Indoor Localization Using FM Radio Signals: A Fingerprinting Approach

Transcription

1 Indoor Localization Using FM Radio Signals: A Fingerprinting Approach Vahideh Moghtadaiee, Andrew G. Dempster, and Samsung Lim School of Surveying and Spatial Information Systems University of New South Wales Sydney, Australia v.moghtadaiee@student.unsw.edu.au Abstract Indoor positioning has become highly important because of the failure of GPS in such areas. Many Wireless Local Area Networks (WLAN) indoor localization studies use the fingerprinting technique. In this study, a new positioning system is proposed based on broadcast FM as a signal of opportunity, with significant benefits for indoor positioning. This localization system uses FM signal strength fingerprinting. The deterministic approach of fingerprinting is considered, and several algorithms are compared. The results demonstrate a minimum mean distance error of 2.96m for the K-Weighted Nearest Neighbors (KWNN) algorithm with K=6. The comparison between using fingerprinting for FM and Wi-Fi is also discussed. Keywords-component; formatting; style; styling; insert (key words) I. INTRODUCTION The Global Positioning System (GPS) and Global Navigation Satellite Systems (GNSS) in general have been adopted as the de facto positioning technology because of the highly accurate location information they provide globally; however, this technology fails in particular environments such as indoors or urban canyons. These failures of GNSS are mainly due to the low received signal power and low visibility of satellites in urban/indoor areas. Therefore, non-gnss navigation technologies are essential for such regions [1]. Utilizing signals of opportunity is a promising alternative navigation means of providing adequate geo-location [2]. Signals of opportunity are existing (non-navigation) radio frequency (RF) signals around us, which tend to have much higher power levels and wider coverage in urban and indoor environments than GNSS signals. Furthermore, they can penetrate buildings due to their lower frequencies [3]. Although such signals cost less than GNSS in terms of the system implementation [4], there are significant problems that need to be considered in employing each of them for positioning, owing to the fact that such terrestrial signals were not designed for location estimation. Signals of opportunity in previous research include analogue/digital television and analogue/digital audio signals transmitted from commercial radio and television broadcasting towers [3]. They also include Global System for Mobile communication (GSM) signals from mobile telephone base stations [3]. Other types of signals of opportunity are Ultra- Wide Band (UWB), ZigBee, and Wireless Local Area Networks (WLAN) such as Wi-Fi and Bluetooth [2]. Many studies have been especially done on employing Wi- Fi signals for indoor positioning [5], [6]. In this paper, however, we focus on indoor navigation using broadcast FM signals - an analogue audio signal with outstanding advantages for urban/indoor positioning purposes. These benefits are the ability to be received both indoors and outdoors, dense coverage in urban areas where GNSS behaves poorly, availability without installing additional transmitters, low-cost and low-power usage hardware with no complicated technology, high received signal power, and finally the large number of transmitters that can provide good geometry for positioning [7]. The geometry of transmitters is highly significant in the distance-based positioning techniques. The most crucial problem when using an FM signal, however, is that they do not carry any timing information, which is a critical factor in range calculation [8]. Moreover, navigation using FM signals would be degraded, as are most radio-navigation systems, by the effects of multipath and Non- Line-of-Sight (NLOS) signals. Measurements that can be taken from signals of opportunity for navigation purposes are based on: Time of Arrival (TOA), Time Difference of Arrival (TDOA), Angle of Arrival (AOA), and Received Signal Strength (RSS). For the first three methods, the lack of timing information in FM signals is critical. Hence, some of the previous studies proposed utilizing an independent fixed position observer unit that receives all FM signals from the FM transmitters in the area, and observes the difference in synchronization between them and broadcasts transmitted signals relative information [8]. Performing TDOA range measurements for a set of unsynchronized FM subcarrier signals by making use of an additional observer module is discussed in [9]. In addition, in order to extract timing and direction information, particular hardware with a multidirectional antenna is required [10]. The latter positioning technique is localization based on RSS and signal propagation modeling. There are two general approaches to wireless localization using the RSS technique: signal propagation modeling and location fingerprinting. The /11/$ IEEE

2 former is not discussed in this paper since indoor radio signal propagation is very difficult to model usefully. Indoor received signals have been attenuated by distance from the transmitters, multipath, and penetration through walls, buildings, floors. Hence, no useful model of signal propagation can present the characteristics of the signal in reality. Fingerprinting, on the other hand, is capable of alleviating some of the problems related to multipath and NLOS propagations [11]. Fingerprinting has two stages: training and positioning. A database of the location dependent parameters collected at the reference points (RPs) is generated in the training stage, and in the positioning stage, the parameter vector ( fingerprint ) of the user is compared with the database. By applying different matching algorithms, the best match is found and considered to be the user position estimate [11]. In this paper, the fingerprinting technique is chosen for indoor FM-based positioning since it defeats some of the difficulties of using FM signals. It is an economical method, as it does not need any additional hardware or infrastructure [10]. Secondly, fingerprinting is independent of the timing problem of FM signals. Finally it does not suffer from the effects of multipath (arguably, it exploits multipath) compared to other methods based on distance measurements [12]. However, it should be noted that constructing the database for fingerprinting is always time-consuming and labor-intensive [13]. More recently, the use of location fingerprinting has been investigated using FM broadcasting signals in a campus area in [10]. The authors employed the probabilistic approach of fingerprinting, producing an accuracy of 35m at 67th percentile of the cumulative distribution function (CDF). The experiment, however, was limited to a campus area and was carried out for outdoor purposes only. The technique in [14] utilized some commercial short-range FM transmitters in an indoor area, and compared the positioning results with Wi-Fi positioning. The performance of the two positioning technologies was evaluated using the leaveone-out approach and K-Nearest Neighbor (KNN) methods. The median accuracy is about 1.0 m and 5.0 m at 95% confidence level. To the best of our knowledge, there is no research on indoor positioning using FM broadcast signals. The rest of the paper is organized as follows. Section II describes the methods utilized for FM positioning. Section III presents the experimental setup used to fulfill the FM-based positioning. The experimental results and the analysis are discussed in Section IV. Section V compares the results of our FM positioning with Wi-Fi positioning which has been formerly carried out in the same test bed. Finally, Section VI provides the conclusions and discusses some ideas for future work. Q fingerprinting measurements of all P FM channels sensed at each RP in a specified period of time, which is the vector of {RSS rq = [RSS r1 RSSS r2... RSS rp ], r = 1, 2,..., R, q = 1, 2,, Q}. The average of all measurements of each FM channel is calculated according to and is logged as the reference data of that RP in the database. Fig. 1 illustrates the whole procedure of the training stage of FM-based fingerprinting. B. Positioning Stage In this stage, the unknown location will be estimated by comparing the average of Q observed measurements {rss q = [rss 1 rss 2... rss p ], q = 1, 2,, Q}at an unknown point according to with the database established in the training stage. Various matching algorithms can be used to find the best match indicating the estimated position. The process of the positioning stage is shown in Fig. 1. In the positioning phase of fingerprinting, there are two main ways to estimate a location: the deterministic and probabilistic approaches [11]. In this paper, we just analyze the deterministic approach. Three different algorithms are applied for this purpose. The first is the Nearest Neighbor (NN) algorithm [15], in which the position of the nearest RP to the unknown point is considered as the estimated location. The nearest RP is determined with the shortest distance to the unknown point. Such distance calculation is based on Manhattan distance and Euclidean distance between the observed fingerprint and those recorded in the database [16]. In this paper, we use Euclidean distance which is defined as follows: The second method is K-Nearest Neighbor (KNN), in which the estimated location is the average of the coordinates of K nearest points. The third algorithm which is similar to KNN is K-Weighted Nearest Neighbors (KWNN). In this method the weighted average of the coordinates of K nearest points is calculated rather than the average. The weights are the inverse of the Euclidean distance. (1) II. FINGERPRINTING TECHNIQUES A. Training Stage The first stage of the work is to establish a RSS database for a set of R reference points (RPs) with known (X,Y) position located indoors in 2D space, which will be used as training samples in the positioning stage. Such a database includes the Figure 1. Two stages of FM-based location fingerprinting

3 It has been shown that the KNN and KWNN methods provide higher accurate results than the simple NN algorithm [11]. However, it should be noted that the NN method performance is better than the other methods in case that the distances between RPs are large, but since some of the nearest neighbors may be too far from the estimated points, considering too many nearest neighbors can drop the accuracy of the estimator. The KWNN method presents higher accuracy, which shows the weighting scheme effect on improving the positioning results [11]. III. EXPERIMENTAL SETUP A. Experimental Test bed The experimental test bed is located on the 4th floor of Electrical Engineering building at our university. The layout of the test bed is shown in Fig. 2. The test bed has dimensions of 11m by 23m, and consists of 7 rooms, which are a typical indoor office environment, and the corridor. The crosses represent RPs and the squares are the test points (TPs). This is the same test bed as used for Wi-Fi positioning previously in [5], so we can compare our result with Wi-Fi positioning results. In our experiment, there are 150 RPs (R=150), and 28 test points. The RPs are distributed as evenly as possible. They are mostly close to or aligned with the corners, doors and windows, which could be easily identified on the map and in the real test bed. The points that users are most likely to require are chosen as the TPs. There are 17 FM channels sensed at each RP (P=17), which cover whole FM bandwidth from 88MHz to 108MHz. At each RP the user faces north first, and records the RSS of the sensed FM channels. Then the orientation is changed to south and the RSS values are logged. The reason for this is that the antenna (see Fig. 3) was wide to fit in the narrow corridor for east and west orientations. A total of 120 measurements are made at each point within 12 seconds (Q=120). Since our FM-based positioning is two-dimensional positioning, the height of the FM antenna kept constant in all measurements (height=75cm). The data were collected in one day over the weekend, so that not many people were present. B. Data Acquisition The investigation of signal of opportunity data is based on a Software Defined Radio (SDR) approach. The equipments utilized for this purpose are shown in Fig. 3. The first one is the Universal Software Radio Peripheral (USRP2) manufactured by Ettus Research LLC [17]. It captures and stores the raw spectrum for post-processing. USRP2 is used with GNU Radio source codes written in Python (with some signal processing capability) and run on the Ubuntu platform. The maximum sampling rate that USRP2 can provide is 100 MS/s. Secondly the TVRX daughterboard manufactured by Ettus Research LLC was used with USRP2. It can receive 50 to 870 MHz, which covers the FM signal bandwidth. Finally, for the receiver antenna, we use a v-shaped rabbit ear antenna, an effective indoor antenna for both FM and television broadcast signals, which takes advantage of the nearly universal use of circular polarization on FM [18]. Figure 2. Experimental layout for FM-based fingerprinting Figure 3. The equipments used for data acquisition The output of the USRP2 is a binary file including all samples of the signal for every RP and TP. This file was converted to complex data in MATLAB and the 120 RSS data sets for each sensed FM channels were extracted and averaged. The average is then logged in the database for online positioning purposes. In practice, the sampled data might not be the same even when the measurements are taken at a fixed location or at the same frequency, as changes in the indoor environment influence the propagation of the RF signal and then the RSS fluctuates. Fig. 4 shows a typical histogram of the measured RSS for the strongest channel at one of the RPs. It shows that the strongest channel level is distributed from -47 to -52 dbm for one location. Other channels and other locations suffer similar variation of the RSS of the FM radio signal, which is around 5 db for different FM channels in all RPs.

4 Figure 4. A typical probability density of RSS for one FM channel measured at one location Figure 5. Mean distance error using NN, KNN, and KWNN methods for different K values IV. EXPERIMENTAL RESULTS AND ANALYSIS To evaluate the performance of our indoor FM-based positioning, the experimental results of the mentioned test bed using NN, KNN, KWNN algorithms are analyzed and compared from three different aspects: K value, data acquisition time, and sensed channel numbers. The errors between the estimated and true location are calculated as a Euclidian distance. The mean distance error using the three algorithms is calculated. Fig. 5 shows the relationship between mean distance error and the value of K for these three algorithms. K equals to 1 represents the NN method. When K is equals to 3, the error is less than 20cm higher than K equals to 2, which can be ignored. However, the general trend of the figure is downwards to a minimum error and then going up. The optimal value of K is when the mean distance error is minimum. The results indicate that this happens when K equals 6 for both KNN and KWNN methods. Also, K equals 5 is slightly less accurate than the results of K=6. Hence, by taking six (or five) nearest neighbors of the test point into account for positioning, the best location estimation can be achieved, whereas using fewer nearest neighbors might miss some useful information. Moreover, Fig. 5 demonstrates that the weighted average of nearest neighbors gives improved results compared to NN and KNN. The CDF of the error is plotted in Fig. 6 for NN, KNN, and KWNN when K=6. It shows that the NN algorithm has its best 50% of estimates more accurate than the other techniques. The maximum error of the NN algorithm is about 13 m. On the contrary, the KNN and KWNN algorithms can grant more accurate estimation than the NN algorithm in higher probabilities. The maximum error is about 8.5 m for both KNN and KWNN algorithms which is 4.5 m less than NN algorithm. It can be observed that the error at the 67th percentile is 5, 3.8, and 3.7 m for NN, KNN, and KWNN methods, respectively. When the 95th percentile is considered, the error is 11, 6.5, and 6.4 m (see Table I.). Figure 6. CDF of Deterministic approach results for different algorithms (K=6) As a result, the NN algorithm would not be an appropriate option for FM-based positioning when reliable high accuracy results are needed. In this case, the KWNN algorithm provides the best results. TABLE I. THE ERRORS FOR DIFFERENT ALGORITHMS (K=6) NN KNN KWNN Mean Distance Error (m) Median error (m) CDF 67% (m) CDF 95% (m)

5 A second approach is to evaluate the effect of the number of measurements made at each RP or similarly the corresponding period of data acquisition time. In our experiment, 120 samples are taken for 12 seconds for each point. Even though a higher number of measurements could be expected to give better accuracy, the number of samples used for all points (RPs and TPs) resulted in a huge amount of data (260 GB) and took significant effort. Thus, finding out whether lower sampling time would suffice is helpful. The mean distance error for different data acquisition times is illustrated in Fig. 7 for NN, KNN, and KWNN algorithms. We expected that the mean distance error would decrease by increasing the data acquisition time, but the figure does not match that expectation, especially when the NN algorithm is considered. The results indicate that 12 seconds of data provide only slightly higher accuracy than fewer ones for KNN and KWNN algorithms, but in general, the three algorithms do not show particular sampling-time-dependent behavior. The highest error differences for all algorithms from 0.8 to 12 seconds of sampling time are less than 50 cm, which is quite low. It is an important result since helps reduce the labor and amount of recorded data in the training stage. The fact is that there is a tradeoff between the accuracy obtained in positioning stage and data storage, and fingerprinting time in the training stage. The comparison between data collection time of 0.8 and 12 seconds is shown in Table II. Sampling time and data storage for all RPs are greatly reduced while the accuracy is only slightly degraded. Therefore, less than 1 sec data acquisition is enough for FM-based positioning systems to provide accuracy of around 3 m. The number of sensed FM channels in our experiment is 17. However, there are some channels which are broadcast from the same antenna. Therefore, a third approach is to examine the effects of frequency diversity as separate from transmitters location diversity. We just select one channel among other channels broadcasting from one location, which decreases the number of sensed channels to 9. TABLE II. SAMPLING TIME COMPARISION Sampling time for each RP (Sec) Sampling time for all RPs (Sec) Data storage for all RPs (GB) Mean distance error for NN (m) Mean distance error for KNN (m) Mean distance error for KWNN (m) The comparison between the mean distance errors for P=17 and P=9 are shown in Table III. The results show that the number of sensed FM channels does have an effect on the positioning accuracy. Without changing the number of transmitter locations, the extra signals improve the mean distance error from about 4 m to about 3 m. Fig. 8 also illustrates the CDF of KWNN algorithm when the number of sensed channels is 17 and 9. It clearly shows the accuracy reduction when the sensed channel number decreased. In addition, the figure depicts that we never achieve less than 1 m accuracy in the 9 channel case. The maximum error worsens from 8.5 m to 11.5 m by removing 8 FM radio signals broadcasting from one transmitter tower. TABLE III. MEAN DISTANCE ERROR FOR DIFFERENT FM CHANNEL NUMBERS NN KNN KWNN 17 channels channels Figure 8. CDF of KWNN algorithm for 17 and 9 sensed channels (K=6), retaining the spatial diversity Figure 7. Mean distance error in different data acquisition time for different algorithms (K=6)

6 V. COMPARING RESULTS OF FM VERSUS WI-FI POSITIONING In order to have a good comparison between navigation using FM signals and Wi-Fi signals, we chose the same test bed as was used for Wi-Fi positioning in [5]. There are 132 RPs in [5], so we decreased the number of RPs in our system to 132 in order to have the same conditions as the Wi-Fi case. To investigate the effect of the number of RPs in location estimation, the number of RPs is intentionally reduced to 99, 66, 33, and 16, as was done in [5]. But the RPs are still gridded as evenly as possible in the test area. Hence, in total 6 and 5 fingerprinting databases were generated for FM and Wi-Fi cases, respectively, since there is no data for accuracy of Wi-Fi for 150 RPs. Again we consider K equals 2, 3, 4, 5, and 6. Fig. 9 shows the average of the positioning error using all the NN, KNN, KWNN algorithms for FM and Wi-Fi positioning systems. When the number of RPs increases, the accuracy of the estimated location increases. When the density of the RPs is high, the rate of increase of the accuracy decreases. The similar rate of decreasing the error by higher number of RPs can be easily seen for both Wi-Fi and FM. This figure also indicates that Wi-Fi positioning has higher accuracy. This can be expected since the Wi-Fi transmitters are very close to the receiver and because they are short-range and have a significantly shorter wavelength, their RSS varies more rapidly in a small region. However, FM-based positioning gives reasonable results for indoor positioning and will cost much less than Wi-Fi in terms of receiver simplicity, FM availability and coverage, and sensitivity to the environment. FM stations, unlike Wi-Fi access points, are not turned on and off and moved, requiring frequent database updates. VI. CONCLUSION AND FUTURE WORK No positioning technology has good performance in all environments. GPS as a globally accepted technology has operational difficulties in urban and indoor areas. Utilizing terrestrial signals of opportunity instead of or in addition to GPS can be helpful for navigation in such regions. Figure 9. Mean of average distance errors for FM and Wi-Fi positioning in the same test bed In this study, a new positioning method based on broadcast FM radio signals as a signal of opportunity is proposed because of their considerable benefits for indoor positioning, such as the ability to be received both indoors and outdoors, dense coverage in urban areas where GNSS behaves poorly, availability without installing additional transmitters, low-cost and low-power usage hardware with no complicated technology, high received signal power. For navigation purposes, the fingerprinting method is utilized, in which the RSS of the FM signals are taken to build a data map and postprocessing. The experiment for estimating unknown locations by the deterministic approach of location fingerprinting was carried out based on NN, KNN, and KWNN algorithms. They are compared from three different aspects: K value, data acquisition time, and sensed channel numbers. The results determine the KWNN method as the best one here, in which a minimum mean distance error of about 3m is obtained. The comparison between our FM-based positioning results and Wi-Fi positioning results from the same test bed show the higher accuracy of Wi-Fi positioning; however, FM-based positioning not only gives reasonable results for indoor positioning, but also costs much less than Wi-Fi in terms of receiver simplicity, FM availability and coverage, and sensitivity to the changes in the environment. Moreover, FM stations, unlike Wi-Fi access points, are not turned on and off and moved, requiring frequent database updates. For the next stage, we plan to analyze the probabilistic approach of location fingerprinting for our FM-based positioning and examine more fully the effects of transmitter location and frequency diversity. Furthermore, we will examine outdoor applications, and transitions from indoors to outdoors. In addition, the positions at which the errors are largest will be analyzed in details. This can be important input to choosing new reference points. Finally, we will try to combine FM radio signals and Wi-Fi signals so that we could achieve higher accuracies for indoor positioning. REFERENCES [1] J. Raquet and R. K. Martin, Non-GNSS radio frequency navigation, in Proc. of IEEE ICASSP, 2008, pp [2] H. TIAN, E. MOK, L. XIA, and Z. WU, Signals of opportunity assisted ubiquitous geolocation and navigation technology, in Proc. of SPIE, 2008, pp [3] J. F. Raquet, M. M. Miller, and T. Q. Nguyen, Issues and approaches for navigation using signals of opportunity, in Proc. of ION NTM, San Diego, CA, 2007, pp [4] Chunpeng Yan and H. Howard Fan, Asynchronous diffrential TDOA for non-gps navigation using signals of opportunity, in Proc. of IEEE ICASSP, 2008, pp [5] B. Li, Y. Wang, H.K. Lee, A.G. Dempster and C. Rizos, "Method for yielding a database of location fingerprints in WLAN," Proc. IEE Communications, vol. 152, no. 5, pp , Oct [6] M. Youssef, A. Agrawala, and U. Shankar, WLAN location determination via clustering and probability distributions, in Proc. IEEE PerCom, 2003, pp [7] V. Moghtadaiee, S. Lim, and A.G. Dempster, System-level considerations for signal-of-opportunity positioning, presented at Int. Symp. GPS/GNSS, Taipei, Taiwan, 2010.

7 [8] D. C. Kelly, J. Cisneros, and L. A.Greenbaum. Navigation and positioning systems and method using uncoordinated beacon signals. U.S. Patent , Jan. 18, [9] J. M. Janky and J. F. Schipper, Location of emergency service workers, U.S. Patent , Mar. 09, [10] S.-H. Fang, J.-C. Chen, H.-R. Huang, and T.-N. Lin, Is FM a RF-based positioning solution in a metropolitan-scale environment? A probabilistic approach with radio measurements analysis, IEEE Transactions on Broadcasting, vol. 55, no. 3, pp , [11] B. Li, Terrestrial mobile user positioning using TDOA and fingerprinting techniques, PhD thesis, School of Surveying & Spatial Information Systems, University of New South Wales, Sydney, Australia, [12] T. King, T. Haenselmann, and W. Effelsberg, Deployment, calibration, and measurement factors for position errors in based indoor positioning systems, in Proc. Location-and Context-awareness, 2007, pp [13] L. Anthony et al., Place Lab: Device positioning using radio beacons in the wild, Lecture Notes in Computer Science, vol. 3468, pp , [14] A. Matic, A. Papliatseyeu, V. Osmani, and O. Mayora-Ibarra, Indoor positioning using FM radio, International Journal of Handheld Computing Research (IJHCR), vol. 1, no. 3, pp , [15] P. Bahl and V.N. Padmanabhan, RADAR: an in-building RF-based user location and tracking system, in Proc. of IEEE INFOCOM, 2000, vol.2, pp [16] T. Cormen, C. Leiserson, and R. Rivest. Introduction to Algorithms. New York: MIT Press, [17] Ettus Research LLC. Internet: [May. 11, 2011] [18] Circularly Polarzied Rabbit Ears. Internet: [April. 20, 2011]