor together with Time-Difference-of-Arrival (TDoA) between two sensors. Hence, 2 sensors are sufficient for the computation of a 3D coordinate if they can receive the signal transmitted by a particular tag. The tag comprises a flat UWB antenna oriented to the front side. The size of the tag is comparable to a cigarette packet. As a result, a hemispherical characteristic of the tag's transmit antenna can be assumed. Therefore, two sensors can detect the tag's location when the normal of the Tx antenna is perpendicular ± 4º to their base line. (Cf. Fig. 1: for sensors S and S1, the base line is parallel to the x-axis and the tag orientation angle ϑ must nominally be in the range [4º, 13º].) The additional two sensors of a location cell increase the coverage and reduce the orientation dependence of the tag's location. B. Room and Location Cell Our room for location experiments is a class room for computer education with a size of 9.6 m x 8 m x 2.8 m. A full metal H-beam extends from floor to ceiling in the center

2 2 IEEE International Conference on Ultra-Wideband Paper # of the room, here is the origin of the coordinate system as indicated in Fig. 1. The sensors are installed close to the ceiling with their antenna normal pointing to the center of the room at an angle (pitch) of about 2º with respect to the ceiling. The room is equipped with desks whose top has a height of 73 cm or 83 cm at certain locations. The desks in the left part of the room are provided with LCD screens. Particular is the presence of large areas of metal at about % of the wall area due to metal doors, metal light weight construction walls, a switch cabinet and large heating radiators. As a consequence, we have to cope with a highly reflective "multipath" environment for the UWB pulses. This is in contrast to the usual assumption of an office environment but representative for any kind of industrial building or warehouse. III. MEASUREMENT SETUPS A. Received Signal Strength The system provides a mode to display the received signal strength (RSS) of the pulses received from the tags. These values are provided 'as is' ( Activity ) without any information on scaling or unit. We located a tag at several distances from a sensor and recorded the RSS. Additionally, we tried to estimate the attenuation of the UWB signal by a human hand. A tag is completely invisible to the system at any distance between tag and sensor when the tag is covered closely by a hand. We address this effect to a change of the antenna characteristic rather than simple attenuation. To nevertheless estimate the attenuation due to the human hand, we measured the RSS when the hand is approaching the tag and extrapolated to zero distance. B. Coordinate Measurements We introduced only one tag into the room in order to not irritate the system with UWB noise. Therefore, the measured results show a best case. Any kind of coordinate filtering by the system was switched off ("raw" coordinates). Such filtering is available to smooth the results especially of moving objects, but it also considerably increases the measurement time and prevents any output when nonsensical coordinates have been measured. The tag has been put to six locations on the wooden desks as shown in Fig. 1. Its orientation (angle ϑ) has been changed in 4º steps and more than 1 coordinate readings have been recorded at each position and orientation. To see the effect of more than two sensors in a cell, the whole procedure was repeated for two sensors operating (S + S1, S + S2 and S + S3, S is the master sensor which cannot be switched off), three sensors operating (S1, S2 or S3 switched off) and for all four sensors activated. From these readings, we obtained histograms of the x, y, z coordinate estimates and calculated mean value µ, standard deviation σ and bias (= abs(mean correct position)). Before the calculation of mean and standard deviation, we removed outliers as identified by a "3σ" rule. Finally, we recorded trajectories of a moving tag carried by a person at an update rate of 7 Hz without motion filtering. IV. OBSERVATIONS AND MEASUREMENT RESULTS A. Observations The best location accuracy is achieved when only two sensors receive the tag's signal and the tag is approximately equally far from both sensors. In this case, standard deviation can be as low as 1 cm in the direction parallel to the sensor's base line and well below 1 cm in the two other directions. A tag can be hidden from the system by covering it by a hand. This means that an attenuation of about 1 db to 2 db by human tissue is sufficient to reduce the RSS below the sensitivity limit. The Tx antenna characteristic might be changed when it is covered by a hand such that no signal can be detected anymore at the sensors even if the distance is lower than the sensitivity limit given by the undisturbed, attenuated RSS. The same effect has been observed in an IEEE 82.11g WLAN system (2.4 GHz) with a distance of about 3 m between Access Point (AP) and a Laptop. Attenuation through human hands without touching the AP was 2 db, an additional loss of 1 db occurred when touching the AP antenna. In [3] the attenuation of microwave signals by human tissue has been measured. The results indicate that an attenuation of 1 db to 2 db by a human hand (1 cm to 2 cm of tissue) at 6 GHz is a realistic assumption. When an object, including a human torso, is obstructing the line-of-sight (LOS) path between the tag and a sensor, this sensor does not receive measurable pulses any more. When this situation occurs, no coordinates are acquired and the tag is "lost" from the system; its movement is not tracked anymore. Certain positions lead to completely wrong coordinates, being in the form of outliers in an otherwise normally distributed histogram or resulting in a normally distributed histogram with a bias of up to several meters. We attribute this to reflections at our metal walls because the effect mainly occurs when the tag is in proximity (1 m to 1. m) to a metal wall and the Tx antenna points to the wall. Due to the described effects, the "felt" location accuracy is on the order of 1 to 2 meters in our environment. Strong filtering can reduce this at the cost of a considerably increased measurement time (decreased update interval). Under such conditions, "real-time" tracking is not available anymore. B. Measurement Results The obtained Received Signal Strength as a function of tag-sensor distance d is shown in Fig. 2 together with a least squares fit to M/d (M = 9433). The proportionality to 1/d indicates that the "Activity" readings correspond to electrical field strengths. Our noise level is between 4 and in these Arbitrary Units (AU), hence the indicated detection limit. Attenuation by 13 db (this is the value we achieved as the attenuation of a human hand by extrapolating the attenuation at several tag hand distances to zero) reduces tag visibility to < 4 m. Since any location in the room has a minimum distance of 4 m to at least one of the two required sensor, this effectively hides a tag. The readings at 1.7 m show a reduced activity due to the "near field" condition.

3 2 IEEE International Conference on Ultra-Wideband Paper # The transition between near-field and far-field is approximately at d =.8 m (estimated from d = 2D 2 /λ with D: largest antenna size =.14 m, λ: wavelength = cm at 6 GHz). Figs. 3 and 4 show histograms of the location estimations at certain positions demonstrating particular features. Fig. 3 was taken at P (3.2,,.73), ϑ = [4º, º, 31º] with all sensors active. Fig. 3a, 3b and 3c present the measured x, y and z coordinate readings, respectively. For all three figures, the upper graph was taken at ϑ = 4º, the middle graph was taken for ϑ = º, and the lower graph for ϑ = 31º. The upper graphs (ϑ = 4º) of Figs. 3a c present the best case in terms of measurement accuracy we could achieve: after removal of outliers which are identified by abs((x, y, z) (µ x, µ y, µ z )) < 3(σ x, σ y, σ z ), bias (= (x, y, z) (µ x, µ y, µ z )) and standard deviation σ are better than 1 cm, even better than cm for the x and y directions. The removed outliers are marked by circles in Figs. 3a 3c and Fig. 4. Because the z direction must fully depend on the AoA measurement while the x and y coordinate can profit from the TDoA estimation, it is to expect that the estimate of the z direction is worse. However, while the 3σ - interval of the x and y measurements is below 1 cm, this is not true anymore for the z coordinate: even in the best measurements we achieved, the 3σ location accuracy of the z coordinate is only approximately 3 cm. The lower graphs (ϑ = 31º, same position) of Figs. 3a c represent a situation that we could observe in a significant number of measured histograms: a Gaussian distribution with outliers and with a very small standard deviation (σ < cm) but at a very large bias of approximately. m to over 1 m. We observed such measurements with bias up to several meters. Since only one Gaussian distributed histogram without outliers is present, this bias cannot be removed by an outlier detection algorithm. The middle graphs (ϑ = º) of Fig. 3a and Fig. 3c can be interpreted as the overlay of two independent measurements. In fact we could observe that, since two sensors are sufficient for a location estimate and each sensor is calibrated independently against the master (S), different bias can result from each sensor pair. The system does obviously not calculate a common estimate of all sensor pairs. Rather, if the measurements of a pair of sensors can be solved to a location estimate, this estimate is sent to the server and can be read as a coordinate. Due to this procedure, an overlay of the independently received measurements results with a bias between the two original values and very poor standard deviation (σ >> 1 cm). This finding will be verified and further explained with Fig. 4. The middle graph of Fig. 3b represents the second case of frequent measurement errors: a Gaussian distributed histogram with σ << 1 cm and small bias but with a large contribution of outliers far away from the correct position. Such outliers could be removed from a large population of coordinate readings by an outlier detection algorithm that removes all location estimates outside a 3σ interval of the mean value. We did this in Matlab with our coordinate readings. After such a procedure, the majority of measured histograms have a bias and standard deviation better than 1 cm. We address both observed effects (outliers and false distributed readings) to multipath effects due to reflections at the metal parts of the walls in our class room environment. The results of simulations with channel models for UWB provided by the IEEE WPAN TG4a yield approximately the same result [4] and support the assumption that multipath effects are the cause for the observed outliers. Fig. 4 (tag at P6 = (1.7; ; 1.12), ϑ = 4º) is the proof that the observed overlay of histograms is caused by two pairs of sensors obtaining independent location estimates that are sent to the positioning server without further processing. All three histograms (top, middle, bottom) were taken at the same tag location and rotation. The difference is that in the bottom graph only the sensors S and S2 have been active and in the middle graph only the sensors S and S1 have been active. Both yield differently biased histograms with comparable standard deviation. The tag orientation has been chosen such that S3 does not receive a signal, therefore the combination S, S3 did not yield any location readings. In the top graph, all sensors have been active. It is clearly visible that the top histogram is the superposition of the middle and bottom histograms. Therefore we can conclude that in the case of two sensor pairs receiving a tag s signal, the system does not calculate a common estimate. In the shown case, the resulting bias is 11 cm and the resulting standard deviation is 1 cm. In most of the observed situations of this kind (e.g. Fig. 3a middle or Fig. 3c middle), the standard deviation is much worse than 1 cm. C. Conclusion of the Measurement Results The mentioned "typical location accuracy" of 1 cm corresponds to a 1σ interval (68% of all readings) around the mean value of more than 1 readings after outlier removal. The measurement time to achieve this accuracy is well above 1 seconds. The 3σ (99.4% of all readings) accuracy is only 4 cm, corresponding to the "felt" accuracy in the absence of mis-readings due to reflections. In the presence of reflections, the bias can increase to several meters. Three or more active sensors can yield worse results than only two active sensors due to the effective overlay of the differently biased coordinate readings, increasing the standard deviation at decreasing bias. A clear LOS between the tag's antenna plane and at least two sensors must be present for location. A human torso obstructing this LOS already prevents location. A tag can be completely hidden from the system when the antenna is covered by hand. V. IMPLICATIONS ON APPLICATION SCENARIOS A wide range of applications and user scenarios for UWB-based 3D LT technologies has been announced by companies in the UWB business which have been collected by the IEEE a WPAN Task Group [] and the EU FP6 IP "PULSERS" [6]. Basically, these can be subsumed under two main categories: Person Tracking (PT) and Asset Tracking (AT). Typically mentioned PT applications are security applications like firefighter tracking, military deployment or training tracking, patient tracking or access authentication. Another field for PT is expected in entertainment environments like customer tracking and guidance,

4 2 IEEE International Conference on Ultra-Wideband Paper # theme park guidance or gaming. AT applications are typically of the type inventory control/tracking, warehouse management (RFID substitute), shoplifting protection or machine guidance. The PT applications have in common that they a) cannot assume a supporting user that is willing to provide a line-ofsight path between the tracked tags and the antennas. Be it because the "user" purposely wants to disturb the system (thief, unauthorized personnel) or because the system must not interfere with his basic tasks (rescue, emergency and military scenarios). The entertainment and retail applications must provide location of objects in crowds of people. Our experiences with the installed UWB location technology render it questionable whether such applications will get viable without substantial improvements of UWB LT technology, mainly in terms of sensitivity and robustness to attenuation through human tissue (torso and hand). The AT applications on the other hand must usually operate in highly reflective environments full of large metal objects and/or walls. Apart from potentially obstructing the LOS path (which in this case could not be circumvented by increased sensitivity or transmit power), a LT system must be capable of dealing with reflected signals. Partially, this could be achieved by a means of outlier detection. However, we observed locations in our reflective environment where the false readings do not have the characteristics of outliers (Figs. 3a c). Such situations could probably be detected by improved hardware (spherical antenna characteristics) and low-level software (signal processing operating on the results of all sensors instead of only two for a location estimate). VI. SUGGESTED COUNTERMEASURES Currently, the promised location accuracy of 1 cm (standard deviation and bias) can only be achieved with averaging of over 1 position readings and subsequent outlier detection. For static or quasi-static applications, this could be achieved by software using the current location technology. For real-time tracking of moving objects, the current update rate of < 1 Hz is not at all sufficient. At least ten times higher update rate is required here which, in combination with a 1-fold averaging, results in a requirement of at least a 1 times increased update rate on position readings (1 khz). A substantial improvement of location accuracy would be possible if the different results of two or more sensor pairs would not simply be overlaid but calculated into a common location estimate. Different biases among the different, overlapping partial sensor cells must be avoided by a common calibration procedure. To accommodate the attenuation by human tissue (hand or torso) and therefore get applicable in PT scenarios, an increase in sensitivity on antenna level must be achieved. Transmit power might be gradually increased by making use of the full available UWB spectrum (3-6dB). Increasing the transmitted power level would be a technical but most probably not a legal option due to the already difficult standardization and licensing process. Increased sensitivity might also be achieved by the development of more advanced (coherent) modulation and detection schemes. Presently, to our knowledge, a correlation with threshold detection is used. Measuring the channel impulse response and matched filtering might be an option. Finally, to operate in reflective environments, a method of distinguishing between reflections and true positions must be developed. Such measures might range from outlier detection over channel response measurements to different methods of ranging measurements (e.g. additional TDoA measurements using all 4 sensors of a cell). Also a highly redundant sensor cell with quasi-randomly distributed sensors spread over the location might help here: If a number of sensors receive a LOS signal and can solve their equations to a common coordinate, they could "outvote" those other sensors which do not receive a LOS signal. VII. CONCLUSIONS The promoted location accuracy of 1 cm manifests itself in actual measurements as the standard deviation of position readings from a single sensor pair. Systematic errors (bias) as well as random errors can be substantially larger (range of meters) due to reflections and overlay of location results from different sensor pairs. The effective measurement time to achieve 1 cm accuracy is above 1 seconds. At the maximum update rate of our system of 1 Hz, the resulting accuracy is at most 4 cm to cm (3σ interval) and even as worse as some meters in the presence of reflections. Attenuation by human tissue (hand or torso) obstructing the line-of-sight path between tag and sensor prevents any location reading. The observed behavior does not approve 3D location in its present form for most of the published application scenarios. However, technical improvements might be possible to circumvent some or most of the issues seen in the available system. For personal tracking, the attenuation through tissue is crucial. For asset tracking, the response in reflective environments must be improved. -. S1 S2 y [m] radiator, h=.7m 4. P4 LCD P3 metal H-beam metal wall, h=1.6m metal door ϑ=9º P P6 Desk switch cabinet, h= m. x [m] metal door radiator, h=2m -4. metal door Fig. 1. Room footprint. Sensors S to S3 are located at the corners. Six different positions have been used to take coordinate readings with the tag s orientation angle ϑ rotated by 36º in steps of 4º as indicated at P. P1 P2 P S S3 ϑ

5 2 IEEE International Conference on Ultra-Wideband Paper # RSS ( Activity ) [AU] least squares fit to M/d, M= dB attenuation (human hand) 1 detection limit: Distance d [m] Fig. 2. Received Signal Strength ( Activity ) of pulses received at sensor versus tag-sensor distance d. hits [ϑ = 4 ] hits [ϑ = ] hits [ϑ = 31 ] real position, bias =.3 std =.4 1 real position, bias =.14 1 std = real position, bias =.7 std =.3 x position [m] Fig. 3a. Readings of the x-coordinate, tag is located at P = (3.2; ;.73) and turned by 9º in steps of 4º. Rotation angle ϑ is defined in Fig. 1. Thick line: Gaussian distribution function with mean µ and σ = std after removal of outliers as indicated by circles. hits [ϑ = 4 ] hits [ϑ = ] hits [ϑ = 31 ] 4 2 real position, bias =.4 std = real position, bias =.21 std =.7 2 strong outlier real position, bias = 1.41 std = y position [m] Fig. 3b. Rotated tag: y-coordinate. Circle indicates outlier removed before calculation of mean µ and σ = std of Gaussian distribution (thick line). hits [ϑ = 4 ] hits [ϑ = ] hits [ϑ = 31 ] 1 real position, bias =.8 1 std = real position, bias =.4 std = real position, bias =.44 std = z position [m] Fig. 3c. Rotated tag: z-coordinate. Circle indicates outlier removed before calculation of mean µ and σ = std of Gaussian distribution (thick line). hits [S, S1, S2, S3] hits [S, S1] hits [S, S2] real position, bias =.11 std = real position, bias =.24 std = real position, bias =. std = y position [m] Fig. 4. Readings of y-coordinate at P6 = (1.7; ; 1.12), ϑ = 4º taken by two active sensors (S and S2 or S and S1, respectively) and taken by four active sensor (S to S3).Tag orientation was ϑ = 4º such that S3 did not receive any signal. Circle indicates outlier removed before calculation of mean µ and σ = std of Gaussian distribution (thick line). REFERENCES [1] Institute for Electronics and Electrical Engineers, "IEEE 82.1 WPAN Low Rate Alternative PHY Task Group 4a (TG4a)", Last update: [2] N.N., "Pulsers - Public Documents", documents_deliverables.shtml, Last update: 24 [3] Yamaura, I., "Measurements of Ghz Microwave Attenuation in the Human Torso," Microwave Theory and Techniques, IEEE Transactions on, vol. 2, no. 8, pp , Aug [4] Tüchler, M., Schwarz, V., and Huber, A. "Location Accuracy of an UWB localization system in a multi-path environment". Accepted for presentation at the International Conference on Ultra-Wideband ICU