Radio Map Update Automation for WiFi Positioning ... - IEEE Xplore

10 downloads 36522 Views 723KB Size Report
Radio Map Update Automation for WiFi Positioning Systems. Jun-Sung Lim, Woo-Hyuk Jang, Gi-Wan Yoon, and Dong-Soo Han. Abstract—This paper presents ...
IEEE COMMUNICATIONS LETTERS, VOL. 17, NO. 4, APRIL 2013

693

Radio Map Update Automation for WiFi Positioning Systems Jun-Sung Lim, Woo-Hyuk Jang, Gi-Wan Yoon, and Dong-Soo Han

Abstract—This paper presents a novel method to reduce the recalibration costs of a radio map by automatically updating the radio map. The appearance frequencies of access points (APs) detected from user feedback data are mainly used for the update. The proposed method appeared superior to previous methods, especially in its ability to update newly installed APs in the radio map. According to the experiment conducted for the radio map of 233 Seoul subway stops, the proposed method was effective for updating APs with weak as well as strong signal strengths. Index Terms—Indoor environments, wireless LAN, WiFi positioning, radio map, database updating, implicit user feedback.

I. I NTRODUCTION ITH the proliferation of smartphones and the explosively increasing WiFi hot spots, the so-called WiFi positioning system (WPS) has recently received great attention. In the WPS, due to its superior accuracy, the fingerprintbased localization technique is preferred rather than trilateration. In fingerprint-based localization, the location of a device is usually estimated by referring to a radio map, which is the collection of the fingerprints and their collected location information. Thus, the radio map is essential for the fingerprint-based localization technique. A fingerprint is a set of pairs of the received signal strength (RSS) and the MAC address (BSSID) of access points (APs). Recalibration should be made occasionally in the fingerprint-based localization to maintain the initial accuracy of the WPS, irrespective of the changes of the APs. However, the recalibration usually requires considerable time and effort, especially with the WPS covering a large-scale space. Some researchers have attempted to address this issue by introducing additional referencing devices such as RF[1], Bluetooth[2], or WiFi based indoor positioning systems[3], [4]. Using such referencing devices is simple and effective, but its cost is too high to cover a wide area. There has been another research stream on the recalibration. In this stream, researchers have tried to implicitly or explicitly acquire the fingerprints from users’ mobile devices as feedback data, then they substituted the newly captured fingerprints for the old ones for updating their radio maps[5], [6], [7]. However, one noticeable critical problem of this approach is how to correctly update the newly captured fingerprints into the radio map. In this work, we propose a novel method to more effectively reduce the recalibration cost based on an automatic update of

W

Manuscript received December 23, 2012. The associate editor coordinating the review of this letter and approving it for publication was H. Wymeersch. J.-S. Lim, W.-H. Jang, and D.-S. Han are with the Department of Computer Science, and G.-W. Yoon is with the Department of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), 291 Daehakro, Yuseong-gu, Daejeon, Korea. D.-S. Han is the corresponding author (email: [email protected]). Digital Object Identifier 10.1109/LCOMM.2013.022213.122885

radio maps in response to possible changes of WiFi environments. The appearance frequencies of the Access Points (APs), collected from the user feedback data, can be the main sources for updating radio maps. The updating experiments have been comprehensively performed on the radio map constructed across the 233 Seoul subway stops. As compared to previous methods, the proposed method turned out to be superior in its ability to update newly installed APs into the radio map. Also, the proposed method could more effectively be used for the APs with weak signal strengths as well as for the APs with strong signal strengths. II. M ETHODOLOGY A. Problem description Suppose a preliminary radio map has already been constructed, and a WPS is running on the radio map. When a WiFi environment changes due to the removal of any existing APs and/or the installation of new APs, the accuracy of the WPS may gradually deteriorate. Thus, to prevent such accuracy degradations, the changes of the WiFi environment should be reflected in the radio map from time to time. The goal of our proposed method is focused on automatically updating the radio map using user feedback. Therefore, we assume that the fingerprints are captured by smart phones, and the captured fingerprints are delivered to a server along with the information of time stamps, events, and confidence values. Based on the fingerprints collected from any users, we need to decide from the feedback data which APs should be added into the radio map and also which APs should be removed from the radio map. B. Appearance frequency based WiFi radio map update The easiest way of updating the radio map is to overwrite the old fingerprints with the new ones captured through manual recalibration. Even though we ignore the drawback of tremendous recalibration effort, but a problem still remains: it may miss some installed APs with weak signal strengths or even include temporally presented APs like mobile hotspots in the update process. Gallagher et al.[5] proposed an algorithm in an effort to automatically update a radio map using the fingerprints contributed by users. In order to update the radio map, they added two new fields; score and pending to each AP as its attributes in the existing radio map. The score value of an AP increases or decreases whenever the AP is present or absent, respectively, in user feedback. If the score of an AP reaches max, the AP’s pending value is set to f alse, and the pending value remains f alse until the score decreases to 0. If the score of an AP is under max, the pending value is true.

c 2013 IEEE 1089-7798/13$31.00 

694

IEEE COMMUNICATIONS LETTERS, VOL. 17, NO. 4, APRIL 2013

Algorithm 1 UpdateRadioMap(apsuser ) 1: apsuser = {ap|ap obtained by a user feedback, where ap =

(bssid, rssuser )}

2: apsDB = {ap|ap is included in the radio map, where ap =

(bssid, rssDB , score, base, f reqin , f reqout , δ, pending)}

(a) Gallagher et al.

(b) Current study Fig. 1. Plots of the variation of AP scores as the amount of user feedback increases.

Gallagher et al. used only the APs whose pending states are f alse for location estimation. In addition, this method delays the adding of the newly captured APs, which is necessary to prevent any temporal APs from being included for the updating of the radio map. This method turned out to be very effective in some situations, but it is too simple to cover various situations in reality. Moreover, the APs with weak signals were not taken into account in their method. For this reason, the APs whose signal acquisition probability is below 50% (about -80dBm) cannot be added into the radio map. As illustrated in Fig. 1(a), there is no chance for AP3 to be added into the radio map with Gallagher et al.’s method. In contrast, in our method, the APs with the weak signals were considered as well as the APs with strong signals (alg. 1). In fact, the signal acquisition probability of an AP varies depending mainly on its RSS, even if it is captured at a fixed point. When we analyzed the 2,200 AP signals collected from a large-scale shopping mall (COEX, Seoul, Korea, 189,000m2), we found that the lower the signal strength of an AP is, the lower its signal acquisition probability is. Hence, the APs with weak signal strengths should be given more chances to be updated into the radio map. In order to update the newly installed or removed APs into the radio map, we have devised more refined scoring functions. Basically, the score of an AP may increase or decrease, depending on the AP appearance patterns in user feedback. For instance, when an AP is newly detected in a users feedback, or when an AP which may already exist in the radio map is detected in the user feedback, the score increasing function increases the score of the AP (lines 9, 18). The increment is defined by the

3: for all ap such that ap ∈ apsuser ∪ apsDB do 4: if ap ∈ apsuser and ap ∈ apsDB then 5: if f reqin = 0 then 6: f reqin ← 1, f reqout ← 0, base ← score 7: rssDB ← (rssDB + rssuser )/2 8: δ ← 1/acquisition probability(rssDB ) × 100 9: score ← base + (ef reqin ++ − 1) 10: if pending = true and score ≥ min then 11: pending ← f alse 12: else if score > max then 13: score ← max 14: else if ap ∈ apsuser and ap ∈ / apsDB then 15: base ← 0, pending ← true, f reqin ← 1, f reqout ← 0 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27:

rssDB ← rssuser δ ← 1/acquisition probability(rssDB ) × 100 score ← base + (ef reqin ++ − 1) apsDB ← apsDB ∪ {ap} else if ap ∈ / apsuser and ap ∈ apsDB then if f reqout = 0 then f reqout ← 1, f reqin ← 0, base ← score score ← base − (ef reqout ++/δ − 1) if pending = f alse and score < min then pending ← true else if score ≤ 0 then apsDB ← apsDB − {ap}

function: ef reqin ++

(1)

where f reqin is the number of appearances of a particular AP in a row. When an AP which has not appeared in previous user feedback appears in current user feedback, the f reqin of the AP is set to 1 (line 6). Afterwards, if the AP appears again in a consecutive manner, the f reqin of the AP increases one at a time (lines 9, 18). If the AP has not appeared in subsequent user feedback, the f reqin of the AP is set back to 0 again (line 22). Meanwhile, in order to set the initial score value of the undetected AP in user feedback to 0, we subtract 1 from (1). (2) ef reqin ++ − 1 However, the score value of the AP is reset to 0 whenever an AP detected in previous user feedback does not appear in subsequent user feedback. To prevent this, we store the previous score value in base whenever the score value is reset to 0. Thus, the score increasing function is defined by base + (ef reqin ++ − 1)

(3)

where the previous score value is stored in base. The score decreasing function decreases the score of an AP existing in the radio map when the AP has not appeared in current user feedback (line 23). The score decreasing function looks very similar to the score increasing function, but the parameter of the exponential function is not f reqin but f reqout /δ: base − (ef reqout ++/δ − 1)

(4)

LIM et al.: RADIO MAP UPDATE AUTOMATION FOR WIFI POSITIONING SYSTEMS

Fig. 2. Plots of the variation of AP scores for 90 days as the amount of user feedback increases (Dongguk University subway stop, Seoul, Korea).

where f reqout is the consecutively undetected times of an AP in user feedback, and δ controls the decreasing rate of the score value. The f reqout has the opposite role of f reqin in the sense that its value increases as the APs are consecutively undetected in user feedback. If either f reqout or f reqin has a non-zero value, the other must be set to 0 (lines 6, 22). In order to give an opportunity for the APs with low signal strengths to be added into the radio map, we introduced a control value δ to control the decreasing rate of the scoring function. The decreasing rate of score is smaller than the increasing rate by δ. This means that if the APs with weak signal strengths appear in the user feedback once, our algorithm waits for the score of the APs to be increased, even if the APs have not appeared in subsequent user feedback for δ times. In addition, δ allows us to keep the APs with weak signals within the survival window (in between min and max) for a long time until the scores of the APs are less than min. Fig. 1(b) illustrates the situation in which AP3 can be added and then can stably stay within the range of the survival window, compared to the case of Fig. 1(a). The control value of an AP can be set to the reciprocal of the signal acquisition probability of the AP × 100 (lines 8, 17). The acquisition probability at lines 8, 17 is a function which takes the RSS of an AP stored in the radio map and returns its signal acquisition probability. In our algorithm, the signal strength of an AP in the radio map is updated whenever the AP appears in user feedback. If an AP appears in user feedback for the first time, its RSS value becomes the RSS value of the AP in the radio map (line 16); otherwise, we compute the average RSS value of the AP in the radio map and the newly captured one from user feedback (line 7). To take advantage of the capability of our scoring functions, we added the following six new fields into each AP entry in the existing radio map (line 2): score, base, f reqin , f reqout , δ, and pending. The pending state is set to f alse if the score of an AP is within the survival window; otherwise, it is set to true (lines 8, 20). Only an AP whose pending state is false can be used for localization because the AP should be regarded as an installed AP. Our algorithm also introduces two constants, min and max, that need to be tuned depending on the WiFi environments. The min is used to prevent the temporal APs from being added into the radio map. If the value of min is small, the update time for the inclusion of a newly installed AP into the

695

radio map will be reduced. However, it will also increase the possibility for a temporal AP to be included in the radio map. The max is the cap on the score value. The value can be set to be greater than min + (e − 1), where e − 1 is the increment of an AP, when the f reqin value of the AP is 1. Thus, if the score of an AP touches max, the AP is guaranteed to stay in the radio map by at least δ times. If the value of max is small, the removal time of the removed APs is shortened, but instead, the possibility for the existing APs, especially the APs with weak signal strengths, to be incorrectly removed from the radio map will grow. On the other hand, if the value of max is big, it will have the opposite effect. Fig. 2 shows the scores of some APs at Dongguk University subway stop in Seoul, Korea; AP1, AP2, and AP3 represent the installed APs; AP4 represents the removed AP; and AP5 to AP8 represent the temporal APs. As shown in this figure, our algorithm was superior in its ability to reflect the changes of the WiFi environment. III. E VALUATION A. Experiment set up We compared the performance of the three methods - our proposed algorithm, Gallagher et al.’s algorithm, and manual recalibration - in terms of the accuracy of the method reflecting AP environment changes to the radio map. The feedback data were collected from 30,000 users of a subway navigation system, that is, Android app, "Chihacheol Nerimi" 2011 developed for Seoul subways, Korea. This system notifies the final stops on users’ smart phones when the train approaches their final destinations. For the development of the system, we had first collected WiFi fingerprints from 537 Seoul subway stops and then constructed a radio map. Once the WiFi fingerprints are captured by a smart phone, the system can recognize the arrival of the subway at the destination using the radio map. In order to collect the feedback data, we embedded a feedback collecting routine that collects fingerprints every 7 seconds and delivers the data to the server. To prevent any possible inclusion of temporal APs due to any user contributions made in a very short period of time, the server merges the feedback data every 10 minutes and then updates the radio map. The data was continuously collected for a period of over three months. Among the 537 subway stops, we selected only 233 subway stops with a large number of user feedback for more precise evaluation. When we analyzed the accumulated feedback data, it revealed that 3,398 APs already existed, the 244 APs were newly installed, the 93 APs were removed, and the 152,931 APs were temporally presented in the 233 subway stops for the three months. B. Experimental results Fig. 3 plots the accuracy changes of updating the newly installed APs as the amount of user feedback increases. We adjusted the value of min to 21 in our algorithm and the value of max to be 4 in Gallagher et al.’s algorithm. We also set the max to be 23.7 in our algorithm. Those values were determined through an empirical study of the appearance patterns of APs at each subway stop. For the evaluation of

696

IEEE COMMUNICATIONS LETTERS, VOL. 17, NO. 4, APRIL 2013

Fig. 3. Accuracy changes of the three methods in updating the newly installed APs as the amount of user feedback increases.

Fig. 4.

Precision-recall curves of the three methods.

the manual recalibration, we regarded the user feedback as the manual recalibration results. That is, we substituted the old fingerprint with the new one whenever user feedback was received. As shown in Fig. 3, our method was shown to be almost always successful in adding the newly installed APs into the radio map and keeping those APs in the radio map in a stable manner, which is attributed to the adoption of both the refined scoring function and the appropriate size of the survival window. On the other hand, the method suggested by Gallagher et al. was not always successful in updating the newly installed APs into the radio map, and it also has a difficulty in adding the APs with low acquisition probabilities into the radio map. Particularly, for the manual recalibration, the accuracy of updating the newly installed APs into the radio map was severely influenced by the acquisition probability of APs at the recalibration time. This indicates that the accuracy of manual recalibration is very unstable and unpredictable. For more in-depth analysis, we compared the precisions and recalls of the aforementioned three methods. Each AP that appears in the user feedback can be either an actually installed AP or a temporal AP. The outcomes of the update methods will be either positive (i.e., predicting that the AP is an installed AP) or negative (i.e., predicting that the AP is not an installed AP). Thus, an AP detected in the user feedback belongs to one of the following four categories; (i) True Positive (TP): installed AP and correctly identified as Installed AP, (ii)

False Positive (FP): temporal or removed AP and incorrectly identified as installed AP, (iii) True Negative (TN): temporal or removed AP and correctly identified as one of them, and (iv) False Negative (FN): installed AP and incorrectly identified as temporal or removed AP. The precision and recall were computed by the equations T P/(T P + F P ) and T P/(T P +F N ), respectively. For all of the 233 subway stops, we calculated the precision and recall whenever user feedback was received. Then we averaged the values to obtain the final precision and recall values. Fig. 4 shows the precision-recall curves for the three methods. The results were obtained changing the min value in our algorithm, and the max value in Gallagher et al.’s algorithm from 0 to infinite, respectively. In both our algorithm and Gallagher et al.’s algorithm, the recall and precision values changed because the increment of min and max values increased the FN values, but decreased the TP and FP values. On the other hand, the precision and recall values of manual recalibration had static values because these values were not influenced by the min and max values. Overall, our approach achieved much better results compared to Gallagher et al.’s method in that our algorithm had high precision and recall values. On the other hand, the manual recalibration showed the worst performance because all of the APs in the user feedback were considered as the installed APs in this method. IV. C ONCLUSION In fingerprint-based localization, keeping WiFi radio map up-to-date is one of the key issues. For more complete updating of the WiFi radio map, we need to consider the APs with weak signals as well as strong signals. Furthermore, it is necessary to recognize both the temporal and permanent AP changes. The method proposed in this work has been demonstrated to outperform other update strategies, particularly in its capability of dealing with both strong and weak signals. Thus, we expect that the proposed method can be also used for a variety of location-based services such as WiFi-based indoor navigation and bus stop notification. R EFERENCES [1] H.-S. Ahn and W. Yu, “Environmental-adaptive RSSI-based indoor localization,” IEEE Trans. Automation Sci. Engineer., vol. 6, no. 4, pp. 626–633, 2009. [2] A. M. Hossain, H. N. Van, and W.-S. Soh, “Utilization of user feedback in indoor positioning system,” Pervasive and Mobile Computing, vol. 6, no. 4, pp. 467–481, 2010. [3] Y.-C. Chen, J.-R. Chiang, H.-H. Chu, P. Huang, and A. W. Tsui, “Sensorassisted Wi-Fi indoor location system for adapting to environmental dynamics,” in Proc. 2005 ACM MSWiM, pp. 118–125. [4] J. Yin, Q. Yang, and L. Ni, “Adaptive temporal radio maps for indoor location estimation,” in Proc. 2005 PerCom, pp. 85–94. [5] T. Gallagher, B. Li, A. G. Dempster, and C. Rizos, “Database updating through user feedback in fingerprinting-based Wi-Fi location systems,” in Proc. 2010 UPINLBS. [6] R. Hansen, R. Wind, C. S. Jensen, and B. Thomsen, “Algorithmic strategies for adapting to environmental changes in 802.11 location fingerprinting,” in Proc. 2010 IPN, pp. 1–10. [7] J.-g. Park, B. Charrow, D. Curtis, J. Battat, E. Minkov, J. Hicks, S. Teller, and J. Ledlie, “Growing an organic indoor location system,” in Proc. 2010 MobiSys, pp. 271–284.