Transcript
Indoor=outdoor location of cellular handsets based on received signal strength J. Zhu and G.D. Durgin The feasibility of received signal strength (RSS) methods for locating handset calls in a complicated wireless cellular environment is demonstrated. An extensive measurement campaign conducted on the Georgia Tech campus indicates that RSS location techniques can locate handset calls within 100 m error distance to its ground truth 78% of the time for a network with a majority of indoor users. A received signal strength aggregate (RSSA) method for identifying and locating indoor handsets is also proposed.
Introduction: The RSS location technique is a relatively new method for locating handsets in the cellular network. The technique estimates a cellular phone’s location by matching signal strengths measured at the handset with signal strengths recorded in a database of radio frequency (RF) maps. These RF maps are generated through a combination of propagation modelling and field measurement. The technique has found its most common application in indoor wireless local area data networks [1]. Laitinen et al. [2] reported location performance statistics for outdoor users in a GSM network, but there has been no previous research that includes indoor users – which now constitute the bulk of cellular network traffic. Furthermore, RSS location is a low-cost solution for emerging location-based services across the world. In this Letter, we present the properties of handset-measured received signal strength and propose several RSS location algorithms. We also demonstrate a technique for discriminating between indoor=outdoor users with up to 94% reliability. Results from an IS-136 network show that the proposed method can locate outdoor users within 100 m 74% of the time and 300 m 97% of the time. When the system is loaded with indoor users, the accuracy falls to 70% of the time within 100 m and 96% of the time within 300 m. Three different databases of RF maps were compared side by side so that the accuracy effects of different amounts of modelling and measurement can be seen.
for all nearby serving sectors was created from a combination of propagation modelling and varying levels of indoor and outdoor pedestrian-style handset measurements were taken with an Ericsson handset connected to an Ericsson TEMs data collection unit. 24 892 pedestrian handset samples are recorded for the location performance calculations.
Calibration level of RF map database: The database used in the location experiments is a collection of raster-based RF maps (10 by 10 m resolution). These maps are initially populated with received power using a modified Hata propagation model with path loss exponent 3.3 [4]. Then the modelled RF maps are compared to any available scanner measurements, which are also grouped into 10 by 10 m rasters with linear power averaging to remove small-scale fading effects. A dB-offset is added to each predicted RF map so that the mean error between measurement and prediction is 0 dB. The linearly-averaged measurement points are also interpolated directly into the modelled RF maps. Thus, each RF map contains scanner measurements which gracefully blend into modelled, unmeasured regions. Measurement calibration using a Comarco IS-136 scanner with baseband decoding. 26 RF maps are generated for Georgia Tech campus, which are calibrated by 21 851 scanner samples. Our study reports location performance using three different levels of RF map database. The Level 1 database of RF maps is calibrated with only outdoor drive-test scanner measurements. This type of database represents the general purpose RSS position location solution proposed in [5]. The Level 2 database of RF maps is identical to the Level 1 database of RF maps, except raster points corresponding to indoor locations are modified with additional penetration loss based on a simple, georeferenced map of building footprints. The Level 3 database of RF maps is constructed from extensive outdoor and indoor measurements. The Level 3 database shows the upper-limit of performance for an indoor=outdoor location algorithm. indoor signal strength distribution
Description of RSS location: The RSS location algorithm solves for users’ xy-co-ordinates by comparing the signal strength measurements to a well-calibrated database of RF maps [3]. In IS-136 or GSM protocols, this handset measurement of RSS is typically reported in a network measurement report (NMR) that is sent back to the base station on the reverse control channel.
outdoor signal strength distribution
signal strength area shows outdoor probability area shows indoor probability
Fig. 2 Distribution of received signal strength aggregate (RSSA) measured by both indoor and outdoor handsets
Fig. 1 Indoor=outdoor RSS aerial photograph of Georgia Tech campus used for measurement campaign and location analysis, in semiurban Atlanta, GA (1100 800 m area pictured)
Experiment environment: Tests were performed on an 850 MHz IS-136 cellular network in mid-town Atlanta on the campus of the Georgia Institute of Technology (see Fig. 1). The experimental area is 700 by 600 m (marked by a white box in Fig. 1) and contains eight sectors from three base stations. Another six base stations are in the surrounding neighbourhood. The campus approximates a typical semi-urban or dense suburban area with streets, moderate green space, and many four- to five-story academic and office buildings. There are no skyscrapers or ‘urban canyons’ that would be associated with dense urban environments. A database of RF coverage maps
Handset RSS properties: We found that the additional penetration losses experienced by indoor cellular users were large enough to reliably distinguish them from the outdoor users. Based on these handset measurements, the average RSS over the N strongest channels was calculated, which we call the received signal strength aggregate (RSSAN). We assume that the distributions of the six strongest channels are independent and that their dB-sum is log-normal distributed. Based on our experiment, the mean of indoor RSSA6 is 97.8 dBm and the standard deviation is 14.1 dB. For outdoor RSSA6, the mean is 85.5 dBm and the standard deviation is 9.7 dB. Thus, average indoor penetration loss is 12.3 dB. Fig. 2 shows the distribution of RSSA between indoor calls and outdoor calls in side-by-side plots. Interestingly, the standard deviation for RSSA6 is larger for indoor handsets (14.1 dB) compared to outdoor handsets (9.7 dB). If gains and losses among commercial handset RF chains are somewhat similar, then it may be possible to discriminate between all indoor and outdoor handsets using RSSAN. For indoor=outdoor discrimination rate, which is defined as the percentage of handset locations correctly decided to be indoors or outdoors, the proposed algorithm gives correct estimates around 90% and climbing to 94% in some cases, as shown in Table 1.
ELECTRONICS LETTERS 6th January 2005 Vol. 41 No. 1
Table 1: Performance of hybrid-method RSS location algorithm Indoor=outdoor discrimination rate [%] LVL1
LVL2
LVL3
Location error-statistics [%] % error < 100 m
% error < 300 m
LVL1 LVL2 LVL3 LVL1 LVL2 LVL3
M8
90
90
90
56
56
65
96
96
96
M8A
92
92
91
61
64
78
97
98
98
M6
86
86
87
52
53
57
93
93
93
M6A
87
87
87
62
63
70
96
96
96
O6
90
91
89
54
54
56
93
93
94
O6A
93
93
94
68
68
74
97
97
97
Lvl1: RF database calibrated with outdoor measurement Lvl2: calibration with outdoor measurement and indoor modelling Lvl3: calibration with exhaustive outdoor and indoor measurements M: mixed indoor=outdoor sample point (2=3 indoor test points) O: trial includes only outdoor handsets 8=6: eight- or six-sector information available A: linear averaging of 10 consecutive NMRs
Location algorithm details: The input of this algorithm is a calibrated database of RF maps and a sequence of NMRs. An NMR reports signal strengths from N control channels, which we represent as the vector [Ph1, j Ph2, j, . . . , PhN,j], where Phi,j is the jth reported signal strength in dBm of the ith control channel. A set of signal strengths at a particular point in a database of RF maps is also represented similarly as the vector [Pdx,y,1 Pdx,y,2, . . . , Pdx,y,N], where Pdx,y,i is the predicted signal strength in dBm of the ith control channel at the co-ordinate xy. The first step in location estimation is to normalise the NMR vector, removing any dB-biases in the RF chain that would skew the handset measurement relative to a database of RF maps calibrated with a scanner radio. A mean RSS for each control channel is generated by linearly averaging each sector’s power over 10 consecutive NMRs. Then the RSS mean of all control channels is subtracted from the RSS of each reported channel. For each raster point in the database of RF maps, we use the same normalisation procedure, subtracting out the average dBm-power of the same set of received channels from each Pdrx,y,i . The next step is to calculate the Euclidean distance, M(x, y), between these normalised vectors of handset signal strengths and database signal strengths. The xy-co-ordinate that corresponds to the minimum in M is the preestimate for the handset’s location. To further improve our location estimation, we incorporate our pre-estimate above with indoor=outdoor discrimination information. The indoor weight factor is the probability that indoor RSSAN is greater
than a value x in the indoor distribution, which is the area marked with ‘=’ in Fig. 2. The outdoor weight factor is the probability that the outdoor RSSAN is less than a value x in the outdoor distribution, which is the area marked by n’. The matching distances of (3) are multiplied by these weighting factors. In the last step, a location estimate is decided by searching for co-ordinates x, y where the weighted M(x, y) yields the lowest weighted matching distance.
Results and discussion: Table 1 shows results of various location algorithms operating with different types of RF map databases. The linear averaging of 10 consecutive NMRs is very effective in improving the location error statistics, raising the 100 m accuracy numbers between 5 and 18%. With averaging, the performance of the eightsector cases is almost identical to the performance of the six-sector cases. This suggests that the strongest six received signals contribute the bulk of accuracy to the overall system performance.
Conclusion: The results in this Letter demonstrate the feasibility of RSS location techniques for low-cost deployment in IS-136 and GSM networks. The technique is particularly powerful for locating and discriminating indoor users, which is not possible for other triangulation techniques. # IEE 2005 Electronics Letters online no: 20056605 doi: 10.1049/el:20056605
10 September 2004
J. Zhu and G.D. Durgin (School of Electrical and Computer Engineering, Georgia Institute of Technology, 777 Atlantic Drive, Atlanta, GA, USA) E-mail:
[email protected] References 1 2 3 4 5
Pahlavan, K., and Li, J.-P.M.X.: ‘Indoor geolocation science and technology’, IEEE Commun. Mag., 2002, 40, (2), pp. 112–118 Laitinen, H., Lahteenmaki, J., and Nordstrom, T.: ‘Database correlation method for GSM location’. VTS 53rd Vehicular Technology Conf. 2001 Spring, May 2001, Vol. 4, pp. 2504–2508 Weiss, A.J.: ‘On the accuracy of a cellular location system based on RSS measurements’, IEEE Trans. Veh. Technol., 2003, 52, (6), pp. 1508–1518 Rappaport, T.S.: ‘Wireless communications: principles and practice’ (Prentice-Hall Inc., New Jersey, 2002, 2nd edn.) Perez-Breva, L., et al.: ‘Location and determination using RF fingerprinting’ . US Patent No 6,393,294, 22 March 2000 (Issued to Polaris Wireless Inc.)
ELECTRONICS LETTERS 6th January 2005
Vol. 41 No. 1
