sensors Article

Autonomous Landmark Calibration Method for Indoor Localization Jae-Hoon Kim 1, * and Byoung-Seop Kim 2 1 2

*

Department of Industrial Engineering, Ajou University, Suwon 16499, Korea Infrastructure Protection Team, Korea Internet and Security Agency, Seoul 05717, Korea; [email protected] Correspondence: [email protected]; Tel.: +82-10-3129-7996

Received: 21 June 2017; Accepted: 21 August 2017; Published: 24 August 2017

Abstract: Machine-generated data expansion is a global phenomenon in recent Internet services. The proliferation of mobile communication and smart devices has increased the utilization of machine-generated data significantly. One of the most promising applications of machine-generated data is the estimation of the location of smart devices. The motion sensors integrated into smart devices generate continuous data that can be used to estimate the location of pedestrians in an indoor environment. We focus on the estimation of the accurate location of smart devices by determining the landmarks appropriately for location error calibration. In the motion sensor-based location estimation, the proposed threshold control method determines valid landmarks in real time to avoid the accumulation of errors. A statistical method analyzes the acquired motion sensor data and proposes a valid landmark for every movement of the smart devices. Motion sensor data used in the testbed are collected from the actual measurements taken throughout a commercial building to demonstrate the practical usefulness of the proposed method. Keywords: location-based services; data analytics; landmark; dead reckoning

1. Introduction A location-based service (LBS) improves the economic and emotional utility of services in various user applications. LBSs are usually offered by cellular radio providers. In today’s LBSs, the typical cell-based location estimation takes advantage of various techniques: the global navigation satellite system (GNSS), Wi-Fi based fingerprint/triangulation, magnetic field fingerprints, and other complicated techniques for the sensitive measuring of the geographical location of a device. The LBS includes much useful information about specific locations in the user’s neighborhood. Maps and navigation are the essential service capabilities that permit the fundamental location management. Map and navigation services have sufficient expandability to provide asset tracking, autonomous driving and context aware services. Augmented reality (AR) is a prospective LBS service. An accurate geographical location enhances AR applications. More geographically appropriate information or context can be added to the display of a mobile device. The additional context, which is highly related to locations, provides vast opportunities in gaming, advertising, and social applications. The strong usability offered by the LBSs provides huge market expansion. In addition to popular map/navigation services, food delivery, car sharing, and many other Internet of Thing applications are prevailing among the masses. However, the popularity of LBS is generally focused on the outdoor environment. Indoor spaces require high location precision because the environmental contexts change with a much finer spatial granularity. Just a five-meter difference may indicate just two different aisles in a general grocery store. Many research trials have been implemented to achieve affordable location accuracy. Image-based location tracking is a promising approach. Guoyu et al. [1,2] proved that the high location accuracy can be achieved by using captured video for indoor environments. Hongbo et al. [3,4] developed an advanced Wi-Fi fingerprint location

Sensors 2017, 17, 1952; doi:10.3390/s17091952

www.mdpi.com/journal/sensors

Sensors 2017, 17, 1952

2 of 15

environments. Hongbo Sensors 2017, 17, 1952

et al. [3,4] developed an advanced Wi-Fi fingerprint location estimation 2 of 14 method using peer assistance. Moreover, Yuanchao et al. [5] enhanced the Wi-Fi fingerprint method by adopting a gradient for each Wi-Fi fingerprint to eliminate biased measurements. Probabilistic estimation method using peer assistance. Moreover, Yuanchao et al. [5] enhanced the Wi-Fi fingerprint fingerprint location estimation methods are also proposed. The studies of Seco et al. [6], Roos et al. [7], method by adopting a gradient for each Wi-Fi fingerprint to eliminate biased measurements. Probabilistic and Youssef et al. [8] estimated the probability distribution of the measured fingerprints given the fingerprint location estimation methods are also proposed. The studies of Seco et al. [6], Roos et al. [7], user locations. Coverage estimation models were developed by Muller et al. [9] and Koski et al. [10]. and Youssef et al. [8] estimated the probability distribution of the measured fingerprints given the user They estimated the probabilistic coverage of Wi-Fi access points and then picked the user locations locations. Coverage estimation models were developed by Muller et al. [9] and Koski et al. [10]. They using probabilistic intersections of access point coverages. As recent approaches, machine learning estimated the probabilistic coverage of Wi-Fi access points and then picked the user locations using algorithms were applied to enhance fingerprint location estimation. Feng et al. [11] suggested a probabilistic intersections of access point coverages. As recent approaches, machine learning algorithms support vector machine model and an interpolation method to reduce estimation errors. were applied to enhance fingerprint location estimation. Feng et al. [11] suggested a support vector Sanchez-Rodriguez et al. [12] proposed a multiple weighted decision tree model for indoor machine model and an interpolation method to reduce estimation errors. Sanchez-Rodriguez et al. [12] location estimation. proposed a multiple weighted decision tree model for indoor location estimation. The fingerprint location estimation is not restricted to Wi-Fi signals. Magnetic field fingerprints The fingerprint location estimation is not restricted to Wi-Fi signals. Magnetic field fingerprints use use magnetic field distributions inside buildings. The metal structure of the indoor environment magnetic field distributions inside buildings. The metal structure of the indoor environment cause magnetic cause magnetic field disturbances [13,14]. The results of previous studies have provided relatively field disturbances [13,14]. The results of previous studies have provided relatively stable location estimation stable location estimation methods [15–17]. The study of Torres-Sospedra et al. [18] presented stable methods [15–17]. The study of Torres-Sospedra et al. [18] presented stable continuous estimation using continuous estimation using small mobile devices. However, Gozick et al. [15] showed the the small mobile devices. However, Gozick et al. [15] showed the the magnetic field fingerprint restrictions magnetic field fingerprint restrictions caused by limited measuring features: the direction of the caused by limited features: the direction the magnetic field is restricted in two dimensions in magnetic field is measuring restricted in two dimensions in of many real life scenarios. Another challenge if using many real life scenarios. Another challenge if using a magnetic field for location estimation with mobile a magnetic field for location estimation with mobile devices is significant variability of the magnetic devices is significant variability of the field [17]. RecentWang studies tried to address these field [17]. Recent studies have tried tomagnetic address these challenges. et have al. [19] applied a weighted challenges. Wang al. [19] the applied a weighted particle of filter to reducefields. the irregular fluctuation of magnetic particle filter to etreduce irregular fluctuation magnetic Madson and Rahanani [20] fields. Madson and Rahanani [20] suggested an adaptive dipole model for error compensation. suggested an adaptive dipole model for error compensation. Sensor-based Sensor-based location location estimation estimation is is another another general general approach approach for for indoor indoor spaces. spaces. An An inertial inertial measurement unit (IMU) that is embodied in every smart device measures the velocity, orientation, measurement unit (IMU) that is embodied in every smart device measures the velocity, orientation, and and acceleration acceleration of of the the device. device. Using Using the the IMU, IMU, we we can can estimate estimate the the movement movement of ofobjects. objects. This This estimation method is named dead reckoning [21–25]. By using the pedestrian working model, we estimation method is named dead reckoning [21–25]. By using the pedestrian working model, we can can estimate estimate the the movement movement or or orientation orientation of of aa device device without without any any infrastructure. infrastructure. Fuchs Fuchs et et al. al. identified identified the the common common requirements requirements of of indoor indoor tracking tracking in in mission mission critical critical scenarios scenarios and and introduced introduced the the basic basic techniques techniques for for IMU IMU location location estimation estimation [26]. [26]. Beauregard Beauregard et et al. al. proposed proposed aa combined combined approach approach of of pedestrian dead reckoning reckoning and andGPS GPSlocation locationestimation estimation [27]. accelerometer provides signals pedestrian dead [27]. AnAn accelerometer provides signals and and a neural network predicts the length step length for relative indoor positioning. Jiménez et al.aused a neural network predicts the step for relative indoor positioning. Jiménez et al. used lowaperformance low-performance micro-electro-mechanical sensor (MEMS) attached to the foot of a person [28]. micro-electro-mechanical sensor (MEMS) attached to the foot of a person [28]. This This sensor has a triaxial accelerometer, a gyroscope, and a magnetometer. They implemented sensor has a triaxial accelerometer, a gyroscope, and a magnetometer. They implemented and and compared most relevant techniques for user movement such as step detection, stride estimation, etc. compared most relevant techniques for user movement such as step detection, stride estimation, etc. A single use useof of IMU-based reckoning has a characteristic error generation A single thethe IMU-based deaddead reckoning has a characteristic error generation pattern— pattern—accumulation of error. The dead-reckoned tracking is accurate in the beginning, however, accumulation of error. The dead-reckoned tracking is accurate in the beginning, however, the errors the errors from the truth accumulate over time because of thein noise the mobile sensors. Periodic from the truth accumulate over time because of the noise the in mobile sensors. Periodic error error calibrations are necessary to maintain quality estimation.Figure Figure11 illustrates illustrates the the error calibrations are necessary to maintain thethe quality of ofestimation. error generation characteristics of an IMU after appropriate calibration. generation characteristics of an IMU after appropriate calibration.

Figure 1. 1. Calibration Calibration effect. effect. Figure

The error generation pattern of the IMU-based location estimation is simply incremental. The The error generation pattern of the IMU-based location estimation is simply incremental. The blue blue line in Figure 1 shows this incremental error pattern. We can restrict the increment in error line in Figure 1 shows this incremental error pattern. We can restrict the increment in error accumulation

Sensors 2017, 17, 1952 Sensors 2017, 17, 1952

3 of 14 3 of 15

within a certainwithin limit by calibrating The red circle in The Figure showsinthe natural error ofthe a accumulation a certain limitthe byIMU. calibrating the IMU. red1 circle Figure 1 shows calibration method. By using an appropriate calibration method,calibration we can achieve an acceptable limit an of natural error of a calibration method. By using an appropriate method, we can achieve the estimation quality. acceptable limit of the estimation quality. A A popular popular calibration calibration method method uses uses aa Wi-Fi Wi-Fisystem system[29,30]. [29,30]. Affordable Affordable precision precision isis attainable attainable using a pervasive Wi-Fi system with meticulous signal calibration. However, Wi-Fi signal using a pervasive Wi-Fi system with meticulous signal calibration. However, Wi-Fi signal patterns patterns are are usually usually unstable. unstable. The The unstable unstable pattern pattern comes comes at at aa high high cost cost caused caused by by frequent frequent re-establishment re-establishment of of calibration calibration points. points. This This tradeoff tradeoff between between the the calibration calibration overhead overhead and and the the accuracy accuracy has has been been an an important deploying Wi-Fi calibration for indoor location estimation. If one could identify importantchallenge challengetoto deploying Wi-Fi calibration for indoor location estimation. If one could other appropriate means of means calibration, the IMU-based dead reckoning be sufficiently applicable, identify other appropriate of calibration, the IMU-based dead could reckoning could be sufficiently even in indoor environments. applicable, even in indoor environments. Activity Activity recognition recognition and and environment environment sensing sensing demonstrate demonstrate the the ability ability to to recognize recognize the the user user behavior behavior and and spatial spatial significance. significance. For instance, the IMU can detect user activities activities such as walking, walking, turning turning aa corner, corner, and and stepping stepping up up and anddown downstairs, stairs,whereas whereasmagnetometers magnetometers can can detect detect significant significant magnetic magnetic signals. signals. If If these these sensory sensory signatures signatures are are used used for for context context awareness, awareness, they they can can be be referred referred to to as as location location estimation estimation as as well. well. These These signatures signatures can can be be used used as as landmarks landmarks to to calibrate calibrate IMU-based IMU-based dead dead reckoning reckoning [31]. [31]. The The essential essential point point of of localizing localizing the the signatures signatures (i.e., (i.e., finding finding landmarks) landmarks) is is the the more moreaccurate accurate detection detectionof of sensory sensory signatures signatures that that have have concrete concrete uniqueness uniqueness around around the the neighboring neighboring spatial be abe landmark the sensory signaturesignature of a location be statistically differentiated. spatialfield. field.ToTo a landmark the sensory of should a location should be statistically In this article, we a structural designaofstructural indoor spaces andofa indoor statistical procedure identify differentiated. Inpropose this article, we propose design spaces and atostatistical significant indoor dead reckoning. analyze and understand extensive procedure landmarks to identifyfor significant landmarks for First, indoorwedead reckoning. First, wethe analyze and sensory signatures of natural landmarks. Then, we propose an autonomous landmark understand the extensive sensory signatures of natural landmarks. Then, we procedure propose anfor autonomous identification. autonomous procedureThe is especially reinforced by anisarbitrary butreinforced large number of procedure forThe landmark identification. autonomous procedure especially by an mobile devices. They identifyofthe landmarks repeatedly, whichthe canlandmarks enhance the quality of which landmark arbitrary but large number mobile devices. They identify repeatedly, can identification. The proposed statistical process of autonomous procedure localization enhance the quality of landmark identification. The proposed statisticalimproves process the of autonomous accuracy over time. the localization accuracy over time. procedure improves 2. 2. Materials Materials and and Methods Methods 2.1. 2.1. Structural Structural Landmark Landmark Identification Identification A user performs a predictable through a specific structure, such as aasflight of A user performs a predictablemotion motionwhen whenpassing passing through a specific structure, such a flight stairs, elevator, entrance, or escalator, in aninindoor space. The The predictable motion can be to a of stairs, elevator, entrance, or escalator, an indoor space. predictable motion canconverted be converted sensory signature by mobile devices: a multi-dimensional combination of sensory data postulates that to a sensory signature by mobile devices: a multi-dimensional combination of sensory data postulates signatures are likely emerge for specific For instance, fluctuations in the gyroscope that signatures are to likely to emerge for structures. specific structures. For the instance, the fluctuations in the and magnetometer measurements are mapped to their corresponding physical locations such as gyroscope and magnetometer measurements are mapped to their corresponding physical locations an escalator. These spatial sensory determinedetermine the “structural” landmarks.landmarks. To recognize such as an escalator. These spatialsignatures sensory signatures the “structural” To an elevatoran landmark, can monitor the fluctuation on every axis of magnetometer. recognize elevator we landmark, we can monitor thevariance fluctuation variance onthe every axis of the The specific magnetometer fluctuations arefluctuations observed for stopfor over an stop elevator magnetometer. The specific magnetometer are every observed every over movement an elevator (see Figure 2). movement (see Figure 2).

(a)

(b)

Figure2. 2. Accelerator/magnetometer Accelerator/magnetometer fluctuation to elevator: Figure fluctuation pattern pattern corresponding corresponding to elevator: (a) (a) Accelerometer, Accelerometer, (b)magnetometer. magnetometer. (b)

Sensors 2017, 17, 1952

4 of 14 Sensors 2017, 17, 1952 Sensors Sensors 2017, 2017, 17, 17, 1952 1952

of 15 444 of of 15 15

Similar to an elevator landmark, we can categorize other types of typical structural landmarks: Similar to an elevator landmark, we can categorize other types of typical structural landmarks: Similar to landmark, categorize of Similar to an an elevator elevator landmark, we can can categorize other other types of typical typical structural structural landmarks: corridors, corners, stairs, and escalators. Table 1 lists we the observation andtypes visualization of the landmarks: measured corridors, corners, stairs, and escalators. Table lists the observation and visualization of the corridors, corridors, corners, corners, stairs, stairs, and and escalators. escalators. Table Table 111 lists lists the the observation observation and and visualization visualization of of the the sensory signatures for measured each structural landmark. sensory signatures for each structural landmark. measured measured sensory sensory signatures signatures for for each each structural structural landmark. landmark.

Table 1. Determinants and visualization of sensory signature. Table 1. Determinants visualization of sensoryof Table 1. and sensory Tableand 1. Determinants Determinants and visualization visualization ofsignature. sensory signature. signature. Structural Landmark

Structural Structural Structural Observation Observation Observation Landmark Landmark LandmarkObservation

Visualization Visualization

Visualization Visualization

Typical Typical Typical

Corridor

vibration/fluctuation on vibration/fluctuation on vibration/fluctuation on Typical vibration/fluctuation on the Corridor Corridor Corridor the z-axis of the the the z-axis z-axis of of the the z-axis of the accelerometer accelerometer accelerometer accelerometer

Changes corresponding to Changes Changes corresponding corresponding to to

Corner

corner are measured by aaa corner measured corner are areto measured by Changes corresponding a cornerby Corner the yaw-axis (z-axis) of the Corner the yaw-axis (z-axis) of Corner by thethe yaw-axis (z-axis) of the the are measured yaw-axis (z-axis) gyroscope over the gyroscope the gyroscope over the of the gyroscope over the ±over 1.5 rad/s 1.5 rad/s ±±± 1.5 1.5 rad/s rad/s

A large fluctuation A A large large fluctuation fluctuation

variance in the z-axis of the variance in of A large fluctuation variance in the variance in the the z-axis z-axis of the the accelerometer accelerometer z-axis of the accelerometer accelerometer (approximately greater (approximately greater (approximately greater than twice the (approximately greater than twice the fluctuation than the fluctuation variance corresponding than twice twice the fluctuation fluctuation variance corresponding to corresponding to avariance corridor) variance corresponding to to corridor) aaa corridor) corridor)

Stair Stair Stair Stair

The change change in in the the direction direction The

The direction change in measured the direction The change in the measured by the yaw-axis measured measured by by the the yaw-axis yaw-axis by the yaw-axis of the gyroscope of the the gyroscope (Turning of the gyroscope gyroscope (Turning (Turning (Turningof into a corner is into aaa corner corner is observed into corner is is observed observed observedinto additionally) additionally) additionally) additionally)

Sensors 2017, 17, 1952

5 of 15

Escalator

duration (less than 20 s)

1 15 29 43 57 71 85 99 113 127 141 155 169 183 197 211 225 239 253 267 281 295 309 323 337 351 365 379 393 407 421 435 449 463 477 491 505 519 533

The on–off pattern of the

The on–off pattern of the fluctuation fluctuation variance on the variance on the z-axis of the Escalator z-axis of the accelerometer accelerometer over a relatively long over a relatively long duration (less than 20 s)

Acceleration of gravity (g)

Accelerometer 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1 -1.2 -1.4

x-axis

y-axis

z-axis

To confirm each sensory signature as a valid structural landmark, we propose a validity test

To confirm each sensory a valid criteria. structural we propose a validity test using using thesignature landmark as confirming The landmark, landmark confirming criteria listed in Table 2 are obtained by conducting extensive empirical tests in indoor spaces. We develop a custom sensory the landmark confirming criteria. The landmark confirming criteria listed in Table 2 are obtained signature collecting application for three types of mobile devices: a Samsung Galaxy S7 (Samsung Electronics, Seoul, Korea), a Sony Xperia XZ (Sony Corporation, Tokyo, Japan), and an LG G5 (LG Electronics, Seoul, Korea). The raw sensor data measured by the embodied IMU and magnetometer are transformed into sensory signature values for each measurement time interval (usually 100 ms). The gathered sensory signatures are transferred to a Microsoft Excel program for visualization after the whole signature gathering process. The tested indoor space is the second floor of a college building that includes 15 corridors, 10 corners, three sets of stairs, two escalators, and three elevators.

Sensors 2017, 17, 1952

5 of 14

by conducting extensive empirical tests in indoor spaces. We develop a custom sensory signature collecting application for three types of mobile devices: a Samsung Galaxy S7 (Samsung Electronics, Seoul, Korea), a Sony Xperia XZ (Sony Corporation, Tokyo, Japan), and an LG G5 (LG Electronics, Seoul, Korea). The raw sensor data measured by the embodied IMU and magnetometer are transformed into sensory signature values for each measurement time interval (usually 100 ms). The gathered sensory signatures are transferred to a Microsoft Excel program for visualization after the whole signature gathering process. The tested indoor space is the second floor of a college building that includes 15 corridors, 10 corners, three sets of stairs, two escalators, and three elevators. We collect sensory signatures at least 20 times for each structure. An experimenter holds mobile devices with a handheld pose (generally, when a person looks for a way using navigation application in an indoor space, he or she holds a mobile device in hands and looks at the screen with his or her eyes). By using the statistical analytics (i.e., one-way ANOVA), we define the key criteria to identify the valid structural landmarks. The key criteria (e.g., z-axis accelerometer for corridors or stairs) are identified within a significance level of 5%. If an observed average sensory signature for the key √ criteria is within the confidence interval (i.e., average ± z0.025 × variance, z0.025 = 1.96), the observed signature can be considered a structural landmark under a statistical confidence of 95%. Table 2. Central values for landmark confirming criteria. Signature Stationary Corridor Corner Stair Escalator Elevator

Key Criteria (m/s2 )

z-axis accelerometer z-axis accelerometer z-axis gyroscope (rad/s) z-axis accelerometer z-axis gyroscope z-axis magnetometer (µT)

Mean

Maximum

Minimum

Variance

8.973 9.1099 0.0151 8.7039 0.0298 −10.5621

9.32 14.9578 2.0107 19.6085 3.0201 −5.1028

8.707 5.9023 −1.603 2.6965 −2.703 −21.8373

0.009 2.3549 0.1741 4.0408 0.2132 7.721

Notes: (1) A sensory signature for stationary state is included for comparison. (2) The z-axis accelerometer value is generally less than 1 G (9.8 m/s2 ) because an experimenter holds mobile devices with a slightly tilted handheld pose. (3) The significant change of gyroscope value is caused by a turning movement. To prevent confusion caused by changing direction (left turn or right turn), we use absolute gyroscope measurement values.

Structural landmarks provide effective calibration points for indoor location estimation. However, the number of structural landmarks is often not sufficient to sustain the desirable accuracy of location estimation. The increasing density of landmarks will reduce the location estimation error by frequent calibration. Thus, we focus on another type of landmark—spontaneous landmarks. 2.2. Spontaneous Landmark Identification Indoor spaces offer ambient sensory signatures across multiple sensors. For instance, the metallic objects in a certain location may generate unique and reproducible patterns on a magnetometer. These ambient sensory signatures generally may not be recognized a priori. A different indoor space will have a different distribution of signatures. Although there is no theoretical proof, empirical observations indicate that an indoor space has sufficient ambient signatures from sound, light, magnetic field, temperature, and radio signals. An effective method to identify the ambient signatures is definitely useful for frequent calibrations. The identified ambient signatures can be used to determine the “spontaneous” landmarks that supplement the limited amount of structural landmarks. Recognizing distinct sensory signatures is essential for discovering spontaneous landmarks. Sensory signatures are characterized by the mean, max, min, and variance similar to structural landmarks. Our objective is to identify distinct signatures that have less similarity to the signatures of neighboring locations. Distinct signatures are generally identified using a sensory signature gathering process. A sensory signature gathering application installed in smart devices collects sensory signatures continuously according to the walking path of a pedestrian. A similarity test is applied to two time-consecutively collected sensory signatures: one signature at time interval t and the other at t − 1.

Sensors 2017, 17, 1952

6 of 14

If the correlation between the two consecutive signatures is greater than a similarity threshold, we consider the signature at time interval t as statistically distinct, and the location of the distinct signature is considered as the candidate for a spontaneous landmark. The sensory signatures measured at time interval t are described using a vector form that has three tuples. The first tuple comprises the signatures of an accelerometer as expressed by (1): t f accelerometer = ( atx−mean , atx−min , atx−max , atx−variance , aty−mean , atx−min , aty−max , aty−variance ,

atz−mean , atz−min , atz−max , atz−variance )

(1)

where atx−mean denotes the mean value of the x-axis accelerometer at time interval t and the other three elements of the x-axis accelerometer denote the min, max, and variance. Similarly, the elements for t y-axis and z-axis accelerometers are contained in the first tuple, f accelerometer . The tuples for gyroscope and magnetometer are defined as shown in (2) and (3): t f gryoscope = ( gtx−mean , gtx−min , gtx−max , gtx−variance , gyt −mean , gtx−min , gyt −max , gyt −variance ,

gzt −mean , gzt −min , gzt −max , gzt −variance ) t f magnetometer = (mtx−mean , mtx−min , mtx−max , mtx−variance , mty−mean , mtx−min , mty−max , mty−variance ,

mtz−mean , mtz−min , mtz−max , mtz−variance )

(2)

(3)

t t t A typical sensory signature is a combination of three tuples: st = { f accelerometer , f gryoscope , f magnetometer }. The proposed similarity test is conducted with two sensory signatures at consecutive time intervals. Each time interval holds separated portion on the time axis. The two consecutive time intervals are not overlapped. The square of the Euclidean distance (i.e., d2 (st−1 , st ) given by (4)) of the two time-consecutive signatures (st−1 , st ) presents the statistical difference of the two sensory signatures: 2

d2 (st−1 , st ) = (st−1 − st )

(4)

Each element of a sensory signature has a normally distributed measurement error. The central limit theorem [32] guarantees the normal distribution of relatively large number (usually, more than 20) of independent sensor measurements. Thus, each element of sensory signature follows a normal distribution. Then, the vector difference of two signatures (st−1 − st ) also follows a normal distribution. Each element of vector st−1 − st follows a normal distribution. Then, d2 (st−1 , st ), the square summation of elements of vector st−1 − st follows chi-squared distribution [33] with degrees of freedom m, where m is the number of independent elements of a vector (i.e., d2 (st−1 , st ) ∼ χ2 (m), mean of χ2 (m) is m and variance is 2m). The chi-square distribution with m degrees of freedom is the distribution of a sum of p the squares of m independent normal random variables. By the statistical finding of √ Fisher [34], 2d2 (st−1 , st ) follows approximate normal distribution with mean 2m − 1 and a unit variance. The significant difference p between two sensory signatures can be determined by its statistical characteristics. When a value of 2d2 (st−1 , st ) is over the given threshold, the two sensory signatures, st−1 and st , are considered significantly different. Note that, the time interval is usually 100 ms. Because of the highly sensing frequency of embodied IMU and magnetometer (100 Hz–400 Hz for usual mobile devices), 100 ms time interval provides 10–40 sensing opportunities for single time interval. It is the sufficient number to obtain valid sensory signatures for every time intervals. For obtaining the spontaneous landmarks, we subject the sensory signatures to the continuous process of a similarity test. Once the similarity test has approved, each of the resulting distinct sensory signatures is expected to contain a unique location. We detect the unique sensory signatures and record their ground truths to localize the spontaneous landmarks. We build an entire map of landmarks by the combining both the structural and spontaneous landmarks. Figure 3 shows an example of landmarks. The illustrated floor plan is obtained from a commercial departmental store in Incheon, Korea. We detected 11 structural landmarks and 10 spontaneous ones.

process of a similarity test. Once the similarity test has approved, each of the resulting distinct sensory signatures is expected to contain a unique location. We detect the unique sensory signatures and record their ground truths to localize the spontaneous landmarks. We build an entire map of landmarks by the combining both the structural and spontaneous landmarks. Figure 3 shows an example of1952 landmarks. The illustrated floor plan is obtained from a commercial departmental store Sensors 2017, 17, 7 of 14 in Incheon, Korea. We detected 11 structural landmarks and 10 spontaneous ones.

Figure3.3.Example Exampleof oflandmarks landmarks detection. detection. Figure

The ground truths that are used to localize the spontaneous landmarks are collected by The ground truths that are used to localize the spontaneous landmarks are collected by voluntary voluntary activities. A simple sensory signature gathering application generates an alarm whenever activities. simplesignatures sensory signature gathering generates an alarm whenever distinct distinct A sensory are detected. A user application can record the geographical ground truths for the sensory signatures are detected. A user can record the geographical ground truths for the detected detected signatures by pinpointing over the floor plan. In addition, the ground truth can be estimated signatures by pinpointing overof thea floor plan. In addition, the ground truth The can be estimated by the by the autonomous actions sensory signature gathering application. sensory signature autonomous actions of aautomatically sensory signature gathering application. The sensory signature gathering gathering application records the current location obtained by dead reckoning for a application automatically records the current location obtainedlarge by dead a detected detected signature. The autonomous recording has a relatively errorreckoning compared for to voluntary signature. The autonomous recording has a relatively error compared to voluntary pinpointing. For the estimation of a location by deadlarge reckoning, repetitive collection andpinpointing. averaging For the estimation of a location by dead reckoning, repetitive collection and averaging can reduce8 the can reduce the error of the ground truth estimation. Figure 4 illustrates the entire process of landmark Sensors 2017, 17, 1952 of 15 error of the ground truth estimation. Figure 4 illustrates the entire process of landmark identification and location estimation. The identified landmarks are continuously usedare for continuously recalibration. Periodic identification and location estimation. The identified landmarks used for recalibration achieverecalibration an acceptable bound of locationupper errors.bound of location errors. recalibration.can Periodic canupper achieve an acceptable

Figure Figure4.4.Process Processofoflandmark landmarkidentification identificationand andlocation locationestimation. estimation.

The floor plan with the entire landmark information is downloaded when a user reaches an indoor space. We can dead reckon our indoor locations with frequent recalibration (see Figure 5). The recalibration is activated by both the structural and spontaneous landmarks.

Sensors 2017, 17, 1952

8 of 14

Figure 4. Process of landmark identification and location estimation.

The is downloaded downloaded when when aa user user reaches reachesan an Thefloor floorplan planwith with the the entire entire landmark landmark information information is indoor space. We can dead reckon our indoor locations with frequent recalibration (see Figure 5). indoor space. We can dead reckon our indoor locations with frequent recalibration (see Figure 5). The The recalibration is activated by both the structural spontaneous landmarks. recalibration is activated by both the structural and and spontaneous landmarks.

Figure 5. User location estimation by continuous calibration. Figure 5. User location estimation by continuous calibration.

3. Results 3. Results To demonstrate the practical usability of the proposed landmark calibration method, we To demonstrate the practical usability landmarks of the proposed landmark method, we identified identified the structural and spontaneous for two differentcalibration types of buildings. A large area the structural and spontaneous landmarks for two different types of buildings. A large area cannot cannot be appropriate to test and enhance the applicability of the landmark identifying process. Thus, be to test and enhance the test applicability of the landmark identifying process. Thus, we weappropriate first apply the developed similarity in a relatively small area. The test area is the second floor first apply the developed similarity test in a relatively small area. The test area is the second floor of of a college building. Figure 6 illustrates the details of the floor. The bottom-left point is the origin ofa college building. Figure 6 illustrates the details of the floor. The bottom-left point is the origin ofthe the the coordinates. Sensors 2017, 17, 1952 The distances from the origin are represented in centimeters. To identify9 of 15 coordinates. The distances from the origin are represented in centimeters. To identify the landmarks, landmarks, we repeatedly wander around the test floor. After the determination of landmarks, we we repeatedly wanderestimation around the test for floor. After the paths. determination of landmarks, we measure the measure the location error total seven Each path has ten location measuring location estimation error for total seven paths. Each path has ten location measuring points. points.

Figure 6. Test floor of small area. Figure 6. Test floor of small area.

The similarity test processes on two time-consecutive sensory signatures (one signature is The similarity time-consecutive sensory signatures signature is measured at time t test andprocesses the otheronis two measured at t  1 ). The difference of two(one signatures (i.e., measured at time t and the other is measured at t − 1). The difference of two signatures 1 t √ unit variance (i.e.,   1 2d 2p (st2d ,2s(s)t−in1 ,Section  1 and 2.2) is normally distributed with meanwith2m (i.e., st ) in Section 2.2) is normally distributed mean 2m − 1 and unit variance ). Then, we apply 2m 1  k as a useful threshold value. The constant value, k , determines the strictness of the similarity test. Suppose that k is equal to 1, the two sensory signatures are determined to be statistically different only when they have a difference belonging to the upper 15.8% of the difference distribution. In the cases of k  1.5 and k  2 , we consider the upper 6.7% and 2.2% of the total difference distribution as the similarity threshold, respectively. In this experiment, we

Figure 6. Test floor of small area.

The similarity test processes on two time-consecutive sensory signatures (one signature is measured at 1952 time t and the other is measured at t  1 ). The difference of two signatures9 (i.e., Sensors 2017, 17, of 14

2d 2 (st 1, st ) in Section 2.2) is normally distributed with mean

2m  1 and unit variance (i.e.,   1 √ (i.e., σ = we 1). Then, − 1a + k as athreshold useful threshold The constant k, determines m 1  2m k as ). Then, applywe 2apply useful value. value. The constant value,value, the k , determines the strictness of the similarity test. Suppose that k is equal to 1, the two sensory signatures strictness of the similarity test. Suppose that k is equal to 1, the two sensory signatures are are determined to be statistically different only when they have a difference belonging to the upper 15.8% determined to be statistically different only when they have a difference belonging to the upper 15.8% of the difference differencedistribution. distribution.InIn cases 2, we consider the upper and of the thethe cases of of andand consider the upper 6.7%6.7% and 2.2% k k 1= .5 1.5 k k 2=, we 2.2% of the total difference distribution as the similarity threshold, respectively. In this experiment, of the total difference distribution as the similarity threshold, respectively. In this experiment, we we indicate results corresponding 1.5,and and2 2totodemonstrate demonstratethe theusefulness usefulnessof of the the similarity similarity indicate thethe results corresponding to tok k= 1,1,1.5, test. A mobile application tool is developed for the test. The application obtains the landmarks, setssets up test. A mobile application tool is developed for the test. The application obtains the landmarks, auptest path, andand performs dead reckoning withwith calibration as shown in Figure 7. 7. a test path, performs dead reckoning calibration as shown in Figure

Figure 7. Test tool. Figure 7. Test tool.

Figure 8 illustrates the difference corresponding to different k values: k  1 has the largest Figure 8 illustrates the difference corresponding k values: k = 1 has thelandmarks. largest number number of spontaneous landmarks and thedifferent least number of spontaneous The k  2 has to of spontaneous landmarks and k = 2 has the least number of spontaneous landmarks. The actual actual footprints of users are illustrated using the green dots. The black lines show the average traces footprints ofexperiments users are illustrated using for multiple per single path.the green dots. The black lines show the average traces for multiple experiments per single path. Sensors 2017, 17, 1952

10 of 15

Figure 8. Traces k value: 1, (b)kk =11.5, Figure 8. Traces for kfor value: (a) (a) (c)k = k k 1=, (b) . 5 ,(c) k 2.  2.

The location errors for each k value are listed in Table 3. The number of spontaneous landmarks increases with a decrease in k values, whereas the number of structural landmarks is constant. The location estimation errors are not minimized when the number of landmarks is the largest. However, the location errors are minimized at the median of k.

Sensors 2017, 17, 1952

10 of 14

The location errors for each k value are listed in Table 3. The number of spontaneous landmarks increases with a decrease in k values, whereas the number of structural landmarks is constant. The location estimation errors are not minimized when the number of landmarks is the largest. However, the location errors are minimized at the median of k. Table 3. Error distribution according to landmark density (unit: cm). k=2 Error Path1 Path2 Path3 Path4 Path5 Path6 Path7

k = 1.5

Number of Landmarks Structural

Spontaneous

3 3 4 4 5 5 7

5 4 6 7 7 8 10

127.69 126.66 131.16 84.29 78.47 127.69 86.95

Error 80.33 94.71 97.94 64.38 71.15 116.28 59.17

k=1

Number of Landmarks Structural

Spontaneous

3 3 4 4 5 5 7

7 7 7 8 9 10 13

Number of Landmarks

Error

Structural

Spontaneous

3 3 4 4 5 5 7

10 9 9 11 13 14 16

127.24 106.93 111.71 98.05 84.04 129.73 132.32

Figure 9 illustrates the patterns of location errors for each experimental path. Location errors are minimized at the median of k. We expand the k value into {1, 1.2, 1.4, 1.6, 1.8, 2} to enhance the visibility of the trends. If k = 1.2, we consider the upper 11.5% of the total difference distribution as the similarity threshold. Similarly, upper 8.1% when k = 1.4, 5.5% when k = 1.6, and 3.6% when k = 1.8 are considered to similarity thresholds. Next, we extended the landmark calibration method to a commercial department store located in Incheon, Korea. The floor area of the target floor is 9428 m2 . Three escalators and four elevators are included in the target floor. In addition, the corners are included over the seventy places in the target floor Sensors(see 2017,Figure 17, 19523). 11 of 15

140.00 120.00

Path 1

100.00

Path 2

80.00

Path 3

60.00

Path 4 Path 5

40.00

Path 6

20.00

Path 7

0.00 k =2

k=1.8

k=1.6

k=1.4

k=1.2

k =1

Figure 9. Error charts for k values (unit of y-axis: cm). Figure 9. Error charts for k values (unit of y-axis: cm).

We listed the results of three representative similarity thresholds: k  1, 1.5, and 2. Table 4 lists We listed resultsand of three representative similarity thresholds: = 1, 1.5, and 2. Table 4 lists the number of the structural spontaneous landmarks, recalibration, andk errors of location estimation the number of structural and spontaneous landmarks, recalibration, and errors of location estimation for ten paths. for ten paths. Figure 10 Table illustrates thedistribution number ofaccording recalibrations for eachfor experimental path. 4. Error to recalibrations large area (unit: cm).The number of recalibrations increases with decreasing k values. Error Path1 Path2 Path3 Path4 Path5 Path6 Path7

121.31 98.24 135.12 75.87 96.19 119.56 131.29

k=2 Number of Recalibration 11 15 16 18 18 20 20

Error 106.31 86.24 119.12 62.87 82.19 102.56 112.29

k = 1.5 Number of Recalibration 14 19 21 20 21 22 24

Error 114.31 94.24 130.12 89.87 94.19 118.56 140.29

k=1 Number of Recalibration 16 21 22 22 22 24 26

Figure 9. Error charts for k values (unit of y-axis: cm).

We listed the results of three representative similarity thresholds: k  1, 1.5, and 2. Table 4 lists the number of structural and spontaneous landmarks, recalibration, and errors of location estimation for ten paths.

Sensors 2017, 17, 1952

11 of 14

Table 4. Error distribution according to recalibrations for large area (unit: cm). Table 4. Error for large area (unit: cm). k = 2 distribution according to recalibrations k = 1.5 k=1 Number of

Error Path1 Path2 Path3 Path1 Path4 Path2 Path5 Path3 Path6 Path7 Path4 Path8 Path5 Path9 Path6 Path10 Path7

Error

k =Recalibration 2

121.31 Error 98.24 135.12 121.31 75.87 98.24 96.19 135.12 119.56 131.29 75.87 67.31 96.19 71.49 119.56 114.35 131.29

11

Number of 15 Recalibration 16

Number of

k = 1.5 Recalibration

106.31

Error 86.24 119.12

14

Number of 19 Recalibration 21

Error

Number of

k = 1 Recalibration

114.31

16

Number of 21 Recalibration

Error 94.24 130.12

22

11 106.31 1420 114.31 1622 18 62.87 89.87 15 86.24 1921 94.24 2122 18 82.19 94.19 16 119.12 2122 130.12 2224 20 102.56 118.56 20 112.29 140.29 18 62.87 2024 89.87 2226 23 48.31 84.31 18 82.19 2127 94.19 2229 22 53.49 57.49 20 102.56 2228 118.56 2431 25 102.35 105.35 20 112.29 2429 140.29 2634 Path8 67.31 23 48.31 27 84.31 29 Path9 53.49 28each experimental 57.49 Figure 10 71.49 illustrates the22number of recalibrations for path. The31number of Path10 114.35 25 102.35 29 105.35 34

recalibrations increases with decreasing k values. 40

Path 1

35

Path 2

30

Path 3

25

Path 4

20

Path 5

15

Path 6

10

Path 7

5

Path 8

0

Path 9 k =2

k =1.8 k =1.6 k =1.4 k =1.2

k =1

Figure 10. Number of recalibrations for k values in large test area (unit of y-axis: recalibration Figure 10. Number of recalibrations for k values in large test area (unit of y-axis: recalibration frequency). frequency).

Sensors 2017, 17, 1952

12 of 15

However, does not not guarantee the quality of location estimation. The location However,frequent frequentrecalibration recalibration does guarantee the quality of location estimation. The errors areerrors minimized under the median recalibrations (see Figure We 11). alsoWe expand the k value location are minimized under the of median of recalibrations (see11). Figure also expand the into {1, 1.2, 1.4, 1.6, 1.8, 2} to indicate a more explicit trend of recalibration. k value into {1, 1.2, 1.4, 1.6, 1.8, 2} to indicate a more explicit trend of recalibration.

160

Path 1

140

Path 2

120 100

Path 3

80

Path 4

60

Path 5

40

Path 6

20

Path 7

0 k =2

k =1.8

k =1.6

k =1.4

k =1.2

k =1

Path 8

Figure 11. Error charts for k values in large test area (unit of y-axis: cm). Figure 11. Error charts for k values in large test area (unit of y-axis: cm).

Figure 11 illustrates the patterns of location errors for each experimental path. The median of k minimizes the location errors. The results of the calibration method using a Wi-Fi system [29,30] are illustrated in Figure 12 to show the excellence of the proposed method. The number of Wi-Fi recalibrations are slightly different for each test path: the maximum number of recalibration is 27 times for path 10 and the minimum is 11 times for path 1. Figure 12 shows the performance

40

Path 6

20

Path 7

0 k =2

k =1.8

k =1.6

k =1.4

k =1.2

Path 8

k =1

Sensors 2017, 17, 1952

12 of 14

Figure 11. Error charts for k values in large test area (unit of y-axis: cm).

Figure 11 11 illustrates illustrates the the patterns patternsof oflocation locationerrors errorsfor foreach eachexperimental experimentalpath. path.The Themedian medianofofk Figure k minimizes the location errors. The results of the calibration method using a Wi-Fi system [29,30] minimizes the location errors. The results of the calibration method using a Wi-Fi system [29,30] are are illustrated in Figure show excellenceofofthe theproposed proposedmethod. method. The Thenumber number of of Wi-Fi Wi-Fi illustrated in Figure 12 12 to to show thethe excellence recalibrations are each testtest path: the the maximum number of recalibration is 27 times recalibrations areslightly slightlydifferent differentforfor each path: maximum number of recalibration is 27 for path 10 and the minimum is 11 times for path 1. Figure 12 shows the performance comparisons of times for path 10 and the minimum is 11 times for path 1. Figure 12 shows the performance Wi-Fi calibration and the proposed autonomous landmark calibrationlandmark (k = 1.4 and k = 1.6). The comparisons of Wi-Fi calibration and the proposed autonomous calibration (k =proposed 1.4 and method illustrates the superior results for average 17.8% enhancement. k = 1.6). The proposed method illustrates the superior results for average 17.8% enhancement. 160 140 120 100 80 60 40 20 0 Path1

Path2

Path3

Path4 k=1.6

Path5 k=1.4

Path6

Path7

Path8

Path9

Path10

Wi-Fi Calibration

Figure 12. Comparison with Wi-Fi calibration (unit of y-axis: cm). Figure 12. Comparison with Wi-Fi calibration (unit of y-axis: cm).

To concretely determine a landmark in the extended department store experiment, we apply To concretely determine a landmark in the extended department store signature experiment, collective intelligence by voluntary activities. Repeated detections of same on we the apply same collective intelligence by voluntary activities. Repeated detections of same signature on the same location is needed to be a landmark. In addition, when we detect a landmark signature during location is needed to be a landmark. In addition, when we detect a landmark signature during navigation, we check the stored ground truth of the landmark. If the ground truth of the detected navigation, thecurrent stored ground the landmark. If the ground truth landmark of the detected landmark is we far check from the locationtruth of theofdevices, we determine the detected as an landmark is far from the current location of the devices, we determine the detected landmark as an improper one. Then, we inactivate the calibration process. improper one. Then, we inactivate the calibration process. 4. Discussion 4. Discussion The proliferation of mobile communication and smart devices has drawn significant attention to Location-Based Services. Accurate location estimation is one of the most essential components for the usability of LBSs. Especially in an indoor environment, the traditional location estimation methods such as pure dead reckoning have the limitation of error accumulation over time. Moreover, solution providers do not share significant technical advances publicly. The dead reckoning with frequent recalibration is a beneficial technique for expanding the LBS capability. The landmark points, structural or spontaneous, activate recalibration. The corner points, stairs, escalators, and elevators are representative structural landmarks. The usefulness of structural landmarks can stimulate the expansion of spontaneous landmarks. A statistical similarity test identifies various spontaneous landmarks. The sensory signatures of geographical locations are represented using a vector form. The difference between the two sensory signature vectors can be evaluated by conducting a similarity test. This autonomous landmark finding procedure is used to reduce cost while keeping a affordable estimation quality. The theoretical achievement of statistical analysis strengthens the advantage of LBS application especially for indoor location estimation. In addition, the proposed landmark calibration can be tightly coupled to traditional Wi-Fi calibration. The well-trained Wi-Fi calibration points can be

Sensors 2017, 17, 1952

13 of 14

added as important landmarks in our proposed method. All the sensory signatures were acquired through the actual measurements, which validates the practicality of the proposed method. Acknowledgments: This work was supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No. 2016-0-00160, Versatile Network System Architecture for Multi-Dimensional Diversity). Author Contributions: Jae-Hoon Kim and Byoung-Seop Kim developed the proposed method and designed the experiments; Byoung-Seop Kim performed the experiments; Jae-Hoon Kim analyzed the experimental results and wrote the article. Conflicts of Interest: The authors declare no conflict of interest.

References 1.

2.

3.

4. 5. 6.

7. 8.

9.

10.

11. 12.

13. 14.

15.

Lu, G.; Yan, Y.; Kolagunda, A.; Kambhamettu, C. A Fast 3D Indoor-Localization Approach Based on Video Queries. In Proceedings of the International Conference on Multimedia Modeling, Miami, FL, USA, 4–6 January 2016; pp. 218–230. Lu, G.; Yan, Y.; Sebe, N.; Kambhamettu, C. Knowing Where I Am: Exploiting Multi-Task Learning for Multi-View Indoor Image-Based Localization. In Proceedings of the British Machine Vision Conference, Nottingham, UK, 1–5 September 2014; pp. 1–12. Liu, H.; Gan, Y.; Yang, J.; Sidhom, S.; Wang, Y.; Chen, Y.; Ye, F. Push the Limit of WiFi Based Localization for Smartphone. In Proceedings of the 18th Annual International Conference on Mobile Computing and Networking (MobiCom’12), Istanbul, Turkey, 22–26 August 2012; pp. 305–316. Liu, H.; Yang, J.; Sidhom, S.; Wang, Y.; Chen, Y.; Ye, F. Accurate WiFi Based Localization for Smartphones Using Peer Assistance. IEEE Trans. Mob. Comput. 2014, 13, 2199–2214. [CrossRef] Shu, Y.; Huang, Y.; Zhang, J.; Coué, P.; Cheng, P.; Chen, J.; Shin, K.G. Gradient-Based Fingerprinting for Indoor Localization and Tracking. IEEE Trans. Ind. Electron. 2016, 63, 2424–2433. [CrossRef] Seco, E.; Jimenez, A.R.; Prieto, C.; Roa, J.; Koutsou, K. A survey of mathematical methods for indoor localization. In Proceedings of the IEEE International Symposium on Intelligent Signal Processing, Budapest, Hungary, 26–28 August 2009; pp. 9–14. Roos, T.; Myllymäki, P.; Tirri, H.; Misikangas, P.; Sievänen, J. A probabilistic approach to WLAN user location estimation. Int. J. Wirel. Inf. Netw. 2002, 9, 155–164. [CrossRef] Youssef, M.; Agrawala, A. The Horus WLAN location determination system. In Proceedings of the 3rd International Conference on Mobile System Application Services, Seattle, WA, USA, 6–8 June 2005; pp. 205–218. Müller, P.; Raitoharju, M.; Piché, R. A field test of parametric WLAN-fingerprint-positioning methods. In Proceedings of the 17th International Conference on Information Fusion, Salamanca, Spain, 7–10 July 2014; pp. 1–8. Koski, L.; Piché, R.; Kaseva, V.; Ali-Löytty, S.; Hännikäinen, M. Positioning with coverage area estimates generated from location fingerprints. In Proceedings of the 7th Workshop on Positioning Navigation Communication, Dresden, Germany, 11–12 March 2010; pp. 99–106. Feng., Y.; Minghua, J.; Jing, L.; Xiao, Q.; Ming, H.; Tao, P.; Xinrong, H. An improved indoor localization of WiFi based on support vector machines. Int. J. Future Gener. Commun. Netw. 2014, 7, 191–206. Sánchez-Rodríguez, D.; Hernández-Morera, P.; Quinteiro, J.M.; Alonso-González, I. A low complexity system based on multiple weighted decision trees for indoor localization. Sensors 2015, 15, 14809–14829. [CrossRef] [PubMed] Haverinen, J.; Kemppainen, A. Global indoor self-localization based on the ambient magnetic field. Robot. Auton. Syst. 2009, 57, 1028–1035. [CrossRef] Berkovich, G. Accurate and reliable real-time indoor positioning on commercial smartphone. In Proceedings of the International Conference on Indoor Positioning and Indoor Navigation, Busan, Korea, 27–30 October 2014; pp. 670–677. Gozick, B.; Subbu, K.P.; Dantu, R.; Maeshiro, T. Magnetic maps for indoor navigation. IEEE Trans. Instrum. Meas. 2011, 60, 3883–3891. [CrossRef]

Sensors 2017, 17, 1952

16.

17. 18.

19.

20. 21.

22.

23. 24. 25.

26. 27. 28.

29.

30.

31.

32. 33. 34.

14 of 14

Abadi, M.J.; Luceri, L.; Hassan, M.; Chou, C.T.; Nicoli, M. A collaborative approach to heading estimation for smartphone-based PDR indoor localization. In Proceedings of the International Conference on Indoor Positioning and Indoor Navigation, Busan, Korea, 27–30 October 2014; pp. 554–563. Shu, Y.; Bo, C.; Shen, G.; Zhao, C.; Li, L.; Zhao, F. Magicol: Indoor localization using pervasive magnetic field and opportunistic WiFi sensing. IEEE J. Sel. Areas Commun. 2015, 33, 1443–1457. [CrossRef] Torres-Sospedra, J.; Montoliu, R.; Mendoza-Sinva, G.M.; Belmonte, O.; Rambla, D.; Huerata, J. Providing databases for different indoor positioning technologies: Pros and cons of magnetic field and Wi-Fi based positioning. Mob. Inf. Syst. 2016, 2016, 1–22. [CrossRef] Wang, X.; Zhang, C.; Liu, F.; Dong, Y.; Xu, X. Exponentially Weighted Particle Filter for Simultaneous Localization and Mapping Based on Magnetic Field Measurements. IEEE Trans. Instrum. Meas. 2017, 66, 1658–1667. [CrossRef] Madson, R.; Rajamani, R. Adaptive Dipole Model Based Disturbance Compensation in Nonlinear Magnetic Position Systems. IEEE/ASME Trans. Mechatron. 2017, 22, 794–803. [CrossRef] Renaudin, V.; Combettes, C.; Peyret, F. Quaternion based heading estimation with handheld MEMS in indoor environments. In Proceedings of the IEEE/ION on Position Location Navigation Symposium, Monterey, CA, USA, 5–8 May 2014; pp. 645–656. Combettes, C.; Renaudin, V. Comparison of misalignment estimation techniques between handheld device and walking directions. In Proceedings of the International Conference on Indoor Positioning and Indoor Navigation, Banff, AB, Canada, 13–16 October 2015; pp. 1–8. Harle, R. A survey of indoor inertial positioning systems for pedestrians. IEEE Commun. Surv. Tutor. 2013, 15, 1281–1293. [CrossRef] Gu, Y.; Lo, A.; Niemegeers, I. A Survey of Indoor Positioning Systems for Wireless Personal Networks. IEEE Commun. Surv. Tutor. 2009, 11, 13–32. [CrossRef] Turetzky, G. Indoor Positioning with SiRFusion for Next Generation E-911. Available online: https: //transition.fcc.gov/bureaus/pshs/docs/summits/911%20Location%20Acuracy/Greg_Turetzky.pdf (accessed on 24 August 2017). Fuchs, C.; Aschenbruck, N.; Martini, P.; Wieneke, M. Indoor tracking for mission critical scenarios: A survey. Pervasive Mob. Comput. 2011, 7, 1–15. [CrossRef] Beauregard, S.; Haas, H. Pedestrian Dead Reckoning: A Basis for Personal Positioning. In Proceedings of the 3rd workshop on positioning, navigation and communication, Hannover, Germany, 16 March 2006. Jim’enez, A.R.; Seco, F.; Prieto, C.; Guevara, J. A Comparison of Pedestrian Dead-Reckoning Algorithms Using a Low-Cost MEMS IMU. In Proceedings of the IEEE International Symposium on Intelligent Signal Processing, Budapest, Hungary, 26–28 August 2009. Wang, H.; Lenz, H.; Szabo, A.; Bamberger, J.; Hanebeck, U.D. WLAN-based pedestrian tracking using particle filters and low-cost MEMS sensors. In Proceedings of the Workshop on Positioning, Navigation and Communication, Hannover, Germany, 22 March 2007; pp. 1–7. Korbinian, F. Development and evaluation of a combined WLAN & inertial indoor pedestrian positioning system. In Proceedings of the 4th International Symposium on Location and Context Awareness, Savannah, GA, USA, 22–25 September 2009. Wang, H.; Sen, S.; Elgohary, A.; Farid, M.; Youssef, M.; Choudhury, R.R. Unsupervised Indoor Localization. In Proceedings of the International Conference on Mobile Systems, Applications, and Services, Low Wood Bay, UK, 25–29 June 2012; pp. 499–500. Wikipedia. Central Limit Theorem. Available online: https://en.wikipedia.org/wiki/Central_limit_theorem (accessed on 21 July 2017). Wikipedia. Chi-Square Distribution. Available online: https://en.wikipedia.org/wiki/Chi-squared_ distribution (accessed on 21 July 2017). Fisher, R.A. Statistical Methods for Research Workers; Springer: New York, NY, USA, 1954. © 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).