使用調頻及數位電視訊號的室內定位方法研究

全文

(1)NATIONAL TAIWAN N ORMAL U NIVERSITY C OMPUTER S CIENCE AND I NFORMATION E NGINEERING. Indoor Localization using FM and DVB-T Signals. Supervisor:. Author:. Dr. Ling-Jyh C HEN. Shan-Ho YANG. February 2, 2015.

(2) Indoor Localization using FM and DVB-T Signals Shan-Ho Yang Dept. of Computer Science and Information Engineering, National Taiwan Normal University [email protected] Abstract Accurate indoor localization has been an objective of the ubiquitous computing research community, and numerous indoor localization solutions based on Wi-Fi, Bluetooth, ZigBee, and infrared technologies have been proposed. Owning to both of these signals are high frequency, the ability of penetration are weak and incline to be affected by the presence of obstacles such as desk, shelf, and etc. Furthermore, both of these signals require extra hardware to localize. This work using FM and DVB-T as signal resources. Overcoming the defects caused by previous signals is the target, and verifying the result by proceeding several consecutive experiments. Discovering the possibilities of indoor localization by combining FM with DVB-T. We investigated the temporal effect on the training data in terms of age and period. Due to the hybrid system, we can achieve sub-meter accuracy.. Keywords: FM, DVB-T, Indoor Localization.

(3) C ONTENTS I. Introduction. 1. II. Related Work. 4. II-A. Wi-Fi Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . .. 4. II-B. FM Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . .. 6. II-C. Other Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . .. 7. II-D. Hybrid Fingerprints . . . . . . . . . . . . . . . . . . . . . . . .. 8. III. IV. Methodology. 10. III-A. Testbed Setup . . . . . . . . . . . . . . . . . . . . . . . . . . .. 11. III-B. Site-Survey Phase . . . . . . . . . . . . . . . . . . . . . . . . .. 11. III-C. Testing Phase . . . . . . . . . . . . . . . . . . . . . . . . . . .. 12. Evaluation. 13. IV-A. Indoor NTNU-CSIE 2F for extensive site-survey . . . . . . . .. 14. IV-B. Indoor NTNU-CSIE 2F for flash site-survey . . . . . . . . . . .. 16. IV-C. Indoor AS-IIS 7F for flash site-survey . . . . . . . . . . . . . .. 18. V. Discussion. 22. VI. Conclusion and Future Work. 25. Reference. 26.

(4) I. I NTRODUCTION Indoor localization has been researched for several decades. What we focus on the territory is the accuracy of indoor localization. The accuracy of indoor localization is not in our expectation until now. Many approaches are proposed to fulfill high accuracy. However, localization techniques are limited by some restrictions due to the signal characteristics and environment factors. We have discovered that there are two dissimilarities of approach applied to indoor localization technique. The one is model-based approach derived from transmission equation of Physics, and the other one is fingerprint-based approach which has been extensively used now. Model-based approach, however, is hedged about the presence of obstacles between anchor node and measuring node. According to path loss propagation model, there are three kinds of significant noises in realistic indoor environments which involves multipath fading, signal occlusions and signal diffractions. By contrast, fingerprint-based approach is adaptable to set up the environment settings. The basic algorithm of this approach is simply comparing the most resembling signal pattern from thorough site-survey. Wi-Fi-based indoor localization technologies has been developed for decades [2, 22, 26, 29]. The number of Wi-Fi APs (Accessing Points) are more enormous than decades before. Owning to the standard provided by 802.11/g/n, the accuracy of indoor localization is not capable of achieving satisfied error distance. There are some limitations existing in these standard. 802.11/ac which is the latest Wi-Fi standard has more features to localize objects. Due to the updating of the standard, Wi-Fi APs applied today has been changed into new standard increasingly. Therefore, indoor localization techniques have been evolved equally. Despite the emerging, Wi-Fi-based indoor localization techniques are still incomplete so far. Thus, we must seek sufficient indoor localization techniques which have more precise accuracy of localization. Fortunately, indoor localization researchers have developed other techniques based on different signal resources synchronously. Apart from Wi-Fi, several indoor localization works have been proposed including FM [5, 17, 20, 28], GSM [25], magnet [18], Bluetooth [3], visible light [11, 15],. 1.

(5) and other signal resources. The frequency of FM signal is around 100 MHz which wavelength is longer than Wi-Fi signal wavelength and GSM signal wavelength. Due to this fact, FM signal is not easily affected by the obstacles in comparison to WiFi or GSM. Popleteev et.al. claim that FM radio signals are less affected by weather condition, less sensitive to terrain conditions, and penetrates walls more easily than WiFi or GSM [20]. FM radio is popular and well-established technology. In comparison to FM, magnet indoor localization technique is prone to be affected by natural magnet field or artificial magnet field. We have to install extra sensor if we apply light indoor localization, Bluetooth indoor localization, and ZigBee indoor localization. In terms of [7], we can infer that the accuracy of DVB-T localization is better than GSM localization. We prefer signals without deploying extra anchor node. Hence, FM and DVB-T signals which based on existed launch station are the resources we primarily focusing. We aim to provide a novel system which based on FM and DVB-T signal resources. First, we have no requirements to build anchor node which can propagate signals. Second, due to both of FM and DVB-T signals are using low frequency, they have more ability of penetration obstacles which can provide stable signal resources. Third, the terminal equipments are cheap and relatively small. Indoor localization utilized the method which have employed in outdoor localization in the beginning. Previous works mainly applied model-based approach to perform localization. Because of the characteristics of indoor environment, the interactions of signal and obstacles result in unpredictable divination. Thus, some researchers do localization in another way which counts the error factors in the collected data. It was fingerprint-based approach that resolves the position in indirect solution. Rather the traditional model-based approach, the fingerprint-based approach shows the better result. Therefore, recent studies mainly utilize fingerprint-based approach. Nevertheless, fingerprint-based approach needs previous site-survey which is labor-intensive. Even so, FM indoor localization are proposed by previous work. Nonetheless, we propose a better solution than early research. We build a novel solution to localize indoor objects by using both FM and DVB-T signals. Then, we investigate the temporal. 2.

(6) effect on the data we collect in terms of age and period. We investigate the reason why window effect can be verified in our collecting data.. 3.

(7) II. R ELATED W ORK The techniques used in indoor localization can be roughly divided into two approaches. The one is model-based technique and the other one is fingerprint-based technique. Model-based techniques are including i) Time of Arrival (TOA) [4, 8, 21], ii) Angle of Arrival (AOA) [13, 14, 19], iii) Time Difference of Arrival (TDOA) [9], and iv) Frequency Difference of Arrival (FDOA). In order to gain enough information and perform localization, locations of three launch stations must be known in the beginning. Otherwise, these techniques are based on path loss propagation model, which is hard for deducing the accurate location. In realistic indoor environment, there are some significant noises which includes multipath fading, signal occlusions and signal diffractions. Therefore, the precise location in indoor environment is arduous to resolve. Generally, model-based approach described above needs prerequisite emitter (i.e., anchor node), and these corresponding localization results have been huge influenced by i) the obstacles between anchor node and measuring node, ii) the relative position between anchor node and measuring node, and iii) the time granularity provided by measuring nodes. By contrast, fingerprint-based approaches has more flexible to set environment settings. The basic algorithm of this pattern is simply to find out the most resembling signal pattern. Fingerprint-based approach can be categorized into four types according to their signal resources: 1) Wi-Fi-based 2) FM-based 3) other signal-based 4) multiple signal-based. A. Wi-Fi Fingerprints Almost every mobile device has embedded Wi-Fi module, and most of modern buildings today contain Wi-Fi Access Point (AP). Hence, the utilization of these WiFi signals should be an important objective. However, [2] is a pioneering research in type 1), and the system employs Received Signal Strength Indicator (RSSI) as the fingerprint of indoor localization. The result demonstrated the effectiveness of Wi-Fi fingerprinting by achieving localization accuracy in the range of 2 meters. The work has an overhead which broadcasting packets from Wi-Fi APs at each door location. In [29], authors proposed new probability-based model that identified an unknown. 4.

(8) location from RSSI probability map. Probabilistic fingerprinting properly modeled the stochastic variation of WiFi signals at the fingerprinting stage. Clustering of locations will improve performance. Furthermore, it significantly improve the accuracy of localization. However, recent study [22] showed that we could exploit the physical layer to fingerprint wireless channel. By fingerprinting wireless channel, we could enhance accuracy to below 1m. Chintalapudi et.al. in [6] proposed a novel system that doesn’t need to build fingerprints by exploiting existed LDPL equations. The objective of [6] is combining site information of Wi-Fi APs with model-based solution. However, the defect of this work requires enough number of Wi-Fi APs in order to cover the indoor environment and the model-based equation is not feasible to the indoor scenario. In [26], they compares the core localization algorithms with previous works. There are three kinds of pattern matching algorithm including kNN, MLP neural network and GRNN. kNN algorithm is outperformed than the other two algorithms. MLP neural network and GRNN require further time to train patterns. However, the result shows that if the pattern database is huge, and the accuracy obtained by three algorithm can be similar. In [16], not only discussed the model-based approach but also referred to the fingerprint-based approach. Fingerprint-based approach contains at least five localization algorithms. Besides previously conferred algorithms, there are probabilistic method, support vector machine (SVM), and smallest M-vertex polygon (SMP). Each of them contributes to the different accuracy, but the most feasible algorithms are kNN and SVM. The most of the important fact is that both of these algorithms are effective enough to compare with patterns. In [27], Wu et. al. claim that RSSI is not suitable for indoor localization. The reasons that not feasible lies in the variance of RSSI and the characteristic of RSSI which is tend to be affected by multipath effect. Thus, Wu et. al propose to employ Channel State Information (CSI) as indicator of indoor localization. The result of this work shows that it is more feasible to utilize CSI as collecting pattern indicator. However, CSI still requires to be calibrated which involves the path loss propagation model.. 5.

(9) B. FM Fingerprints Although Wi-Fi fingerprinting displayed impressive accuracy, it still has some limitations. It is possible to be affected by the presence of human or obstacles. Besides, the commercial APs show different power to localize objects. The RSSI varies with time, and blind spots are hard to sense. Thus, we seek for another signal resources to eliminate the limitations occurring in Wi-Fi fingerprinting. There have been growing research interests regarding FM-based outdoor localization. The advantages of outdoor localization is that the signals can cover extremely wide area (up to 100 km). Consequently, the FM-based signals can be retrieved in the indoor environment. Recent works [5, 28] show environment may contain two to three kinds of signals which can work together. By collaborated together, the result showed different compound of signals can achieve dissimilar results. In [5], the system use crowdsourcing technique for collecting location-annotated wireless signal fingerprints that form the localization pattern database. In [28], Yoon et. al. collect experiment data with both online and offline phases. Constructing fingerprint signature during offline phase, and autocalibrating when online phase. The former lacks of continuing calibration, and the latter lacks of verifying the dependency of parameter calibration. In [17], authors apply two kinds of machine learning methods including k-nearest neighbors (kNN) classification and Gaussian Process (GP) regression to analyze data, and the system performs autorecalibration afterwards. The deficiency of calibration phase lies in the process which is labor-intensive, and it requires expert knowledge and cannot be carried out often. In [20], they claim that FM is better than GSM or Wi-Fi in indoor localization. FM radio signals are less affected by weather conditions, such as rain or fog. This work gives us an idea that if the wavelength is close to the obstacles, the interaction can be strong resulting in complex interference patterns. The small indoor objects are transparent to long FM radio waves. The result shows that the accuracy is inconsistent when considering different environments of scale. Popleteev et. al. collect data by using three receivers simultaneously. Owning to using different receivers, the variance of three models must be concerned. The result also indicates that large-scale variations are caused by free-space path loss, while small-scale variations caused by radio wave. 6.

(10) interactions (reflections, diffractions, interferences). C. Other Fingerprints In [3], the solution using Bluetooth aims at determining room-level which is a meaningful granularity for a wide range of indoor applications. By considering the facts, the collected RSSI can result in imprecise location prediction. [3] Mortaza S et. al. think the RSSI of as unsuitable indicator. The work proposed a new indicator and used relative entropy measure (i.e., Kullback-Leibler function) and its extension (i.e., Jensen-Shannon distance measure) to estimate the location of the target device. In [12], Kamol et. al. confer the factors that affect the experiment result by using ZigBee signal. The work uses ZigBee Cluster Library (ZCL) to perform data collection. ZigBee localization system can be implemented using four different techniques as outlined by ZigBee Cluster Library Specification, which are the lateration, the proximity or the signposting, the RF fingerprinting, and out-of-band localization device. The accuracy of this work can be 77 cm error distance. We have made a research on GSM signals, which frequency is between 850 MHz and 1900 MHz. The main merit of utilizing GSM signals is that we could directly sense existed infrastructure. Each GSM cell broadcasts control packets at the maximum power through the Broadcast Control Channel (BCCH). As shown in [25], the system collects GSM fingerprints in a dense grid with 1.5 meters granularity. Moreover, in this work it collected wide fingerprints that include up to 29 different GSM channels in addition to the 6-strongest GSM cells. This work fulfills 5-10 meters in room-level localization. Besides, the result shown in acoustic background sound fingerprints [24] demonstrates that sound is feasible to room-level localization. It introduces the Acoustic Background Spectrum (ABS) fingerprints, which are compact, easily computed, robust to transient sounds and are able to distinguish remarkably similar rooms. Although it could achieve room-level localization, it still has some parameters to tune. In other words, the learning phase may take a long time. In [11], it deploys light sensors on predefined area that users without requiring extra infrastructure. The design of the pattern vector contains location and orientation.. 7.

(11) The prediction made by the system eliminates the possibilities of error and faces new situation. In the end, the system provides calibration algorithm to fix the error distance. In [15], it introduces the orientation error that should be considered. Luxapose uses optical angle-of-arrival (AoA) localization principles based on an ideal camera with a biconvex lens. Their positioning algorithm assumes that transmitter locations are known. This allows them to express the pairwise distance between transmitters in both transmitters’ and receivers’ frames of reference. Magnet Field can be one of signal resources [18]. Navarro, Danilo and Benet, Gines propose a testbed that can sense 3-dimensions magnetic, accelerometer and MPU vector. By means of these information and specific parameter, it could resolve location and height which can be used to localize 3-dimensions position. Nevertheless, the accuracy significantly is influenced by specific parameters. D. Hybrid Fingerprints As shown in [7], Fang et. al. utilized four different kinds of signals including DVBT, GSM, FM and WiFi signals. In addition, this work combined two of four kinds of signals to compare with accuracy. The result displayed another analysis of collected data, and authors added DVB-T signals to make the result more precise; however, it doesn’t support online phase to calibrate those collected data. The technique applied in this work is probabilistic model regression. In [1], the result proposed another type of fingerprint-based solution, and the pattern comes from variety of resources. Due to the multiple resources, the system requires different kinds of sensors and evaluation metrics. The result shows that the accuracy is far better than the result by using single fingerprint. Nevertheless, it is not feasible for existing devices because the schemes are so numerous that it is hard for combining or collecting now. In [30], Zheng et. al. employ RF and ultrasound signals as localization signal resources. This system proposed calibration-less solution by using ultrasound calibration. However, it exploits TOA in small scale which measures the distances between ultrasound receivers and target sensors. In this work, RF localization utilizes semisupervised learning technique in order to perform calibration-less solution. The basic. 8.

(12) idea of semi-supervised technique for manifold learning is that, if two data samples (RSS vector) are similar, then their labels (positions) should be similar. The work proposed a general method that how to incorporate dissimilar types of signals. In [10], they present a idea that indoor fingerprints should wisely choose pattern vector. Study shows that the accuracy won’t increase with number of AP increasing. This work applies Wi-Fi and Bluetooth as the localization signal resources. This work mainly addresses on the following areas of a typical fingerprint-based location system. i) Rather than utilizing absolute signal strength as location fingerprint, we argue that differences of signals perceived at APs would provide a more stable signature for any mobile device irrespective of its hardware used. ii) By smartly choosing good APs, a location system can benefit.. 9.

(13) III. M ETHODOLOGY The proposed localization system is based on fingerprints. To perform this process of localization, the system must be built a fingerprint pattern database which contains vectors of RSSI corresponding to each channel. Before building database, our system should be employed channel scanning for confirming how the number of channels are in this area. The scale of the database may result in different localization accuracy. After the database is set up, the system performs localization by collecting RSSI vector and comparing this vector with fingerprint pattern in database. The comparing metric is using nearest-neighbor method.. Fig. 1. Localization Flow Chart As shown in figure 1, the proposed system is composed of two phase including i) site-survey phase and ii) test phase. When the system is deployed to a new expanse, it requires channel information through channel survey step in i). The requirements in channel survey step is obtaining the number of channel. Once the system gains enough channel information, it begins to build fingerprint pattern which contains RSSI of each channel. After the fingerprint pattern is built by the system, it stores the pattern into database. The scale of the database depends on the period that collecting the RSSI in one place. The proposed system will sum up all the pattern vectors and take the average of it. In ii), end-user applying our system will collect vector of RSSI corresponding to each channel. The system will compare the measuring vector with pattern vector in. 10.

(14) Fig. 2. TerraTec Cinergy TStick+. Fig. 3. Data Collector. the database by exercising the euclidean distance. As generating corresponding values of distance, the system will employ nearest-neighbor method to resolve the location. A. Testbed Setup As shown in figure 2, the testbed we using is Terratec Cinergy TStick+ which is constructed by Realtek RTL2832U chip. There is an open source project developed on the chip which is called Software Defined Radio (SDR) rtl-sdr [23]. By applying SDR, the system controls which frequency should be detected. The merits of this library are 1) all of three operating system platforms work well and 2) the frequency bandwidth which can be perceived is abundant, including FM, DVB-T, GSM, ADS-B, GPS, LTE, etc. The RTL2832U outputs 8-bit I/Q-samples, and the highest theoretically possible sample-rate is 3.2 MS/s, however, the highest sample-rate without lost samples that has been tested so far is 2.56 MS/s. The frequency range is highly dependent of the used tuner, dongles that use the Elonics E4000 offering 52 - 2200 MHz with gap from 1100 MHz to 1250 MHz. B. Site-Survey Phase Before continuing site-survey, the system scans all the FM and DVB-T channels in one expanse. After the system confirms the number of channels, the system schedules to collect RSSI records according to α FM channels and β DVB-T channels every predetermined distance on a floor or in a small area. The reason for collecting RSSI records every predetermined distance is that the system requires to collect (α + β) RSSI entries to form a feature pattern which is described by RSSI record vector. The collected vectors are stored in database after labeling for every point of interest (POI).. 11.

(15) However, the RSSI record vectors the system collected are affected by the signal samples and the time range. Each RSSI record can be generated by specific number of samples. On the other hand, the vector contains several RSSI record. The system can collect a huge number of RSSI records on each channel and retrieve the average of RSSI records as one of values in a pattern vector. C. Testing Phase After the system builds up a fingerprint pattern database according to the site-survey, end-user collects RSSI record vector for localization. The resolving of localization is nearest-neighbor method. The system apply euclidean distance metric for judging the similarity of vector. Therefore, the system generates some distance, and the corresponding site label with smallest value of distance is the position it localized. As shown in equation 1, α is the number of FM channels, β is the number of DVB-T channels, and 0. Si is the collected vector and Sk,i is the pattern vector on location k. The estimation is the location resolved by our method. v uα+β uX 0 Estimation = avgk (mint (Si − Sk,i )2 ) i=1. 12. (1).

(16) IV. E VALUATION We set up the testbed, and we plan consecutive experiments to evaluate our methodology. The scenarios contains three parts: i) indoor scenario for extensive site survey, ii) indoor scenario for flash site survey, iii) and indoor scenario for flash site survey on inconsistent areas of floor. We make our experiments on 2F of department of computer science, National Taiwan Normal University (NTNU-CSIE) and 7F of Institute of Information Science, Academia Sinica (AS-IIS). In i), the system is focused on the accuracy it achieved. By means of extensive sitesurvey, the system obtains plentiful RSSI records to form fingerprint pattern. Owning to the extensive site-survey, it takes a long time to form a fingerprint pattern which is an overhead. After employing the extensive site-survey, the system can be inferred how accurate it can fulfill. In consequence, the system can be tested further. Therefore, ii) and iii) are proposed. The system exercises the site-survey in a short period of time but will have built the fingerprint pattern database for one month. The system can be examined how temporal effect affects the result. The number of channels are different between i) and ii), iii). We have 16 FM channels and 7 DVB-T channels in scenario i), while 28 FM channels and 6 DVB-T channels in scenario ii) and scenario iii). However, we need a evaluation metric to verify whether results are good. In most of localization works [2, 5, 7], they evaluate the localization by applying same metric for comparing with results. We concentrate on the error distance in all three scenarios. The system generating the result in i) is using five-fold cross validation. On the other hand, the results in ii) and iii) are being focused on temporal effect by using the fingerprint-based validation which is using previous 1 day, 1 week, 2 weeks, 3 weeks as the pattern and rest of data as the test. The system employs the same settings in scenario i) which POI are different from scenario ii). The settings of scenario ii) includes flash site survey and long period of collecting data, and the system applies same settings in scenario iii) which we split the collecting site into open area and lane area. The experiment settings are shown in table I.. 13.

(17) TABLE I E XPERIMENT S ETTINGS Scenario Num. of channel FM Frequency Num. of channel DVB-T Frequency Location Num. of Point Num. of Record Period. A 16 89.3 - 106.5 MHz 7 533 - 599 MHz NTNU-CSIE 2F 45 50 1 day. B 28 88.1 - 107.7 MHz 6 533 - 593 MHz NTNU-CSIE 2F 58 1 30 days. C 28 88.1 - 107.7 MHz 6 533 - 593 MHz AS-IIS 7F 30 1 30 days. Fig. 4. The map is in NTNU-CSIE 2F with 45 POI. A. Indoor NTNU-CSIE 2F for extensive site-survey The system scans all the FM and DVB-T channels in Taipei, and there was 16 FM channels and 7 DVB-T channels. Therefore, each of POI contains 30 RSSI records as one vector of POI. On each channel, we collect 50 RSSI records according to one channel. We perform the five-fold cross validation afterward. The experiments of the section are finished in 1 day. As shown in figure 4, we intend to collect data with 45 POI on this floor. The average distance of POI is 106 cm. We have no permission to make experiments on left-bottom lane. Therefore, we will test right-aisles which no buildings lean on the side. Figure 5 shows that the std is stable just as outdoor scenario. We can gain information from figure 6, and we observe the RSSI of DVB-T is weaker than RSSI of FM. This phenomenon indicates that signal fading in DVB-T is more evident than that of FM. We also discover that the point on channel 0 and location 23 in std figure is dark, and we infer that the signal may be caused by environment noises due to the wall. The. 14.

(18) Fig. 5. RSSI values are displayed in standard deviation figure of the 16 FM channels and of 7 DVB-T channels.. Fig. 6. RSSI values are displayed in mean figure of the 16 FM channels and of 7 DVB-T channels.. Fig. 7. The result of accuracy generated from scenario A is displayed in CDF figure. channel 18 is mapped to CTV in Taiwan which has strong signal in the mean figure. In other words, the programs in CTV is clear to watch. From figure 7, this result implies that the result of indoor localization is good under our setting. The result is make sense, because the size of DVB-T vector is smaller than the size of FM vector. Thus, the localization result of FM is better than the result of DVB-T. Furthermore, we draw a picture map on figure 7, and we can discover that DVB-T can achieve 96 percent of accuracy in our scenario. In section IV-A, the system has to obtain data at one POI for 10 minute which is time-consuming. So in this section, we try to eliminate site-survey time by collecting one RSSI record according to one channel. Moreover, the system re-scanned all the FM and DVB-T channels in Taipei. Thus, our pattern vector contains 28 FM channels and 6 DVB-T channels. The system performs the fingerprint-based validation which. 15.

(19) Fig. 8. The map is in NTNU-CSIE 2F with 58 POI.. Fig. 9. RSSI values are displayed in standard deviation figure of the 28 FM channels and of 6 DVB-T channels.. Fig. 10. RSSI values are displayed in mean figure of the 28 FM channels and of 6 DVB-T channels.. takes 1 day, 1 week, 2 weeks, 3 weeks as the pattern and the rest of data as the test, and draw the error bar figure in the end. Section IV-B collects data from 2014/08/09 to 2014/09/07, and section IV-C collects data from 2014/10/13 to 2014/11/20. B. Indoor NTNU-CSIE 2F for flash site-survey Figure 8 shows that we change the number of POI to 58. We make changes in settings which site-survey can be time efficient. Therefore, we are able to observe how temporal effect affects the accuracy. We have permission to test in all area of 2F. We have discover that near 23 and 46 show the dissimilar result because the wall. Figure 9 indicates that the temporal effect is not apparent on both FM and DVB-T channels. The temporal effect has more influence on high frequency of FM channels. As shown in Figure 10, the figure indicates that FM signal is more sensitive than DVBT signal. The figure implies that the capability of FM signal localization is better than. 16.

(20) Fig. 11. The figure shows that different size of DVB-T patterns generate inconsistent error distance in scenario B.. Fig. 12. The figure shows that different size of FM patterns generate inconsistent error distance in scenario B. that of DVB-T signal. The dark area in this figure corresponds to the POI which is near window. On the other hand, DVB-T signal is hard to be observed window effect. Figure 11 shows that indoor localization using DVB-T signals is not effective, and the highest accuracy appears when we use one week data as pattern. We can discover that if we apply one week training data as our pattern, the proposed system can estimate more accurate. If we apply more than one week training data as our pattern, the accuracy will drop. We infer that DVB-T signal pattern would effective only one week. Besides, the accuracy would drop. Figure 12 indicates that the FM signal can accomplish 400 cm in our settings and the accuracy becomes high when pattern data. Fig. 13. The figure shows that different size of DVB-T and FM patterns generate inconsistent error distance in scenario B.. 17.

(21) Fig. 14. The map is in AS-IIS 7F with 30 POI.. Fig. 15. RSSI values are displayed in standard deviation figure of the 28 FM channels and of 6 DVB-T channels in IIS aisle area.. Fig. 16. RSSI values are displayed in mean figure of the 28 FM channels and of 6 DVB-T channels in IIS aisle area.. gradually increasing. Comparing the result of figure 13 with previous results of single signal, we can discover that the accuracy produced by hybrid signals is higher than that of single signal. We can achieve 300 cm error distance when employing our setting. C. Indoor AS-IIS 7F for flash site-survey We divide this place, as shown in figure 14, into two area: a) aisle area and b) center area. In a), we collect data every 150 cm, on the other hand, we collect data every 50 cm in b). a) contains 18 POI and b) contains 12 POI. The roof of b) is higher than that of a). There are French windows in center area which may produce window effect. Furthermore, we eliminate a antenna on our testbed. We think that the result won’t affect by this factor and the testbed would be light when we test. 1) Aisle Area. 18.

(22) Fig. 17. The figure shows that different size of DVB-T patterns generate inconsistent error distance in aisle area of scenario C.. Fig. 18. The figure shows that different size of FM patterns generate inconsistent error distance in aisle area of scenario C. In figure 15, we can discover that both FM and DVB-T channels remains relatively stable. In low frequency of FM signal, the std is higher than high frequency of FM signal. We can infer from this figure that the ability of indoor localization with FM and DVB-T signals are resembling if we utilizing the same number of channels in the scenario. Figure 16 shows that RSSI of DVB-T channels are higher than RSSI of FM channels. The RSSI records in figure 16 are uniform in FM channels. In this scenario, the RSSI records of DVB-T signal are various, and this feature implies that DVB-T signal may more feasible to localize objects in the indoor environment. In figure 16, we can discover that there is a dark bar which maps to CTV. AS-IIS is near the DVB-T launch station so that the system would sense the inconsistent RSSI. Besides, the channel shows stable when we consider the std figure. Figure 17 shows that the error distance pattern are similar to figure 18. We use 6 channels to perform DVB-T indoor localization which the result similar to the result of FM indoor localization which using 28 channels. The result indicates that the DVB-T indoor localization is more effective than FM indoor. 19.

(23) Fig. 19. The figure shows that different size of DVB-T and FM patterns generate inconsistent error distance in aisle area of scenario C.. Fig. 20. RSSI values are displayed in standard deviation figure of the 28 FM channels and of 6 DVB-T channels in IIS leisure area.. Fig. 21. RSSI values are displayed in mean figure of the 28 FM channels and of 6 DVB-T channels in IIS leisure area.. localization. Figure 19 shows that the result is better than the results of indoor localization using single signal. The result produced by our system shows that DVB-T pattern works well in one week. We can discover that the error becomes high when employing two weeks training data as our pattern. 2) Leisure Area Figure 20 shows that the RSSI records are unstable for both FM and DVB-T signals. The outperform DVB-T channel remains high RSSI and stable. Figure 21 indicates that RSSI records of DVB-T signal are more sensitive to the obstacles than that of FM signal. Location index between 8 to 12 is near sofa, and we discover that the RSSI records generated by DVB-T signals are more diverse than RSSI records generated by FM signals. We can discover that FM and DVB-T are complementary, and DVB-T is stronger than FM which the result is dissimilar in NTNU-CSIE. We infer that the launch station is close to AS-IIS.. 20.

(24) Fig. 22. The figure shows that different size of DVB-T patterns generate inconsistent error distance in leisure area of scenario C.. Fig. 23. The figure shows that different size of FM patterns generate inconsistent error distance in leisure area of scenario C. Figure 22 implies that the error distance generated by DVB-T signal are larger than error distance produced by FM signal. The error distance can achieve under 50 cm when using both FM and DVB-T signals as shown in figure 24. The localization utilized by FM is rapidly increasing the accuracy. Combining DVBT and FM can achieve greater result.. Fig. 24. The figure shows that different size of DVB-T and FM patterns generate inconsistent error distance in leisure area of scenario C.. 21.

(25) V. D ISCUSSION As described in section IV, we use one day, one week, two weeks and three weeks as the location pattern saved in the fingerprint database. We use the rest of data as our test case by applying nearest-neighbor method to perform localization. Based on the experiment result, we have the following speculation: 1) The result shows that we can achieve the best accuracy if we choose one week data as our fingerprint pattern. The result produced by one day pattern also implies that we cannot get the representative data owing to the influence made by other factors. Otherwise, our system will make the wrong prediction if we over-sampled the data as our pattern. 2) We also realize that the accuracy in AS-IIS 7F center area is better than that of in aisle area. We infer that signal in aisle area tends to produce multipath fading and diffractions. Therefore, the system caused error when proceeding site-survey phase and test phase. We demonstrate an experiment which evaluates the localization according to the lane area. As shown in figure 26, we can reveal that the accuracies in lane 1 and lane 3 are better than the accuracies in lane 2, lane 4, and lane 5. The result indicates that the window effect has evident influence. Lane 1 and lane 3 are on the place with no buildings to lean on where others are not. Figure 28 shows the similar result as well. Lane 2 in AS-IIS 7F is obviously accomplishing the greater result than the prediction produced by lane 1 and lane 3. We discover that the result shows better if the system make the localization near windows. The result in AS-IIS 7F is better than NTNU-CSIE 2F which we think the window size is one of reasons that caused the different result. AS-IIS 7F has French windows which has better window effect where NTNU-CSIE 2F has only traditional windows. On the best situation, we can achieve 50 cm below accuracy by using FM and DVB-T signals. 3) As shown in table II, we compare our work with the result in [5]. The scenario in previous work is in campus office rooms and our work is in office open area. Chen et. al. employ the FM signal which contains four indicators, including. 22.

(26) Fig. 25. Map of NTNU-CSIE 2F. Fig. 26. CDF of NTNU-CSIE 2F according to Lane Area. Fig. 27. Map of AS-IIS 7F. Fig. 28. CDF of AS-IIS 7F according to Lane Area. Fig. 29. CDF of campus office with sensors deployed in rooms. 23. Fig. 30. CDF of office open area with finegrained data collection.

(27) TABLE II T HE COMPARISONS BETWEEN PREVIOUS WORK [5] Testbed Channel Type Num. of FM Channel Num. of Vector Evaluation Method. Previous work [5] SI-4735 Radio Receiver FM (RSSI, SNR, MULTIPATH, FREQOFF) Wi-Fi 32 32 * 4 + M Manhattan Distance. AND OUR WORK. Our work RTL2832U DVB-T Stick FM (RSSI) DVB-T 28 28 + 6 Euclidean Distance. RSSI, SNR, Multipath, Frequency Offset, each of them contains 32 channels. On top of that, this work also employs Wi-Fi signals and collects the RSSI from existing APs. Our work using FM and DVB-T signals, but we only employs RSSI in FM signals. Our sensing channels are dynamically changed in different environment which is more robust than the result in previous work. As shown in figure 29 and figure 30, we can discover that the result produced by our work is better than the result generated by previous work. Our system uses less number of channels and achieve the better accuracy. 4) We choose FM and DVB-T signal as our resource of indoor localization. In the beginning, we think the accuracy achieved by FM signal is good enough. If we want to accomplish more accurate result, we have to combine multi signals to form a hybrid system [1]. We consider the following conditions: i) The launch stations should be built by authority, ii) The signal should work well in indoor environment. The reason why we consider the first condition is that we think the information of launch station should be open for us to exploit and be prone to manage the number of stations. We think GSM, LTE and DVB-T signal should fit the first condition, but the GSM and LTE may not work well in indoor environment [20]. Thus, we think DVB-T should be the best choice. According to the evaluation result, we think DVB-T is a weak but stable signal. We think DVB-T signal should work well with FM signal in our system. However, our localization work flow doesn’t consider the channel diversity, and it is tend to cause our system to make incorrect anticipation.. 24.

(28) VI. C ONCLUSION AND F UTURE W ORK We propose a novel indoor localization system using both FM and DVB-T signals. We use single receiver to sample both FM and DVB-T signals thanks to the advance of software radio receiver hardware. We investigate how FM and DVB-T signals are working in different scenarios. We investigate the temporal effect on the training data in terms of age and period. The fingerprint pattern we collect more, the accuracy we achieve higher. On the other hand, we investigate the window effect of signals by evaluating the experiment in AS-IIS aisle area and open area. We think our methodology can be applied to the 3-D space. In section V, we have discussed the channel diversity. We believe that the system can be more accurate if the system applies the channel diversity. By reason of the SDR, the system testbed can be shaped in a smaller size. The system can be further developed to fingerprintless type by combining site information and model-based approach. Before manufactured, the system has to solve the presence of human or obstacles. We made experiment in a vast indoor environment without considering presence of human. We think that fingerprintless approach can solve this issue by applying model-based approach.. 25.

(29) R EFERENCES [1] M. Azizyan, I. Constandache, and R. Roy Choudhury. Surroundsense: mobile phone localization via ambience fingerprinting. In Proceedings of the 15th annual international conference on Mobile computing and networking, pages 261–272. ACM, 2009. [2] P. Bahl and V. N. Padmanabhan. Radar: An in-building rf-based user location and tracking system. In INFOCOM 2000. Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, volume 2, pages 775–784. IEEE, 2000. [3] M. S. Bargh and R. de Groote. Indoor localization based on response rate of bluetooth inquiries. In Proceedings of the first ACM international workshop on Mobile entity localization and tracking in GPS-less environments, pages 49–54. ACM, 2008. [4] Y.-T. Chan, W.-Y. Tsui, H.-C. So, and P.-c. Ching.. Time-of-arrival based. localization under nlos conditions. Vehicular Technology, IEEE Transactions on, 55(1):17–24, 2006. [5] Y. Chen, D. Lymberopoulos, J. Liu, and B. Priyantha.. Fm-based indoor. localization. In Proceedings of the 10th international conference on Mobile systems, applications, and services, pages 169–182. ACM, 2012. [6] K. Chintalapudi, A. Padmanabha Iyer, and V. N. Padmanabhan. Indoor localization without the pain. In Proceedings of the sixteenth annual international conference on Mobile computing and networking, pages 173–184. ACM, 2010. [7] S.-H. Fang and T.-N. Lin. Cooperative multi-radio localization in heterogeneous wireless networks. Wireless Communications, IEEE Transactions on, 9(5):1547– 1551, 2010. [8] S. Hara, D. Anzai, T. Yabu, K. Lee, T. Derham, and R. Zemek. A perturbation analysis on the performance of toa and tdoa localization in mixed los/nlos environments. Communications, IEEE Transactions on, 61(2):679–689, 2013. [9] K. Ho and Y. Chan. Geolocation of a known altitude object from tdoa and fdoa measurements. Aerospace and Electronic Systems, IEEE Transactions on,. 26.

(30) 33(3):770–783, 1997. [10] A. M. Hossain, H. N. Van, Y. Jin, and W.-S. Soh. Indoor localization using multiple wireless technologies. In Mobile Adhoc and Sensor Systems, 2007. MASS 2007. IEEE Internatonal Conference on, pages 1–8. IEEE, 2007. [11] A. Jim´enez, F. Zampella, and F. Seco.. Light-matching: A new signal of. opportunity for pedestrian indoor navigation. In Indoor Positioning and Indoor Navigation (IPIN), 2013 International Conference on, pages 1–10. IEEE, 2013. [12] K. Kaemarungsi, R. Ranron, and P. Pongsoon. Study of received signal strength indication in zigbee location cluster for indoor localization. In Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON), 2013 10th International Conference on, pages 1–6. IEEE, 2013. [13] R. Klukas and M. Fattouche. Line-of-sight angle of arrival estimation in the outdoor multipath environment. Vehicular Technology, IEEE Transactions on, 47(1):342–351, 1998. [14] S. Kotwal, S. Verma, S. Suryansh, and A. Sharma. Region based collaborative angle of arrival localization for wireless sensor networks with maximum range information. In Computational Intelligence and Communication Networks (CICN), 2010 International Conference on, pages 301–307. IEEE, 2010. [15] Y.-S. Kuo, P. Pannuto, K.-J. Hsiao, and P. Dutta. Luxapose: Indoor positioning with mobile phones and visible light.. In Proceedings of the 20th Annual. International Conference on Mobile Computing and Networking, MobiCom ’14, pages 447–458, New York, NY, USA, 2014. ACM. [16] H. Liu, H. Darabi, P. Banerjee, and J. Liu. Survey of wireless indoor positioning techniques and systems. Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, 37(6):1067–1080, 2007. [17] A. Matic, A. Papliatseyeu, V. Osmani, and O. Mayora-Ibarra. Tuning to your position: Fm radio based indoor localization with spontaneous recalibration. In Pervasive Computing and Communications (PerCom), 2010 IEEE International Conference on, pages 153–161. IEEE, 2010. [18] D. Navarro and G. Benet. Magnetic map building for mobile robot localization. 27.

(31) purpose. In Emerging Technologies & Factory Automation, 2009. ETFA 2009. IEEE Conference on, pages 1–4. IEEE, 2009. [19] R. Peng and M. L. Sichitiu. Angle of arrival localization for wireless sensor networks. In Sensor and Ad Hoc Communications and Networks, 2006. SECON’06. 2006 3rd Annual IEEE Communications Society on, volume 1, pages 374–382. IEEE, 2006. [20] A. Popleteev, V. Osmani, and O. Mayora. Investigation of indoor localization with ambient fm radio stations. In Pervasive Computing and Communications (PerCom), 2012 IEEE International Conference on, pages 171–179. IEEE, 2012. [21] H. Schau and A. Robinson. Passive source localization employing intersecting spherical surfaces from time-of-arrival differences. Acoustics, Speech and Signal Processing, IEEE Transactions on, 35(8):1223–1225, 1987. [22] S. Sen, B. Radunovic, R. R. Choudhury, and T. Minka. You are facing the mona lisa: spot localization using phy layer information. In Proceedings of the 10th international conference on Mobile systems, applications, and services, pages 183–196. ACM, 2012. [23] M. Steve, S. Dimitri, and Hoernchen. rtl-sdr http://sdr.osmocom.org/trac/wiki/rtlsdr. [24] S. P. Tarzia, P. A. Dinda, R. P. Dick, and G. Memik. Indoor localization without infrastructure using the acoustic background spectrum. In Proceedings of the 9th international conference on Mobile systems, applications, and services, pages 155–168. ACM, 2011. [25] A. Varshavsky, E. de Lara, J. Hightower, A. LaMarca, and V. Otsason. Gsm indoor localization. Pervasive and Mobile Computing, 3(6):698–720, 2007. [26] G. I. Wassi, C. Despins, D. Grenier, and C. Nerguizian. Indoor location using received signal strength of ieee 802.11 b access point. In Electrical and Computer Engineering, 2005. Canadian Conference on, pages 1367–1370. IEEE, 2005. [27] K. Wu, J. Xiao, Y. Yi, M. Gao, and L. M. Ni.. Fila: Fine-grained indoor. localization. In INFOCOM, 2012 Proceedings IEEE, pages 2210–2218. IEEE, 2012.. 28.

(32) [28] S. Yoon, K. Lee, and I. Rhee.. Fm-based indoor localization via automatic. fingerprint db construction and matching. In Proceeding of the 11th annual international conference on Mobile systems, applications, and services, pages 207–220. ACM, 2013. [29] M. Youssef and A. Agrawala. The horus wlan location determination system. In Proceedings of the 3rd International Conference on Mobile Systems, Applications, and Services, MobiSys ’05, pages 205–218. ACM, 2005. [30] V. W. Zheng, J. Zhao, Y. Wang, and Q. Yang. Hips: A calibration-less hybrid indoor positioning system using heterogeneous sensors. In Pervasive Computing and Communications, 2009. PerCom 2009. IEEE International Conference on, pages 1–6. IEEE, 2009.. 29.

(33)