Activity Sequence-Based Indoor Pedestrian Localization Using ...

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS

1

Activity Sequence-Based Indoor Pedestrian Localization Using Smartphones Baoding Zhou, Student Member, IEEE, Qingquan Li, Qingzhou Mao, Member, IEEE, Wei Tu, Member, IEEE, and Xing Zhang

Abstract—This paper presents an activity sequence-based indoor pedestrian localization approach using smartphones. The activity sequence consists of several continuous activities during the walking process, such as turning at a corner, taking the elevator, taking the escalator, and walking stairs. These activities take place when a user walks at some special points in the building, like corners, elevators, escalators, and stairs. The special points form an indoor road network. In our approach, we first detect the user’s activities using the built-in sensors in a smartphone. The detected activities constitute the activity sequence. Meanwhile, the user’s trajectory is reckoned by Pedestrian Dead Reckoning (PDR). Based on the detected activity sequence and reckoned trajectory, we realize pedestrian localization by matching them to the indoor road network using a Hidden Markov Model. After encountering several special points, the location of the user would converge on the true one. We evaluate our proposed approach using smartphones in two buildings: an office building and a shopping mall. The results show that the proposed approach can realize autonomous pedestrian localization even without knowing the initial point in the environments. The mean offline localization error is about 1.3 m. The results also demonstrate that the proposed approach is robust to activity detection error and PDR estimation error. Index Terms—Activity sequence, hidden Markov model (HMM), indoor localization, smartphone.

I. INTRODUCTION HILE outdoor localization via global positioning system (GPS) is widely used, indoor localization remains a challenge due to the limited visibility of GPS satellites. People spend the majority of time indoors [1], which enables indoor pedestrian localization to become a key technique in locationbased services.

W

Manuscript received January 20, 2014; revised September 14, 2014; accepted November 1, 2014. This work was supported by Shenzhen Dedicated Funding of Strategic Emerging Industry Development Program (JCYJ20121019111128765), Shenzhen Scientific Research and Development Funding Program (ZDSY20121019111146499, JSGG20121026111056204, JCYJ20120817163755063, JCYJ20140418095735587), Major State Basic Research Development Program (2010CB732100), National Natural Science Foundation of China (41201483, 41301511, 41401444), China Postdoctoral Science Foundation (2013M542199, 2014M560671), Navinfo Innovation Funding Program. This paper was recommended by Associate Editor D. Monekosso. (Corresponding author: Q. Li and Q. Mao) B. Zhou, Q. Li, and Q. Mao are with the Department of State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, Wuhan 430079, China (e-mail: [email protected]; [email protected]). Q. Li, W. Tu, and X. Zhang are with the Department of Shenzhen Key Laboratory of Spatial Smart Sensing and Services, Shenzhen University, Shenzhen 518060, China (e-mail: [email protected]; [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/THMS.2014.2368092

Many indoor localization methods are based on wireless radio facility, such as WiFi [2], radio-frequency identification (RFID) [3], Bluetooth [4], and Ultrawide Band (UWB) [5]. These localization methods can be categorized into two types: triangulation and fingerprinting [6]. The former relies on installed expensive hardware, making it neither scalable nor universal. The latter requires pretraining, which is time-consuming. In addition to wireless radio-based methods, dead reckoning (DR) techniques relying on inertial sensors are another way for indoor localization. These methods derive the current location by adding the estimated displacement to the previous estimated one. The biggest advantage of DR method is independence from external infrastructure. DR techniques, widely used for pedestrian localization, known as Pedestrian Dead Reckoning (PDR) [7], leverage lightweight and inexpensive inertial sensors for portable devices, such as accelerometers, gyroscopes, and magnetometers. The devices for PDR include wearable IMU [8], tablet PC, and smartphone [9]. The principle of PDR is integrating inertial sensor measurements over time; therefore, its major drawback is that even small errors in inertial sensors will be magnified by integration [7]. Several solutions have been proposed to prevent the accumulative errors of PDR [10]–[12]. One approach Activity-based Map Matching (AMM) uses activity-related locations as virtual landmarks to eliminate the accumulation of errors [11]–[13]. For example, when a user takes the elevator, there would be an overweight/weightlessness moment and another subsequent weightlessness/overweight moment. These two moments can be detected by the accelerometer, and then, the location of the elevator could be used as the virtual landmark. With the built-in MEMS inertial sensors, smartphones can be considered as primary motion capture sensors, and human activity detection (AD) algorithms based on smartphone have been proposed [14]–[21], which makes AMM a promising method for pedestrian indoor localization. The AMM comprises two basic modules: AD and Map Matching (MM) [22]. The function of the AD module is to detect what a person is doing at a particular instant, such as using an elevator, turning at a corner, or walking upstairs. The MM module identifies the special point on the map where the user is passing based on the detected activity and then matches the estimated position of PDR to the location of the identified special point. Both modules may cause errors. The AD may miss detecting an activity when the activity actually takes place, and it may confuse two different activities and incorrectly detect one when actually the other has taken place. With respect to MM, the exact location of the user in a large indoor environment

2168-2291 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 2

IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS

cannot be determined by the detected activity since there may be more than one special point with the same activity feature. Another neglected factor of the current AMM approaches is the constraint imposed by the topology of the indoor map. For example, a user cannot walk through a wall or other barriers marked on the map. In this paper, we propose a novel activity sequence-based indoor pedestrian localization approach using smartphones.1 To the best of our knowledge, this paper is the first that uses activity sequence for indoor pedestrian localization. The activity sequence consists of several consecutive activities when the pedestrian passes the special points of a building. The approach realizes pedestrian localization by matching the activity sequence to several special points using Hidden Markov Model (HMM). The proposed approach can realize autonomous localization based on PDR even without knowing the initial point. The main contribution is to propose the activity sequence-based map matching (ASMM) model and the ASMM model-based localization approach that takes into account the inertial sensors error and AD accuracy and is robust to a certain degree of error. The remainder of this paper is organized as follows. Section II reviews the related work. Section III presents the activity sequence detection method with an example activity sequence. Section IV introduces the activity sequence-based localization approach. Results and analysis are in Section V. Section VI discusses the proposed method. Section VII concludes the paper. II. RELATED WORK PDR has been applied in indoor localization [23]–[25], which estimates position by accumulating length and heading of each step. The major problem with PDR is that dead-reckoned trajectory accuracy degrades rapidly over time [26]. Therefore, PDR cannot be used on its own for long indoor trips. Some additional mechanisms are required to recalibrate. These approaches can be classified into three generic categories: infrastructure assisted, AMM and indoor map assisted. A. Infrastructure Assisted GPS is one kind of common infrastructure for recalibrating PDR. CompACC [10] triggers periodic GPS measurements to recalibrate the user’s estimated location. Another localization system using GPS as the means to recalibrate and validate the PDR technology is proposed in [27]. However, GPS is unreliable indoors, making it inappropriate for indoor localization. Another alternative approach uses radio frequency devices as the infrastructure to recalibrate the PDR errors. The system proposed in [28] uses RFID technology to recalibrate the PDR errors by placing RFID tags in the environment. RFID technology is also combined with inertial navigation system techniques for accurate pedestrian indoor navigation in [29]. In [30], a system that utilizes HMM to combine WiFi fingerprints localization and DR is proposed. In [31], a constraint approach for PDR and UWB fusion is proposed, which fuses the information of PDR

1 In this paper, indoor localization refers to localization in indoor public spaces, such as office building and shopping mall.

and UWB using a constraint filter with an upper bound in the distance between the estimated positions of both sensors. The proposed infrastructure-assisted approaches for recalibrating PDR can improve the positioning accuracy. However, these solutions rely on infrastructure. Some infrastructure is costly. Others, such as RFID and UWB, have not been installed on smartphones. WiFi fingerprinting is time-consuming and would not work in an environment without WiFi. B. Activity-Based Map Matching An AMM method recalibrates a PDR system by monitoring users’ activities and matching their activities to corresponding specific points. In [11], an indoor positioning approach is proposed based on a combination of Global Navigation Satellite System where available, combining with PDR and AMM. The matching method used in [11] is Nearest Object Matching (NOM), which matches the current estimated location to the nearest object. Inertial sensor features are used as virtual landmarks to prevent accumulation of PDR errors in UnLoc [13]. The matching method in UnLoc is also based on NOM. Another pedestrian tracking system is proposed relying on AMM in [32]. The proposed system uses HMM to estimate pedestrian location and uses detected corners as landmarks to correct the user’s location. In [32], when landmarks (corners) are detected, the pedestrian location is updated with the information at the most similar landmark. The similarity between landmarks and current sensor data is determined based on the distance and heading difference between each landmark and the current location. The criteria of similarity would be invalid, if the sensor error is too large or the distance between different landmarks is too small. Furthermore, the proposed HMM in [32] treats the current location as a hidden state, regarding magnetometer and accelerometer data at the current location as observations. This is different from our system, which uses HMM for activity MM. Because of the sensor’s error, the nearest object is always not the actual one. Therefore, these AMM approaches using NOM as a matching method would encounter mismatch problems. To analyze the mismatch probability of AMM in indoor positioning, a closed-form expression for mismatch probability as a function of PDR sensor error and proximity between two facilities is proposed in [22]. However, this paper only estimates the mismatch probability for a given map and PDR error and does not take into consideration the topology of the interior. Hassan developed a performance model of PDR with activity-based location updates in [33]. He demonstrated that the distance a pedestrian is expected to travel before the PDR is recalibrated is the reciprocal of the density of activity switching points in the indoor environment [33]. ActionSLAM [34] is another approach to activity-based pedestrian localization, which iteratively builds a map of the environment using location-related actions (activities) as landmarks and localizes the user within this map. ActionSLAM is a novel Simultaneous Localization and Mapping (SLAM) method for pedestrian indoor tracking that makes use of bodymounted sensors. ActionSLAM is extended in [35] by introducing heading drift compensation, stance detection adaptation, and ellipse landmarks. The experiments show that the improved

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. ZHOU et al.: ACTIVITY SEQUENCE-BASED INDOOR PEDESTRIAN LOCALIZATION USING SMARTPHONES

ActionSLAM is robust and capable of accurately tracking a user in daily life. SmartActionSLAM [36], another extension of ActionSLAM, uses the integrated motion sensors of the smartphone and an optional foot-mounted inertial measurement unit to track a person. Similar to ActionSLAM, Grzonka et al.[37] incrementally determines the trajectory of a person in a 3-D environment based on motions and activities and is able to accurately recover the trajectory of the person. SLAM technology has been successfully used in real-time smartphone-based indoor navigation [38]. These SLAM-based approaches are different from ours. C. Indoor Map Assisted In pedestrian localization system, the user’s trajectory is restricted by the indoor map and brings opportunities for indoor localization. One common method for localization based on map information is MM, an effective means to improve the accuracy of GPS [39], WiFi [40], and GSM [41] based localization. These MM methods depend on the observations obtained from other means, e.g., GPS, WiFi, and GSM. This is different from our proposed approach, which uses AD as the observation. Other approaches utilize the constraint imposed by the map for independent indoor pedestrian localization. Woodman and Harle proposed a localization approach based on MM using particle filter, which is entirely self-contained and does not rely on infrastructure [42]. An indoor map-assisted pedestrian indoor localization approach is proposed in [43], where the constraint of the indoor map is used to filter out infeasible locations over time. These indoor map-assisted localization approaches use the topology of the map to restrict the pedestrian’s trajectory based on particle filter. Particle filter-based approaches mainly utilize indoor map information for localization. Differently, the purpose of this study is to leverage AD information for pedestrian localization. III. ACTIVITY SEQUENCE Activity sequence implies several consecutive activities when a pedestrian passes the special points of a building, such as a corner, an elevator, an escalator, and a stair, where the pedestrian’s activities are different from walking. These different activities can be detected using AD techniques based on the readings of the built-in sensors in a smartphone. A. Activity Detection Here, we restrict ourselves to structured environments such as office buildings where there are many specific points where pedestrians complete different activities. These activities can be detected using the built-in sensors in smartphones [14]–[21]. In this paper, five types of activities are considered: turning at a corner (normal turn), turning around (U-Turn), taking the elevator, taking the escalator, and walking on the stairs. Some activities only occur at the specific points, which are called “location-related activity” (including turn,2 taking the elevator, 2 Generally, a turn would occur anywhere; in this paper, only the sharp turn is considered, since it usually occurs at a corner.

3

taking the escalator, and walking stairs). Others would occur anywhere, which are called “location-unrelated activity,” such as U-Turn. All activities can be detected by the built-in sensors of the smartphone. Taking the elevator, taking the escalator, and walking on the stairs can be detected using the accelerometer and barometer as pressure changes with altitude, and the acceleration patterns of these activities are different. Normal turn and U-Turn can be detected by the gyroscope and digital compass. The decision tree for AD and signal features of each activity is shown in Fig. 1. The top level classifies the activities into walking and nonwalking based on the standard deviation of acceleration (STD ACC). Walking activity includes walking on the stairs (down and up) and walking normally (on flat ground). Nonwalking activity includes taking the elevator (down and up), taking the escalator (down and up), and keeping still. STD ACC is calculated over a sliding window of size STDwin . A threshold (STD TH) is used to categorize the activity: If STD ACC > STD TH, the activity belongs to walking; otherwise, it belongs to nonwalking. The STDwin is set to 0.8 s, and STD TH is set to 0.5 [44]. The second level divides these two types of activities into pressure changed activities and pressure unchanged activities based on the pressure (Pre.) value measured by the barometer. 1) For walking activities, if the pressure changed, the activity is detected as walking stairs (if the pressure increases, it is walking downstairs; otherwise, it is walking upstairs). Otherwise, it is walking normally. 2) For nonwalking activities, if the pressure changed, the activity is detected as taking the elevator or taking the escalator (if the pressure increases, it is down; otherwise, it is up). To distinguish taking the elevator and taking the escalator, the unique acceleration pattern of taking the elevator is used as the feature. The elevator pattern is caused by the elevator usage, including an overweight period and a weightless period (see Fig. 1). By detecting the overweight and weightless patterns, taking the elevator can be differentiated from taking escalator [45]. If the pressure does not change, the activity is detected as keeping still. Keeping still is not considered in this paper. During the walking process, turns are common activities, including normal turn and U-Turn. Normal turn means turn at a corner, and U-Turn means turn around. Fig. 2 shows the change of the heading and angular velocity when a pedestrian makes a normal turn and U-Turn. The heading value is measured by the digital compass, and angular velocity is measured by the gyroscope. The values of these two sensors change dramatically at the corner. In this paper, the angular velocity measured by gyroscope is used for turn detection, and the heading value measured by digital compass is used to differentiate normal turn or U-Turn. When a pedestrian turns, the turning axis is along the direction of gravity. Hence, angular velocity around the direction of gravity will be generated, detected by the gyroscope. Generally, the acceleration generated when a pedestrian walks normally is much less than gravity. Therefore, the angular velocity of the direction of maximum acceleration can be used to reflect


Fig. 1.


Decision tree for AD and signal feature (acceleration and pressure) of each activity.

change measured by the digital compass is used to distinguish these two activities: (normal) Turn, if ΔH < H T H (2) U-Turn, otherwise where ΔH = abs(mean(H(T − twin : T )) − mean(H(T : T + twin ))), H is the heading measured by the digital compass, T is the turning moment, twin is the time window, which is set to 1.5, mean(H(T − twin : T )) is the average heading value between T − twin and T , H T H is the threshold for distinguish normal turn (called turn) and U-Turn, which is set to 135◦ based on the experiments. Fig. 2. Change of heading and angular velocity when a pedestrian makes turn and U-Turn.

that of the direction of gravity.3 The direction of the maximum acceleration can be obtained by the following equation [32]: axism ax = arg max(accx , accy , accz )

(1)

where axism ax is the axis of the maximum acceleration, accx , accy , and accz are namely the acceleration of x-, y-, and z-axis. Turn is detected using the peak detection algorithm proposed in [46], and the threshold is set to 50 in this paper. The peak detection algorithm is used to find the local maximum or minimum during a period of time [46]. The impact of the threshold to the turn detection result is detailed in [32]. When a turn is detected, there are two possibilities: normal turn and U-turn. The difference between normal turn and U-Turn is the heading change value (see Fig. 2). Therefore, the heading 3 If the angular velocity around direction of gravity was used to calculate the heading change magnitude, the angular velocity measured of the direction of maximum acceleration should be transformed from local coordinate to global coordinate system by multiplying the rotation matrix. However, during our experiments, we detect normal turn and U-Turn using the peak detection algorithm to find the maximum angular velocity rather than using the heading change magnitude. Therefore, we did not use the coordinate transformation.

B. Example of Activity Sequence An activity sequence consists of several activities in chronological order. These activities can be detected by the smartphone carried with the pedestrian. An example activity sequence is shown in Fig. 3. To eliminate the influence of the noise caused by the jitter of the human body, a Butterworth filter of order 4 is used, with a cutoff frequency of 10 Hz. In Fig. 3, the activity sequence includes seven turns, walking down the stair, taking the elevator up, and a U-turn. Based on the detected activities, the pedestrian’s position can be determined by matching these activities to the corresponding special points. IV. ACTIVITY SEQUENCE-BASED LOCALIZATION To eliminate the accumulation of PDR errors, the proposed approach utilizes MM method to find the most likely sequence of special points based on the detected activity sequence. The HMM is used as the MM algorithm. We next introduce PDR and the definition of Indoor Road Network. A. Pedestrian Dead Reckoning PDR is a pedestrian localization scheme that derives the current location by adding the estimated displacement to the previous one. The displacement is obtained from the information of step count and heading. If the previous location is (x, y), the


Fig. 3.

Example of an activity sequence.

Fig. 5.

Fig. 4.

5

Step detection result.

next location is calculated as (x + sl · sc · cos(h), y + sl · sc · sin(h))

(3)

where sl stands for the step length, sc the step count, and h the heading. Step count is obtained by the peak detection algorithm in [46]. Before peak detection, the raw acceleration data should be preprocessed to filter out irrelevant data. For filtering, a Butterworth low pass of order 4 is used, with a cutoff frequency of 10 Hz [46]. The step detection result is shown in Fig. 4. The heading is measured by the compass in the smartphone. The step length is set to a default value added with a random error [30], [43]. B. Indoor Road Network In this paper, each special point where the pedestrian would execute different activities other than walking is defined as the

Node example in the indoor road network.

node. An indoor road network consists of all nodes. The nodes in an office building mainly include corners, elevators, escalators, and stairs. The node attribute is defined as follows: 1) coordinate, coordinate of the node; 2) neighbor nodes; 3) accessible direction (AD); 4) accessible distance of corresponding accessible direction (ADCAD); 5) node type (NT). Fig. 5 is an example of the node; the attribute of node 2 is (x2 , y2 ); {1, 3}; {E, S, W, N }; {dE , dS , dW , dN }; Corner. C. Hidden Markov Model HMM is used to match the activity sequence to the special points of the indoor map, that is, the node of the Indoor Road Network (called node). In this section, we introduce HMM for activity sequence-based localization. The HMM is represented by a finite set of states, each of which is associated with a probability distribution. Transitions among the states are determined


Fig. 6. Example of indoor road network and corresponding transition probabilities.

by a set of transition probabilities. In a specific state, an outcome or observation can be generated by the associated probability distribution. The state is not directly observable to an external observer [47]. An HMM can be represented as λ = (S, V, A, B, π), where 1) S = (S1 , S2 , . . . , SN ) is the set of possible states, N is the number of states in the model; 2) V = (v1 , v2 , . . . , vM ) is the set of observations from the model, M is the number of distinct observation symbols per state; 3) A = {aij } is the state transition probability distribution, aij = pr {qt+1 = Sj |qt = Si } , 1 ≤ i, j ≤ N , where qt denotes the state at time t; 4) B = {bj (k)} is the observation probability distribution in each of the states, bj (k) = pr {vk at t|qt = Sj } , 1 ≤ j ≤ N, 1 ≤ k ≤ M ; 5) π = {πi } is the initial state distribution, πi = pr {q1 = Si } , 1 ≤ i ≤ N . Therefore, under a sequence of observations O = (O1 , O2 , . . . , OT ) where each observation Oi ∈ V, 1 ≤ i ≤ T and T is a system parameter, we want to find the most probable sequence of states Q = (q1 , q2 , . . . , qT ), where qi ∈ S, 1 ≤ i ≤ T. We present our HMM as follows: 1) Hidden States: The hidden states in our HMM are nodes of the Indoor Road Networks. The node is defined as the special points that would make pedestrian complete different activities other than walking. The node attribute includes coordinate and type, such as corner and elevator. 2) Observations: There are two observations in our HMM. The first is the displacement traveled during two consecutive activity moments. The second is the AD result using the AD algorithm. 3) Transition Probabilities: A transition between hidden states is signaled when an activity is detected. To estimate the transition matrix, the indoor road network structure is utilized. Since a pedestrian can only move between adjacent nodes, and each state represents a node, the transition probability is assumed to be uniform over all neighbors of a given node. An example for transition probability estimation is in Fig. 6. 4) Emission Probabilities: The emission probability describes the observation probability distribution at each hidden state. Due to the two observations in our HMM, namely position and activity type, the emission probability consists of two parts: position emission probability and activity type emission


Fig. 7.

Schematic diagram for position emission probability estimation.

probability. As these two observations are independent, the emission probability can be defined as p (zt , mt |ri ) = p (zt |ri ) · p (mt |ri )

(4)

where p (zt |ri ) is the position emission probability, which describes the probability distribution of position observation in a specific hidden state. p (mt |ri ) is the activity detection emission probability, which describes the probability distribution of an activity type given a specific hidden state. According to the principle of PDR, position error is produced by distance estimation error and angle estimation error. Therefore, p (zt |ri ) is made up of two parts: distance observation probability distribution and angle observation probability distribution. Here, these two probability distributions are assumed to be Gaussian distributions [30], [43]. Since distance measurement and angle measurement are independent, the observation probability distribution is defined as p (zt |ri ) = p (dt |di ) · p (φt |φi ) = √ 2

− 2 (φ t −φ i ) 1 2σ ·√ e φ . 2πσφ 1

− 1 2 (d t −d i ) 2 1 e 2σ d 2πσd

(5)

Here, σd is the standard deviation of the measured distance, and σφ is the standard deviation of the measured angle. Based on the distance calculation method of PDR, the distance is in direct proportion to step length; therefore, σd is equal to the standard deviation of step length. dt is the distance between observation and the last matched (determined) state. di is the distance between the i-th state and the last matched (determined) state. φt is the intersection angle between dt and di , as shown in Fig. 7. p (mt |ri ) describes the probability of correct AD for a given hidden state, which is also known as AD confusion matrix. 5) Initial State Distribution: When the first activity is detected, based on the activity type, the initial state distribution is uniform in all candidate nodes. If the start point is unknown, the candidate nodes are the nodes with the same type in the environment; otherwise, the candidate nodes are selected from the neighboring nodes of the start point. 6) Viterbi Algorithm: The Viterbi algorithm is adopted to search for the most probable sequence of hidden states Q = (q1 , q2 , . . . , qT ) for the given observation sequence O = (O1 , O2 , . . . , OT ), a Viterbi variable is defined as δt+1 (j) = max (δt (i) · aij ) · bj (Ot+1 ) , 1 ≤ t ≤ T (6) i


where δt (j) is the highest probability along a single path, at state t, aij is the state transition probability from i to j, and bj (Ot+1 ) is the observation probability at state j. To get the most probable state, ϕt+1 (j) is defined as δt+1 (j) = arg max (δt (i) · aij ), 1 ≤ t ≤ T

(7)

D. Localization Scheme Given the detected activity sequence, our approach aims to find all nodes where the user completes the activities in the activity sequence. The nodes are named “Node Chain” corresponding to the detected activity sequence. During the process of the proposed HMM algorithm, if the number of states is too small, the Node Chain with the highest probability is not always the correct one. Therefore, we adopt a novel method by calculating the probability of every NodeChain candidate using the following equation modified from (6): pt+1 (j) = pt (i) · aij · bj (Ot+1 ) , 1 ≤ t ≤ T

(8)

where pt (j) is the probability of a NodeChain candidate at state t. We adopt the following criteria to select the correct NodeChain from the candidates: phighest /psecondhighest = C

(9)

where phighest means the highest probability of the NodeChain candidate, psecondhighest means the second highest probability, and C is a constant, set to 4 herein. After the correct NodeChain is determined by (9), the user’s location is determined by matching the estimation location of each activity in the activity sequence to the determined NodeChain. The subsequent location can be derived by PDR using the determined node as the starting point. The bias of the smartphone sensors can be inferred from the previous localization process. The step length of the user can be estimated by the detected step number and the distance between the nodes of the determined NodeChain. To estimate the user’s location during the walking process (online localization), (10) is used pest =

N

(pi · pri )

(10)

i=1

where pest is the position estimated by the proposed scheme, pi is the position estimated by every NodeChain candidate, pri is the probability of each NodeChain candidate, and N is the number of the NodeChain candidates. E. System Our approach is summarized with the pseudocode in Algorithm 1. Given the indoor road network and the sensor readings, we first detect the current activity using the AD approach (see line 1). If the activity is a step, we update the position changed after the last activity and calculate the distance traveled after the last activity. The current heading is estimated by averaging the heading data from the detection moment of the last activity and current moment. Based on the distance between the last activity and current heading, the point chain is updated according to the

7

Algorithm 1: System input: Indoor Road Network of the building: IRN input: Sensor readings up to current time t: d1:t output: The activity sequence consisted of all detected activities: ActivitySequence. output: The key point chain consisted of all chain candidates: P ointChain. definition: nchain =number of chain candidates in the P ointChain definition: nactiv ity =number of location related activities in the ActivitySequence 1: Γ=detectCurrentActivities(d1:t ) 2: if stepActivity ∈ Γ then 3: (Δx, Δy)=estimatePositionChangeAfterLastActivity (dt ) 4: DistanceTraveledAfterLastActivity=Δx, Δy 5: DT ALA=DistanceTraveledAfterLastActivity //for better readability 6: CurrentHeading=avg(heading(tlastA citiv ity : t))) 7: CH=CurrentHeading //for better readability 8: if nactiv ity > 0 then 9: for i=1:nchain do 10: if DT ALA · c > ADCAD || CH = AD then 11: delete P ointChain(i) 12: end if 13: end for 14: end if 15: end if 16: if location_related_activity ∈ Γ then 17: AT =ActivityType //for better readability 18: if nactiv ity == 1 then //first detect activity 19: point_candidate=getInitialPoint(DT ALA, CH, AT ) 20: pr_point_candidate=calculatePointCandidate Probabilities 21: addPointtoPointChain(point_candidate,pr_point_ candidate) 22: else //detect more than one activity 23: for i=1:nchain do 24: Neightbors_of_last_points=getNeighbors (PointChain{i}) 25: for j=1:length(Neighbors) do 26: pr_Neightbors=calculateNeighbor Probabilities(IRN ,d1:t ) 27: addNeighbortoPointChain(Neightbors, pr_Neightbors) 28: end for 29: end for 30: end if 31: end if



TABLE I CONFUSION MATRIX OF AD

Turn U-Turn Elevator Stair Escalator

Turn

U-Turn

Elevator

Stair

Escalator

100% 0 0 0 0

0 100% 0 0 0

0 0 100% 0 0

0 0 0 100% 0

0 0 0 0 100%

constraint of indoor road networks. If the DTALA is greater than the ADCAD of the last node (a constant c is used as tolerance for the PDR error, which is set to 0.5 here), or the CH is not equal to the AD of the last node, the NodeChain candidate is deleted (see lines 2–15). As per Section IV-C, if the location related activity is detected, and the activity is the first detected one, the initial node candidates would be obtained based on the DTALA, CH, and AT. The probability of each node is calculated as per Section IV-C5). Then, the node candidates with the probabilities are added to the NodeChain (see lines 16–21). If the activity is not the first detected one, the neighbors of the tail of each chain candidate (last node added when last activity is detected) in the NodeChain is obtained as per Section IV-C5). The probability of each neighbor of the last node is calculated as per Section IV-C6). Similarly, the point candidates with the probabilities are added to the NodeChain (see lines 22–27). V. EVALUATION A. Activity Detection Performance Proof of Concept To evaluate the AD method, a pilot study is conducted. All data were collected using an Android version 4.1.1 Galaxy III smartphone, including accelerometer, gyroscope, magnetometer, and barometer data. The sampling frequency was set to 100 Hz during data collection. Four participants (two females and two males) were asked to complete five activities, according to [18]. Each participant held the smartphone in a hand in front of the body. The sample size of each activity was 20 traces for each activity. For Turn and U-Turn, participants first walked about ten steps, made a turn (U-Turn), and then walked another ten steps. For the elevator, stairs, and escalator, data collection began and ended at two end points of the activity. For the elevator, we collected data for different floors (first floor to the 14th floor), since elevators are stopped by other users in the building. The AD accuracy is calculated using the following equation: Accuracy =

Ti · 100% Ni

(11)

where Ti is the number of the activities that were correctly detected of the i-th-type activities, Ni is the number of all the i-th-type activities. The AD result is shown in Table I. The activity method for a natural track of one participant is shown in Fig. 3. The result shows that the activities can be detected accurately based on the proposed AD approach using a smartphone.

Fig. 8.

Experiment environments. (a) Office building. (b) Shopping mall.

B. Activity Sequence-Based Localization 1) Experiment Setup: To evaluate the overall system performance in real-world environments, we performed experiments in two buildings: an office building, with a 52.5 m × 52.5 m floor plan, and a shopping mall, with a 80 m × 60 m floor plan, as shown in Fig. 8. The proposed system was implemented on the Android platform using the Galaxy S III smartphone, with an accelerometer, a gyroscope, a magnetometer, and a barometer. The participants were asked to walk along six representative routes at constant speed with the smartphone in the hand. Each route was repeated ten times by four participants (two males and two females). Route #1, Route #2, Route #3, and Route #4 are in the office building; Route #5 and Route #6 are in the shopping mall: 1) Route #1 starts from an arbitrary position of the corridor, passes several corners, and arrives at one seat in the office. In this case, the start point is unknown. The traditional PDR scheme cannot work in this case. 2) Route #2 starts from a stair and passes an open area around the elevator. Route #2 includes an open area, where the constraint is poor. 3) Route #3 starts from an elevator and includes a U-Turn activity. It is used to verify the impact of location unrelated activity. 4) Route #4 starts with waking the elevator up to the floor of an office, walking to the office, sitting down for a period of time, and walking to the wash basin. 5) Route #5 and Route #6 are two long routes in the shopping mall: Route #5 starts from an elevator and Route #6 from an escalator. To collect ground truth data, some markers with known coordinates were set along with the routes. When a user walks past a marker, another participant would record the time using another smartphone, which is synchronized with the smartphone used as localization device. Between two markers (the distance is about 10 m), the ground truth is obtained by interpolating using step count. The online localization error is obtained by calculating the Euclidean distance between the estimated position and the ground truth. For offline localization, the error is calculated as follows: N |pei , pg i | (12) Error = i=1 N


Fig. 9.

9

Online localization error results for each route. (a) Route #1. (b) Route #2. (c) Route #3. (d) Route #4. (e) Route #5. (f) Route #6.

where N is the number of the ground truth, pei is the i-th estimated position, pg i is the position of the of the i-th ground truth, |pei , pg i | is the Euclidean distance of pei and pg i . The standard deviation of step length estimation σd is set to 0.1. The standard deviation of measured heading σφ is set to 10◦ . 2) Online Localization Performance: Online localization results of all routes are shown in Fig. 9 for the proposed approach, the proposed approach with known initial point, and the proposed approach with known initial activity. Without initial activity, at the beginning, the average error is high. This is because the initial location is unknown, and the initial position is assumed as a uniform distribution. With increasing step number, the localization error decreases gradually. As the number of encountered activities also increases, after passing a number of steps, the NodeChain consists of the passed nodes determined by the proposed approach (except Route #3), and the location is also determined. For Route #3, if the initial activity is unknown, the trace cannot be determined because the number of activities is insufficient; Fig. 8 shows there are only three turns in Route #3. If the initial activity is known, the localization error decreases faster, as seen in Fig. 9 (there is no initial activity in Route #1). For example, for Route #2, if the initial activity is unknown, the localization error decreases after about 40 steps; if the initial activity is known, it only needs about ten steps for localization error decreasing. This is because the initial activity is special, in Route #2, it is taking the stairs; in Route #3, #4, and #5, it is taking the elevator; in Route #6, it is taking the escalator. The number of these three activity-related nodes is much smaller

than that of the turn. From Fig. 9(f), the location is immediately determined when the initial activity is detected. This is because there is only one up escalator in the shopping mall. Using the special activity-related nodes would help to improve the convergence speed. The fewer the number of activity-related nodes, the faster is the convergence speed. If the initial point is known, based on the AD result, the accumulative error can be eliminated by matching the estimated position of the PDR to the corresponding activity point. There are some special cases in the routes. In Route #2, there is an open area, where turns cannot be detected. In Route #3, there is a U-Turn activity, which is location-unrelated. In Route #4, there is a period of sitting still. The results in Fig. 9 reveal that these cases were addressed. 3) Convergence: Distance traveled before converging to a unique activity chain reflects the convergence speed. The greater the traveled distance, the slower is the convergence speed. Fig. 10 shows the distance traveled of the different routes until the algorithm converges, with and without initial activity. Mostly, with initial activity, the traveled distance is much shorter than without initial activity. Fig. 10 shows that by using AD information, the converge speed on the true location would increase. 4) Performance Versus Activity Detection Accuracy: To analyze the influence of AD accuracy and inertial sensor error to the activity sequence matching result, Route #2 is taken as an example. In Route #2, there are seven activities, including walking upstairs, and six turns. We suppose that walking upstairs would not be detected accurately. In fact, if the barometer is not used (there is no built-in barometer in some smartphones),


Fig. 10. Distance traveled before converging to unique activity chain. Route #1 does not include initial activity, ∞ means Route #3 cannot be converged, 0 means Route #6 can be immediately converged when the initial activity is detected.


when the standard deviation of step length estimation changes from 0.1 to 0.5. As a result of AD error, if σd = 0.1, the matching accuracy is near 100% even if the AD accuracy is 0. The influence of heading error and AD accuracy on matching accuracy is similar to that of step length error, which can be seen from Fig. 11(b). 5) Offline Localization (Tracking) Performance: The offline localization result is obtained by matching the activity position to the NodeChain determined by the proposed approach in Section IV. If the initial position is unknown (Route #1), the trace before first activity is derived retrospectively from the position of the first activity. The tracking trajectory is shown in Fig. 12. The proposed approach tracked pedestrian’s trajectory accurately in the experiment environments. The outcome of the experiments is summarized in Table II (tracking error is the mean of ten trials), and the mean location error of the offline localization is about 1.3 m. VI. DISCUSSION

Fig. 11. Matching accuracy as a function of AD accuracy. (a) Step length error. (b) Heading error.

it would be difficult to distinguish walking stairs from walking normally. Fig. 11 shows the activity sequence matching result as a function of activity (walking upstairs) detection accuracy with different inertial sensors error, expressed by the standard deviation of step length and heading. The matching accuracy is calculated after passing four activities. From Fig. 11(a), when step length estimation error is small, the matching accuracy is not influenced by the AD accuracy. When σd = 0.1, the matching accuracy is near 100, and the AD accuracy is 0. With increases in step length estimation error, the influence of AD accuracy to the matching result is enhanced greatly. The same trend is shown in Fig. 11(b), reflecting the influence of heading error. If the sensor error is small, the activity sequence can match the point well only using turning activity. If the sensor error is large, without the walking upstairs activity, the matching accuracy is low. Fig. 11 shows that activity with a high degree of uniqueness (walking upstairs) is beneficial to activity sequence matching. From Fig. 11, the proposed approach is robust to a certain degree of inertial sensors and AD error. From Fig. 11(a), if the AD accuracy is 100%, the matching accuracy is more than 60%

In contrast with our study, WiFi fingerprinting-based localization requires precalibration of the fingerprints, which is laborintensive and time-consuming [2], [6]. Although autonomous fingerprinting construction approaches are proposed [48], the localization performance based on the autonomous constructed fingerprints is poor. Moreover, WiFi fingerprinting-based approaches rely on WiFi access points. Our approach does not rely on any infrastructure, which can realize autonomous pedestrian localization. We discuss our approach as compared with two state-of-theart calibration-free localization systems: UnLoc [13] and Zee [43]. UnLoc [13] proposed the idea to use inertial sensor features as virtual landmarks to prevent accumulation of PDR errors. However, UnLoc does not consider the ambiguity of the virtual landmark; if there are more than one virtual landmark with the same inertial feature, UnLoc would encounter the mismatching problem. Our proposed approach can avoid the mismatching problem by using the ASMM model. Zee [43] is an indoor map-assisted localization approach which leverages the topology of the map to restrict pedestrian’s trajectory based on a particle filter. However, it is known that particle filtering is timeconsuming, which may be not suitable for online localization based on a smartphone. A. Limitations The proposed approach is based on the assumption that all location-related activities take place at the nodes of the indoor road network. All the nodes are labeled with coordinates. Only the labeled nodes are considered during the localization process. Actually, some activities may take place away from labeled nodes. Our system detects the activity but cannot correctly match the location. This is a limitation of this study, and we are currently investigating methods to address it. The proposed approach works well for structured indoor environments. In these buildings, there are many specific points where pedestrians would complete different activities. Moreover, the pedestrian’s trajectory is restricted by the indoor road


Fig. 12.

11

Offline localization results. (a) Route #1. (b) Route #2. (c) Route #3. (d) Route #4. (e) Route #5. (f) Route #6.

TABLE II EVALUATION RESULTS Activity No. Route No. 1 2 3 4 5 6

Route Length (m)

Detected

Undetected

Location-unrelated

Tracking Error (m)

124.50 106.70 73.25 84.18 161.40 104.50

6 6 5 13 7 6

0 2 0 0 0 0

0 0 1 1 0 0

0.932 1.123 1.012 1.235 1.897 1.581

network in these buildings. The future work needs to address pedestrian localization in open indoor spaces (e.g., lobby). In this paper, step length and heading errors are assumed to be Gaussian distributions [30], [43]. A more realistic PDR error model should be considered.

ACKNOWLEDGMENT The authors would like to thank Danli Li, the editors and the anonymous reviewers for their constructive comments. REFERENCES

VII. CONCLUSION This paper proposed a novel activity sequence-based pedestrian indoor localization approach using smartphones. The activity sequence is first detected using an AD algorithm. Then, the HMM is used to match the activities in the activity sequence to the corresponding nodes of the indoor road network. During the matching process, the constraint of the indoor road network is also taken into account. By the activity sequencebased MM, the proposed approach can realize pedestrian localization even without knowing the starting point in the structured environments. The performance of the proposed approach has been evaluated by experiments in two structured indoor environments. The results show that the proposed pedestrian indoor localization approach can work in these environments using smartphones.

[1] Y. Chon, E. Talipov, and H. Cha, “Autonomous management of everyday places for a personalized location provider,” IEEE Trans. Syst., Man, Cybern., C, Appl. Rev., vol. 42, no. 4, pp. 518–531, Jul. 2012. [2] P. Bahl and V. N. Padmanabhan, “Radar: An in-building RF-based user location and tracking system,” in Proc. 9th Annu. Joint Conf. IEEE Comput. Commun. Soc., 2000, vol. 2, pp. 775–784. [3] L. M. Ni, Y. Liu, Y. C. Lau, and A. P. Patil, “Landmarc: Indoor location sensing using active RFID,” Wireless Netw., vol. 10, no. 6, pp. 701–710, 2004. [4] A. Kotanen, M. Hannikainen, H. Leppakoski, and T. D. Hamalainen, “Experiments on local positioning with bluetooth,” in Proc. Int. Conf. Inf. Technol., Coding Comput. [Comput. Commun.]., 2003, pp. 297–303. [5] R. J. Fontana and S. J. Gunderson, “Ultra-wideband precision asset location system,” in Proc. IEEE Conf. Ultra Wideband Syst. Technol. Dig. Papers, 2002, pp. 147–150. [6] H. Liu, H. Darabi, P. Banerjee, and J. Liu, “Survey of wireless indoor positioning techniques and systems,” IEEE Trans. Syst., Man, Cybern., C, Appl. Rev., vol. 37, no. 6, pp. 1067–1080, Nov. 2007. [7] S. Beauregard and H. Haas, “Pedestrian dead reckoning: A basis for personal positioning,” in Proc. 3rd Workshop Positioning, Navigat. Commun., 2006, pp. 27–35.


[8] J. Bird and D. Arden, “Indoor navigation with foot-mounted strapdown inertial navigation and magnetic sensors [emerging opportunities for localization and tracking],” Wireless Commun., vol. 18, no. 2, pp. 28–35, 2011. [9] A. R. Pratama, R. Hidayat et al., “Smartphone-based pedestrian dead reckoning as an indoor positioning system,” in Proc. Int. Conf. Syst. Eng. Technol., 2012, pp. 1–6. [10] I. Constandache, R. R. Choudhury, and I. Rhee, “Towards mobile phone localization without war-driving,” in Proc. IEEE INFOCOM, 2010, pp. 1–9. [11] D. Gusenbauer, C. Isert, and J. Krosche, “Self-contained indoor positioning on off-the-shelf mobile devices,” in Proc. Int. Conf. Indoor Positioning Indoor Navigat., 2010, pp. 1–9. [12] J. A. B. Link, P. Smith, N. Viol, and K. Wehrle, “Footpath: Accurate mapbased indoor navigation using smartphones,” in Proc. Int. Conf. Indoor Positioning Indoor Navigat., 2011, pp. 1–8. [13] H. Wang, S. Sen, A. Elgohary, M. Farid, M. Youssef, and R. R. Choudhury, “No need to war-drive: Unsupervised indoor localization,” in Proc. 10th Int. Conf. Mobile Syst., Appl. Services, 2012, pp. 197–210. [14] J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Activity recognition using cell phone accelerometers,” ACM SIGKDD Explorations Newslett., vol. 12, no. 2, pp. 74–82, 2011. [15] T. Brezmes, J.-L. Gorricho, and J. Cotrina, “Activity recognition from accelerometer data on a mobile phone,” in Distributed Computing, Artificial Intelligence, Bioinformatics, Soft Computing, and Ambient Assisted Living. New York, NY. USA: Springer, 2009, pp. 796–799. ´ Fábián, and G. Hományi, “An activity recognition system [16] N. Gy˝orb´ıró, A. for mobile phones,” Mobile Netw. Appl., vol. 14, no. 1, pp. 82–91, 2009. [17] K. Ouchi and M. Doi, “Indoor-outdoor activity recognition by a smartphone,” in Proc. ACM Conf. Ubiquitous Comput., 2012, pp. 600–601. [18] S. Khalifa, M. Hassan, and A. Seneviratne, “Adaptive pedestrian activity classification for indoor dead reckoning systems,” in Proc. Int. Conf. Indoor Positioning Indoor Navigat., 2013, pp. 1–7. [19] S. Khalifa, M. Hassan, and A. Seneviratne, “Feature selection for floorchanging activity recognition in multi-floor pedestrian navigation,” in Proc. 7th Int. Mobile Comput. Ubiquitous Netw., 2014, pp. 1–6. [20] L. Pei, R. Guinness, R. Chen, J. Liu, H. Kuusniemi, Y. Chen, L. Chen, and J. Kaistinen, “Human behavior cognition using smartphone sensors,” Sensors, vol. 13, no. 2, pp. 1402–1424, 2013. [21] G. Panahandeh, N. Mohammadiha, A. Leijon, and P. Händel, “Continuous hidden Markov model for pedestrian activity classification and gait analysis,” IEEE Trans. Instrum. Meas., vol. 62, no. 5, pp. 1073–1083, May 2013. [22] S. Khalifa and M. Hassan, “Evaluating mismatch probability of activitybased map matching in indoor positioning,” in Proc. Int. Conf. Indoor Positioning Indoor Navigat., 2012, pp. 1–9. [23] D. Alvarez, R. C. González, A. López, and J. C. Alvarez, “Comparison of step length estimators from weareable accelerometer devices,” in Proc. 28th Annu. Int. Conf. Eng. Med. Biol. Soc., 2006, pp. 5964–5967. [24] Y. Jin, H.-S. Toh, W.-S. Soh, and W.-C. Wong, “A robust dead-reckoning pedestrian tracking system with low cost sensors,” in Proc. IEEE Int. Conf. Pervasive Comput. Commun., 2011, pp. 222–230. [25] U. Steinhoff and B. Schiele, “Dead reckoning from the pocket-an experimental study,” in Proc. IEEE Int. Conf. Pervasive Comput. Commun., 2010, pp. 162–170. [26] R. Harle, “A survey of indoor inertial positioning systems for pedestrians,” IEEE Commun. Surveys Tut., vol. 15, no. 3, pp. 1281–1293, Jul.–Sep. 2013. [27] S. Beauregard, “A helmet-mounted pedestrian dead reckoning system,” in Proc. 3rd Int. Forum Appl. Wearable Comput., 2006, pp. 1–11. [28] S. House, S. Connell, I. Milligan, D. Austin, T. L. Hayes, and P. Chiang, “Indoor localization using pedestrian dead reckoning updated with RFIDbased fiducials,” in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2011, pp. 7598–7601. [29] A. R. Jimenez Ruiz, F. Seco Granja, J. C. Prieto Honorato, and J. I. Guevara Rosas, “Accurate pedestrian indoor navigation by tightly coupling footmounted IMU and RFID measurements,” IEEE Trans. Instrum. Meas., vol. 61, no. 1, pp. 178–189, Jan. 2012. [30] J. Seitz, T. Vaupel, S. Meyer, J. G. Boronat, and J. Thielecke, “A hidden Markov model for pedestrian navigation,” in Proc. 7th Workshop Positioning Navigat. Commun., 2010, pp. 120–127. [31] F. Zampella, A. De Angelis, I. Skog, D. Zachariah, and A. Jimenez, “A constraint approach for UWB and PDR fusion,” in Proc. Int. Conf. Indoor Positioning Indoor Navigat., 2012, pp. 1–9.


[32] K. Park, H. Shin, and H. Cha, “Smartphone-based pedestrian tracking in indoor corridor environments,” Pers. Ubiquitous Comput., vol. 17, no. 2, pp. 359–370, 2013. [33] M. Hassan, “A performance model of pedestrian dead reckoning with activity-based location updates,” in Proc. 18th IEEE Int. Conf. Netw., 2012, pp. 64–69. [34] M. Hardegger, D. Roggen, S. Mazilu, and G. Troster, “Actionslam: Using location-related actions as landmarks in pedestrian slam,” in Proc. Int. Conf. Indoor Positioning Indoor Navigat., 2012, pp. 1–10. [35] M. Hardegger, G. Tröster, and D. Roggen, “Improved actionslam for longterm indoor tracking with wearable motion sensors,” in Proc. 17th Annu. Int. Symp. Int. Symp. Wearable Comput., 2013, pp. 1–8. [36] M. Hardegger, S. Mazilu, D. Caraci, F. Hess, D. Roggen, and G. Tröster, “Actionslam on a smartphone: At-home tracking with a fully wearable system,” in Proc. Int. Conf. Indoor Positioning Indoor Navigat., 2013, pp. 1–10. [37] S. Grzonka, A. Karwath, F. Dijoux, and W. Burgard, “Activity-based estimation of human trajectories,” IEEE Trans. Robot., vol. 28, no. 1, pp. 234–245, Feb. 2012. [38] R. Faragher, C. Sarno, and M. Newman, “Opportunistic radio slam for indoor navigation using smartphone sensors,” in Proc. IEEE/ION Position Location Navigat. Symp., 2012, pp. 120–128. [39] P. Newson and J. Krumm, “Hidden markov map matching through noise and sparseness,” in Proc. 17th ACM SIGSPATIAL Int. Conf. Advances Geographic Inform. Syst., 2009, pp. 336–343. [40] A. Thiagarajan, L. Ravindranath, K. LaCurts, S. Madden, H. Balakrishnan, S. Toledo, and J. Eriksson, “Vtrack: Accurate, energy-aware road traffic delay estimation using mobile phones,” in Proc. 7th ACM Conf. Embedded Netw. Sens. Syst., 2009, pp. 85–98. [41] H. Wang, Z. Wang, G. Shen, F. Li, S. Han, and F. Zhao, “Wheelloc: Enabling continuous location service on mobile phone for outdoor scenarios,” in Proc. IEEE INFOCOM, 2013, pp. 2733–2741. [42] O. Woodman and R. Harle, “Pedestrian localisation for indoor environments,” in Proc. 10th Int. Conf. Ubiquitous Comput., 2008, pp. 114–123. [43] A. Rai, K. K. Chintalapudi, V. N. Padmanabhan, and R. Sen, “Zee: Zeroeffort crowdsourcing for indoor localization,” in Proc. 18th Annu. Int. Conf. Mobile Comput. Netw., 2012, pp. 293–304. [44] A. Brajdic and R. Harle, “Walk detection and step counting on unconstrained smartphones,” in Proc. ACM Int. Joint Conf. Pervasive Ubiquitous Comput., 2013, pp. 225–234. [45] M. Alzantot and M. Youssef, “Crowdinside: Automatic construction of indoor floorplans,” in Proc. 20th Int. Conf. Adva. Geographic Inform. Syst., 2012, pp. 99–108. [46] M. Mladenov and M. Mock, “A step counter service for java-enabled devices using a built-in accelerometer,” in Proc. 1st Int. Workshop ContextAware Middleware Services, Affiliated 4th Int. Conf. Commun. Syst. Softw. Middleware, 2009, pp. 1–5. [47] L. R. Rabiner, “A tutorial on hidden Markov models and selected applications in speech recognition,” in Proc. IEEE, vol. 77, no. 2, pp. 257–286, Feb. 1989. [48] Y. Kim, Y. Chon, and H. Cha, “Smartphone-based collaborative and autonomous radio fingerprinting,” IEEE Trans. Syst., Man, Cybern., C, Appl. Rev., vol. 42, no. 1, pp. 112–122, Jan. 2012.

Baoding Zhou (S’14) received the B.E. degree from School of Information Science and Engineering from Shandong University, Jinan, China, in 2009, and is currently working toward the Ph.D. degree with the State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, Wuhan, China. His research interests include indoor localization and navigation, pervasive computing, and intelligent transportation.


Qingquan Li received the Ph.D. degree in geographic information system and photogrammetry from the Wuhan Technical University of Surveying and Mapping, Wuhan, China, in 1998. He is currently a Professor with Shenzhen University, Guangdong, China and Wuhan University, Wuhan. His research areas include 3-D and dynamic data modeling in GIS, location-based service, surveying engineering, integration of GIS, global positioning system and remote sensing, intelligent transportation system, and road surface checking.

Qingzhou Mao (M’14) received the Ph.D. degree in photogrammetry and remote sensing from Wuhan University, Wuhan, China, in 2008. He is an Associate Professor of Wuhan University, Wuhan, China. His main research interests include satellite navigation system, remote sensing and geographic information system (3S) integrates theory and method, GNSS/IMU navigation and position technology, high-precision laser measurement and point cloud data intelligent processing algorithm, pattern recognition, and vision measurement technology and its application in mapping, road, railways and tunnels and other major projects testing and measurement field.

13

Wei Tu (M’14) received the B.E. and Ph.D. degrees in computer science from Wuhan University, Wuhan, China, in 2007 and 2013, respectively. He is currently a Postdoctoral Fellow with the Shenzhen Key Laboratory of Spatial Smart Sensing and Service, Shenzhen University, Shenzhen, China. His research interests include spatiotemporal data modeling, spatiotemporal data analysis, and spatiotemporal data mining.

Xing Zhang received the B.E. and Ph.D. degrees in geographic information science from Wuhan University, Wuhan, China. He is currently with the Shenzhen Key Laboratory of Spatial Information Smart Sensing and Services, Shenzhen University, Shenzhen, China. His research interests include mobile navigation, visual cognition, ubiquitous computing, and intelligent transportation.