SubTrack: Enabling Real-time Tracking of Subway
Riding on Mobile Devices
Guo Liu
∗
, Jian Liu
†
, Fangmin Li
‡
, Xiaolin Ma
∗
, Yingying Chen
†
and Hongbo Liu
§
∗
WuHan University of Technology, Wuhan, P.R.China
†
Stevens Institute of Technology, Hoboken, NJ, USA
‡
Changsha University, Changsha, P.R.China
§
Indiana University-Purdue University Indianapolis, Indianapolis, IN, USA
Email: guoliuwut@gmail.com, jliu28@stevens.edu, lfm@ccsu.edu.cn, maxiaolin0615@whut.edu.cn,
yingying.chen@stevens.edu, hl45@iupui.edu
Abstract—Real-time tracking of subway riding will provide
great convenience to millions of commuters in metropolitan
areas. Traditional approaches using timetables need continuous
attentions from the subway riders and are limited to the poor
accuracy of estimating the travel time. Recent approaches using
mobile devices rely on GSM and WiFi, which are not always
available underground. In this work, we present SubTrack,
utilizing sensors on mobile devices to provide automatic tracking
of subway riding in real time. The real-time automatic tracking
covers three major aspects of a passenger: detection of entering
a station, tracking the passenger’s position, and estimating the
arrival time of subway stops. In particular, SubTrack employs
the cell ID to first detect a passenger entering a station and
exploits inertial sensors on the passenger’s mobile device to
track the train ride. Our algorithm takes the advantages of the
unique vibrations in acceleration and typical moving patterns of
the train to estimate the train’s velocity and the corresponding
position, and further predict the arrival time in real time.
Our extensive experiments in two cities in China and USA
respectively demonstrate that our system can accurately track
the position of subway riders, predict the arrival time and push
the arrival notification in a timely manner.
I. INTRODUCTION
Subway riding remains as a major convenient means of
public transportation in many years and presents a strong
growing trend as the population of urban cities increases.
Most of the passengers spend their time on reading, playing
games, watching videos, listening to music, or simply dozing
off while taking subways [1]. Current subway administration
only offers trip maps and voice announcements to indicate the
upcoming stops, and such information can be easily ignored
by the passengers in a noisy and crowded train environment.
As a result, people could easily miss their stops unless
they are fully attentive during their subway rides. It will be
convenient and helpful for passengers to obtain the stop times
automatically and track their subway trips. Based on the fact
of the prevalence usage of mobile phones these days, we
seek a solution that can enable real-time tracking of subway
riding on mobile devices, providing intelligent information
of the passenger’s location underground and predicting the
arrival time.
To enable accurate arrival time prediction and notification,
it is critical to keep track of the velocity and position of the
subway train where the passenger rides currently in real time.
Intuitively, the personal mobile devices with built-in GPS
could be utilized to perform the tracking task, however, the
GPS signals are too weak to provide reliable location results
underground. Furthermore, several studies [2, 3] use either
GSM signal strengths or barometer readings along the sub-
way line to track subway train riding. However, GSM signals
are not always available underground, making it hard to scale.
Barometer-based approach is difficult to use when the stations
are built at the similar horizontal planes. New approaches
explore the possibility of using inertial sensors embedded in
mobile devices to detect the train dynamics [4, 5] together
with fixed timetables. However, passengers need to manually
trigger the stop estimation process on their phones. And these
approaches depend on timetables, which cannot reflect the
real-time traffic situation and various dwelling time at train
stops, resulting in only coarse-grained train stop estimation.
In this paper, we make use of the existing inertial sensors
on mobile devices to take one step forward by developing
a real-time fully-automatic passenger position tracking and
arrival notification system without requiring passenger’s in-
volvement. The basic idea is to track the motion states of the
subway train, including both velocity and position, through
accumulated accelerometer readings over time, and then
predict the arrival time to the next stop based on the current
motion state and a subway station map. To facilitate such a
system design, several key challenges should be addressed:
(1) departure time detection and station identification: To
automate the initialization of a subway ride tracking, it is
critical to pinpoint the departure time and station information
as the starting point; (2) smartphone posture alignment:
We need to understand the smartphone’s posture before
proceeding to the inertial sensor data collection, otherwise the
motion states will not be accurately derived; (3) velocity drift
calibration: Velocity drift in inertial systems is inevitable as
indicated in previous investigations [6]. An effective calibra-
tion method should be developed to mitigate the impact of
velocity drift for real-time and accurate tracking.
To trigger the tracking of subway riding automatically
without passenger’s involvement, SubTrack needs to detect a
___________________________________________________________________
This is the author's manuscript of the article published in final edited form as:
Liu, G., Liu, J., Li, F., Ma, X., Chen, Y., & Liu, H. (2017). SubTrack: Enabling Real-Time Tracking of Subway Riding on Mobile Devices. In 2017
IEEE 14th International Conference on Mobile Ad Hoc and Sensor Systems (MASS) (pp. 90–98). https://doi.org/10.1109/MASS.2017.55
passenger enters a subway station and determine the station
information. We find each subway station is associated with
one or two cell tower IDs (Cell-ID). These Cell-IDs are stable
inside each subway station. This useful phenomenon provides
us an unique opportunity to detect whether the passenger has
entered the subway station by examining the specific Cell-
ID connections. In this work, we develop a passenger station
entrance detection mechanism to accurately pinpoint when
and which station the passenger enters.
In order to accurately track the motion states of the subway
train, including both velocity and position, we need to cali-
brate the accumulated velocity and distance which have large
drift errors due to biased motion sensor readings on mobile
phones. Existing studies [6, 7] show that the accumulated
drift error in inertial navigation system increases linearly over
the time, which suggests that such a linear relationship could
be determined if some actual information (e.g., velocity) in
the middle of the subway ride is known. Fortunately, the train
stops from time to time, which indicates that we can capture
the stationary period (i.e., zero velocity) when it stops at each
station. Specifically, we use an acceleration energy based
approach to detect each subway train stopping period, and
exploit it to estimate the velocity drift gradient (i.e., velocity
drift during a sensor reading sample period) and calibrate
the accumulated velocity/distance in real time. With the
calibrated traveling distance, we can then predict the arrival
time to the next stop based on a subway station map. By
integrating all the above components and findings, SubTrack
could perform accurate velocity and position estimation, and
predict the arrival time and send the arrival notification in a
timely manner.
The following contributions are made in this work:
• We propose SubTrack, which leverages inertial sensors
on mobile devices to automatically depict the whole
subway riding trip for each passenger in real-time in-
cluding passenger station entering detection, passenger’s
position estimation during the train ride, and arrival time
prediction.
• We design a passenger station entrance detection mecha-
nism to accurately determine when the passenger enters
a subway station and which station it is.
• We develop a real-time calibration scheme to mitigate
velocity drift error introduced by biased motion sensor
readings to provide accurate passenger position estima-
tion and further enable stop arrival time prediction.
• We align mobile devices held in arbitrary postures by
passengers with the moving direction of the subway
train.
• Extensive experiments are conducted in two cities, one
in China and the other in USA, to validate the perfor-
mance of SubTrack. The experimental results demon-
strate the feasibility and efficiency of SubTrack.
II. CHALLENGES & SYSTEM OVERVIEW
Smartphone-based localization has been widely studied in
many different scenarios [4, 8]. But as far as we know,
Stop
Detection
Distance Estimation
Train Motion State & Time Estimation
Train Departure Detection
User’s Station Entrance Detection
Trigger Trigger
User Motion State Detection
& Data Reconstruction
Coordinate Alignment
Time Prediction
Velocity Drift Correction
Velocity Estimation
Raw Velocity Estimation
Train State Estimation
Database
Station Inner
Cell-ID Database
Next-Stop Info
Output
Arrival Time
Train Position
Arrival Notification
Gyroscope Accelerometer
Fig. 1. Overview of SubTrack system flow.
few work has addressed the problem of localization in
underground public transportation systems, where GPS signal
and wireless infrastructure are not always available. Existing
work mainly relies on the built-in inertial sensors to allow
smartphones to determine their location substantially [4, 5].
But some strong assumptions of these approaches, such
as fixed smartphone postures and reliable sensor readings,
prevent them from practical use. Therefore, we propose a
generic real-time subway rider tracking system leveraging
the build-in inertial sensors of mobile device. In this section,
we first point out several key challenges of the system design,
and then introduce the proposed system work flow.
A. Challenges
To ensure real-time and accurate subway rider tracking, the
following key challenges should be dealt with in the system.
Smartphone Posture Alignment. How the smartphones
is carried along with the passenger has substantial impact on
the inertial sensor readings. Particularly, if the smartphone is
not placed in parallel with the moving direction of the train,
it is difficult to correctly derive the motion state of subway
train based on the raw sensor readings, not even to mention
the smartphone of changing postures. As such, the posture
of smartphone is very critical so that an effective posture
alignment scheme needs to be developed to convert the raw
sensor readings to the coordinate system better serving the
subway rider tracking.
Passenger Station Entrance Detection. As the first step
to track the subway rider, we need to accurately determine
when the passenger enters subway station and which sta-
tion it is. It is intuitive to apply existing GPS-based and
wireless infrastructure-based localization methods to under-
ground transportation system. However, GPS signals can not
penetrate through dense earth, while wireless infrastructures
are rarely available underneath. So an intelligent entrance
detection scheme should be integrated into the SubTrack
system to accurately identify the starting point of the trip.
Unreliable Sensor Readings. Due to the inherent flicker
noise in the electronics and in other components susceptible
to random flickering [6], inertial sensors, including both
accelerometer and gyroscope, will produce biased measure-
ments over time, which lead to inaccurate velocity estimation
results. It is not realistic to remove such bias inside the
sensor themselves. Therefore, an effective data calibration
Time (s)
0 100 200 300 400 500
Acceleration on Z-Axis (m/s
2
)
-1
0
1
Short Time Energy (STE)
0
0.05
STOP
STOP STOP STOP STOP
Fig. 2. Illustration of train stopping detection using short time energy (STE).
scheme will be proposed to compensate the bias in the sensor
measurements, and thereby improve the accuracy of velocity
estimation.
B. System Overview
Given the above challenges, we propose a real time
fully-automatic subway riding tracking system, SubTrack.
It aims to track the motion states of the subway train,
including both velocity and position, through accumulated
accelerometer readings over time, and then predict the arrival
time to next train stop given the current motion state and
subway map. As illustrated in Figure 1, SubTrack includes
two major components: Passenger Motion State Detection
& Data Reconstruction and Train Motion State & Arrival
Time Estimation. In Passenger Motion State Detection &
Data Reconstruction, the smartphone of a particular pas-
senger periodically scan nearby cell-tower IDs (Cell-IDs)
until it finds one matched record in the Subway Station
Inner Cell-ID Database, which is built based on historical
Cell-ID collections from all subway stations. Once the Cell-
IDs match, the departure station of the passenger will be
identified accordingly, and then the SubTrack system will
collect the inertial sensor readings (i.e., accelerometer and
gyroscope) subsequently. Meanwhile, the system will also
remind the passenger to manually input his/her destination,
which is used for arrival time predication and corresponding
notification push.
The raw sensor data is then sent to Train Departure Detec-
tion module to find out whether the passenger is aboard, then
the sensor data before the passenger gets on the train will be
removed. Next, SubTrack converts the sensor readings from
the smartphone’s coordinate to the subway train’s coordinate
via Coordinate Alignment module so that SubTrack can
correctly derive the train’s motion status.
Inside the core component Train Motion State & Arrival
Time Estimation, SubTrack first derives the real-time velocity
of subway train through accumulating the accelerometer
readings over time. Next, we need to identify the stopping
periods of the train, which serve as the stationary reference
points (i.e., zero velocity) to calibrate the velocity estimation.
Integrating both the stationary reference points and specific
historical velocity from Train State Estimation Database, the
velocity drift error will be mitigated based on the linear
model between the drift error and time. The Train State
Estimation Database is built off-line based on the historical
train motion state information (i.e., raw and calibrated ve-
locity, traveling distance and reference points), and it could
Time (s)
100 200 300 400 500
Raw Velocity (m/s)
5
10
15
20
drift
B
drift
A
Reference
Point A
Reference
Point B
Fig. 3. Accumulated raw velocity of a subway train between 5 stations.
be updated automatically after the passenger taking subway.
Given the calibrated velocities, we could keep tracking the
train’s position accurately. Furthermore, we can also inform
the passenger how many train stops remained to his/her final
destination, and predict the arrival time to the next train stop
based on public geographic information (i.e., subway map).
III. TRAIN MOTION STATE TRACKING AND ARRIVAL
TIME PREDICATION
In this section, we focus on detecting whether the train
stops, tracking train’s position and predicting arrival time
using a mobile device (e.g., smartphone). In order to support
the above modules, there are many building blocks includ-
ing Passenger Station Entrance Detection, Train Departure
Detection and Coordinate Alignment which are discussed in
Section IV and Section V.
A. Train Stopping Detection
Train stopping detection aims to detect when the train has
stopped and how long it stays at the train stops (i.e., stopping
periods). More importantly, the stopping periods could serve
as the reference points (i.e., zero velocity) to calibrate the
estimated velocity of the train, which will be discussed in
Section III-B.
We observe that the train has minute vibrations along
with z-axis on the go, which results in the jitter of z-axis
accelerometer readings. We thus can utilize the short time
energy (STE) [9] of the acceleration readings on z-axis to
detect whether the train has stopped. It is important to note
that we use the gravity-aligned acceleration values, which is
discussed in Section V, instead of raw accelerometer readings
to eliminate the impact of mobile device’s arbitrary postures.
In addition, the acceleration readings may also be affected by
the changing postures (e.g., holding the phone and walking
in the train). But it only results in the low frequency sensor
readings, i.e., less than 2 Hz [10], whereas the train on
the go produces a much higher frequency sensor readings.
To eliminate the impact of changing postures, we adopt a
Butterworth high-pass filter with cut-off frequency 50 Hz in
the velocity estimation module. Figure 2 depicts the z-axis
accelerometer readings and corresponding STE on a running
train. It is easy to find that the STE of the accelerometer
readings on z-axis is much lower when the train has stopped.
Therefore, the stopping period of a subway train should be
identified with a carefully designed threshold.
Time (s)
100 200 300 400 500
Velocity (m/s)
0
2
4
6
8
10
12
Off-line On-line
450 500
10.5
11
11.5
300 350
10
12
Fig. 4. Estimated velocity of subway
train between 5 stations.
Time (s)
100 200 300 400 500
Traveling distance (m)
0
500
1000
1500
2000
2500
3000
3500
1069 m
897 m
1019 m
946 m
Station 1
Station 2
Station 3
Station 4
Station 5
Fig. 5. Estimated traveling distance
of subway train between 5 stations.
B. Velocity Estimation
In order to track the train’s position and predict the arrival
time to next train stop, we need to have real-time estimation
on the train’s running speed.
1) Raw Velocity Estimation: The basic idea of raw veloc-
ity estimation is to accumulate the acceleration readings over
time along the moving direction of the train. We assume the
train starts moving at the time τ = 0, and the accelerometer
readings acc(τ) are sampled at a constant rate f sample/sec.
Then the train’s velocity at time τ = t can be expressed as:
vel(t) = vel(0) +
t
X
τ =0
1
f
× acc(τ ), (1)
Figure 3 shows an example of velocity estimation for a five-
stop trip. The estimated velocity should fall back to zero
when the train has stopped. However, due to the biased
accelerometer readings, we still observe non-zero velocities
at the train stops, called velocity drift error, which are
indicated at the ”Reference Points” in Figure 3. Further, such
drift error follows an increasing trend over time.
2) Velocity Drift Calibration: In order to mitigate the
velocity drift error, we develop an on-line (i.e., real-time)
velocity calibration scheme based on the proposition that the
accumulated drift error increases linearly over time, which
has been verified by many existing studies [6, 7].
To derive the linear relationship between the drift error
and time, we need to identify some reference points with
deterministic velocity during the trip, and then fit them to
a linear regression model. As shown in Figure 3, given two
reference points A and B, which correspond to the stopping
period, with zero velocity, the linear velocity drift can be
simply calculated based on the velocity difference between
two reference points. However, the drift error can not be
obtained in real-time, since it has to wait until next reference
point to get the second deterministic velocity to calculate the
velocity difference.
To overcome the above limitation on velocity drift cal-
ibration, we estimate the velocity drift gradient ∆vel (i.e.,
velocity drift during ∆t) based on historical raw velocity
information between neighboring subway stations, which can
be obtained from Equation 1:
∆vel =
1
N
N
X
k=1
drift
k
B
− drift
k
A
L
k
, (2)
where drift
k
A
and drift
k
B
are the velocity drifts of two
neighboring reference points (i.e., stopping periods at two
adjacent stations), while we have N pairs of such neighboring
Time (s)
0 20 40 60 80 100 120 140
Velocity (m/s)
0
5
10
15
20
Approximated
Uniform Motion
s=0
s=1
s=-1
Acceleration
Deceleration
Fig. 6. Typical velocity pattern between two adjacent stations.
reference points from the dataset contributed by passengers’
historical data or crowdsourcing. L
k
is the number of sensor
samples between the k
th
pair of reference points. Based
on the velocity drift gradient ∆vel, the calibrated velocity
vel
cali
(t) that is drift error-free could be obtained in real time:
vel
cali
(t) = vel(t) − drift
A
− (t − t
A
) × ∆vel, (3)
where vel(t) is the estimated velocity at time t, drift
A
is the
velocity drift at reference point A (i.e., the previous reference
point before t) at time t
A
. Figure 4 demonstrates an example
of the velocity estimation for five-stop subway ride with both
our proposed real-time velocity estimation method (i.e., on-
line) and the method in [7] (i.e., off-line). We can find the
calibrated velocities from the two methods are very similar
to each other, so it validates the effectiveness of our proposed
method on eliminating the velocity drift.
C. Traveling Distance Estimation
Given the calibrated velocities, we could derive the current
position of the train through the traveling distance estimation
and thereby predict the arrival time to next train stop.
Specifically, the traveling distance of a subway train can be
obtained based on the integration on the calibrated velocity
over time. The calculated traveling distance at time t from
last train stop can be represented as:
dis(t) = dis(0) +
t
X
τ =0
1
f
× vel
cali
(τ). (4)
Figure 5 gives an example of the estimated traveling distance
over 5 train stops. The estimated stop-to-stop distances are
897m, 1069m, 1019m, and 946m, which achieve the average
error as low as 53m by comparing with the official subway
construction map [11]. Note that there is no accumulated
error in our distance estimation since the distance of each
stop-to-stop segment is estimated independently.
D. Arrival Time Prediction
We next perform arrival time predication to the next train
stop. Accordingly, the proposed system will keep updating
the arrival time to the final destination in real-time, and notify
the riders to prepare to get off near arrival.
In our empirical study in the two cities of China and USA,
we observe that the train usually experiences three motion
phases between two adjacent train stops as shown in Figure 6:
acceleration (s = 1), approximate uniform motion (s = 0)
and deceleration (s = −1). Moreover, the absolute values
of acceleration are almost identical during acceleration and
deceleration, and the motion pattern of the train between two
Cell Tower ID
Connection time percentage (%)
0
5
10
15
20
25
30
35
144195991
144181271
144189253
144189256
144191163
144226642
144191166
41431
144181273
57011
50092
42613
144226648
144226645
144189252
(a) Outside of a station
Cell Tower ID
Connection time percentage (%)
0
20
40
60
80
100
144226642
144226645
144226682
144226685
144226688
144182573
144226722
45478
144226801
144226843
144226846
JDK BTS ZNL XGS PXJ
(b) Inside of 5 stations
Fig. 7. Cell-ID connection time outside and inside of subway stations.
adjacent stations is approximately symmetric with respect to
the middle point of two adjacent train stops.
According to our experimental observations, we assume
that the absolute accelerometer readings during acceleration
and deceleration are a constant value, acc. Based on the
distance to next station that is obtained in Section III-C, we
can predict the arrival time to the next train stop at different
motion phases of the train as follows:
T
arrival
(t) =
vel
uni
acc
+
d
vel
uni
− t, s = 1
vel
uni
acc
+
d−dis(t)−
vel
2
uni
2acc
vel
cali
(t)
, s = 0
vel(t)
acc
, s = −1
, (5)
where acc is the accelerometer readings which can be directly
obtained on the smartphone, and d is the actual distance
between the adjacent stations. vel
cali
(t) and dis(t) are the
calibrated velocity and distance, vel
uni
is the uniform velocity
which can be measured when s = 0 or obtained from
historical velocity when s = 1. Note that the motion status
of the train (i.e., s) can be obtained by applying a predefined
threshold to the short time energy of the train’s well-aligned
moving acceleration, since the short time energy during the
acceleration and deceleration is much larger than that of the
approximated uniform motion.
The arrival time to next train stop can be predicted in real-
time in Equation 5. We can further predict the arrival time
to the final destination by integrating historical remaining
time information to the rest of the stations. In addition,
our algorithm will also intelligently adjust the arrival time
predication at every train stop, and prompt notification when
the train is approaching to the final destination.
IV. DETECTION OF PASSENGER MOTION STATE BEFORE
BOARDING
In this section, we focus on the passenger’s motion state
detection. Specifically, two main topics are included: 1)
Passenger’s station entrance detection. It aims to determine
when the passenger enters the subway station and which sta-
tion it is; 2) Train departure detection. We will detect whether
the train departs by differentiating the accelerometer readings
caused by the moving train from those corresponding to the
walking passenger.
A. Passenger Station Entrance Detection
In the proposed SubTrack system, it is critical to pinpoint
the departure time and departure station as the starting
point of the trip via subway. Since the GPS signal is too
Time (s)
5 10 15 20 25 30
Magnitude of Acceleration (m/s
2
)
0.2
0.4
0.6
0.8
1
1.2
Standing and Walking Start of Subway Train
Walking Standing
Acceleration
Moving
Stopping
Standing
Fig. 8. Acceleration magnitude when the train is accelerating or the
passenger is walking in the station before the train starts to move.
Time (s)
5 10 15 20 25
Average (m/s
2
)
0
0.5
1
Variance (m
1/2
/s)
0
0.05
0.1
0.15
Average
Variance
(a) Before train accelerates
Time (s)
5 10 15 20 25
Average (m/s
2
)
0
0.4
0.8
1.2
Variance (m
1/2
/s)
0
0.05
0.1
0.15
Average
Variance
(b) During train accelerates
Fig. 9. Average and variance of acceleration before and during train starts.
weak to provide reliable location information underground,
instead we present a Cell-ID based approach to perform the
passenger station entrance detection.
Our empirical study observes that the Cell-ID that asso-
ciates with the smartphone is usually unique and stable inside
each specific subway station. It is also much different from
the associated Cell-ID around the subway station entrance but
outside the station, since the base stations deployed inside the
subway station have more reliable wireless links than those
outside.
Figure 7(a) presents the Cell-ID scanning results when the
passenger walks around but outside a subway station entrance
(i.e., JDK station), and the Cell-ID in red color is the one
when the passenger is inside the subway station as shown
in Figure 7(b). We can find that when the passengers are
at the station entrance yet not getting in, their smartphones
could associate with multiple Cell-IDs, and switch among
some of them frequently. Note that it is also difficult for
the passenger to connect to the Cell-ID inside the station.
Figure 7(b) reports the Cell-IDs that the smartphones could
associate with when the passengers stay inside the subway
stations. Note that each color corresponds to one particular
subway station. It is obvious that the smartphones maintain a
long-term connection with one unique Cell-ID (e.g., 98% of
the time connected with 144226642) inside each station. We
could infer that those Cell-IDs must be specifically deployed
inside the stations. Inspired by the above observations, we
can determine whether the passenger enters a subway station
by periodically examining the duration (i.e., λ seconds) that
his/her smartphone has been in association with a unique
Cell-ID inside the subway station.
However, it is a different story for some old subway
stations (e.g., a few in NYC, USA). The Cell-ID associations
around the station entrances are similar as what we observed
previously, but there is no cellular signal inside these stations.
In other words, the smartphone will lose the associations with
