2015 IEEE International Conference on Smart City/SocialCom/SustainCom together with DataCom 2015 and SC2 2015

Enabling Technology for Smart City Transportation

in Developing Countries
Thaw Tar Thein Zan Lounell Bahoy Gueta Toshio Okochi
Research & Development Center Research & Development Center Center for Technology Innovation
Hitachi Asia Ltd. Hitachi Asia Ltd. Hitachi, Ltd.
Singapore, Singapore Singapore, Singapore Tokyo, Japan
tttzan@has.hitachi.com.sg lbgueta@has.hitachi.com.sg toshio.okochi.gj@hitachi.com

Abstract—Traffic in the major cities of developing countries developing urban transportation infrastructure such as Mass
such as the Hanoi and Ho Chi Minh cities is chaotic because Rapid Transit (MRT) system. Smart city is the theme of current
multiple transport modes such as cars, buses, motorbikes, and city development program all over the world. The goal of this
bicycles share the same road lanes together with narrow road research is to bridge the gap between the current technological
infrastructure. In addition, traffic is poorly managed due to lack
of efficient traffic-management system. These problems have
advancement and the infrastructure of developing countries. To
accumulated with the increase in mobility demand because of establish the foundation for smart city development in the
rising urban population. The objectives of this study are to developing countries, this research measures and analyzes
propose a solution to mitigate traffic-congestion problem, to transportation-related data and proposes solutions that support
provide better driving and commuting experience to commuters, sustainable transportation infrastructure. The research goal is to
and to assist in the implementation of efficient transportation- understand transport issues, the cause of traffic jams, and to
management system in the Hanoi city. This paper presents data- propose solution(s) to avoid them. Specifically, the objective is
monitoring and analysis tools to enable smart mobility to analyze the data collected in Hanoi in March 2014. The
capability; adaptive traffic signaling that is the scheduling of analysis result will be the basis to propose solution(s) that will
traffic light timing based on road capacity and congestion
indicator. We propose three congestion indicators: stop count,
benefit the riding public of Vietnam.
velocity, and travel-time variation, which represent the state of
congestion at individual road segment.

Index Terms—Traffic-congestion estimation, stop count,

velocity variation, travel-time variation, CUSUM change
detection.

I. INTRODUCTION
The city population is dramatically increasing due to
urbanization, especially in the developing countries. According
to UN World Urbanization Prospects report [1] in 2014, city
population constitutes 54% of the world population. Vietnam is
a developing nation in South East Asia (SEA) and the
population of its two major cities Hanoi and Ho Chi Minh are
6.4 million and 7.1 million, respectively. The rise in population
has led to an increase in the need for transportation facilities.
In addition, personalized motorized traffic is common in
Vietnam with 85% of transport mode share being contributed Fig. 1. Unmanaged traffic flow in Vietnam
by motorbikes. Moreover, traffic-flow management is
undeveloped, as buses, cars, and motorbikes share the same Conventional data collection in transport research uses
road lanes. This is worsened by the absence of proper flow survey form [2][3] to collect the trip-activity data such as
control of vehicles, especially at intersections (see Fig. 1). In origin, destination, and mode of transport or surveillance and
this figure, traffic builds up not only on the side streets of the traffic camera [4] to collect the traffic information. With the
crossing but also at the crossing center. This effectively leads proliferation of smartphones and development of smartphone-
to narrowing of streets from multiple lanes to becoming single- sensing technology, smartphone-based data collection is
lane only. It both slows down traffic movement and is prone to gaining popularity [5][6][7][8][9]. Smartphones equipped with
road accidents. GPS and other sensors such as accelerometer, magnetometer,
Therefore, it is important to develop an efficient and and gyroscope provide new opportunity for location-based
sustainable transportation-management system in addition to services and traffic-state estimation. While commuting on
vehicles, these phones can be used as probe tools to measure

978-1-5090-1893-2/15 $31.00 © 2015 IEEE 170

DOI 10.1109/SmartCity.2015.66
the location, velocity, and acceleration. This paper presents a staff, and freelancer. In contrast to the conventional travel
method to measure and analyze the traffic condition through survey data, this dataset contains high-resolution data that
android application and data-analysis tools. A new method is include origin-destination (OD) data, trip count, trip purpose,
used to incorporate the traffic-congestion estimation from and travel mode. Trip-activity and social demographic data are
acceleration data into adaptive traffic timing. Cumulative Sum analyzed to extract the trip-making pattern of the commuters
(CUSUM) change detection method is used to detect the whereas sensor data such as acceleration and GPS are used to
deceleration. estimate the traffic-congestion level along the road.
II. DATA-COLLECTION EXPERIMENT Fig. 2 shows the trip purpose and mode share of experimental
Hitachi Ltd (Hitachi) and Hitachi Asia Ltd (HAS) have dataset. Motorbike (M1) contributes 81% of transport mode.
collaborated with Vietnam National University to conduct The trip count by bike and bus are comparable. Subjects tend to
transport-related experiment in Vietnam. Prior to the take bus mainly for home and work while ride bike to home
experiment, a smartphone application is installed on the users' and school. Homebound trips (P1) made up 33% of trip
smartphone and the data were gathered when the user activates purpose and trip count for going to school, for eating, and other
the application. It was developed by Hitachi and modified by purposes are comparable with one another.
HAS. The complete dataset gathered in the Hanoi city
comprises GPS, accelerometer, and trip data for 4,927 trips
M2 M8
made by 144 users within a period of 3 weeks. M1

For each trip, the origin and destination locations were 81%
P1
captured by the app and the users input the trip purpose and
travel mode. The collected data includes social demographic P2
and trip-activity data. The social demographic data comprises
gender, age, occupation, social status, household size, and
house and vehicle ownership. For trip-activity data, two
categories of data, trip information and sensor data, are
recorded. The trip information includes trip profile, the travel
mode, and trip purpose. The trip profile contains the origin and
destination of a trip, its location as well as the time when a P8
subject started the trip, and when she/he has reached a
destination. The travel mode is updated every time the user
changes the mode. The sensor data contains the details of a
Legend
particular trip such as GPS traces from origin to destination, Mode
accelerometer data, and battery percentage of the phone and are M1 – Motorbike M2 – Bike M3 – Bus
collected as long as Trip-Activity Diary (TAD) is running. M4 – Walking M5 – Car M6 – Taxi
M7 – Other M8 – Still
Purpose
TABLE 1. COLLECTED DATA P1 – Home P2 – Other P3 – School
P4 – Eat P5 – Work P6 – Entertainment
Data Type Categories Data Description P7 – Business P8 – Shopping
Personal Name, age, gender, civil
Fig. 2. Trip purpose and transport mode distribution of experimental dataset
information status, occupation
Subject profile
Transport-related Car ownership, motorbike III. DATA ANALYSIS FOR TRAFFIC-CONGESTION ESTIMATION
attributes ownership, residence type The proposed method employs the commuters' smartphone
to estimate traffic congestion. The acceleration and braking are
Origin, destination, trip
Main trips
purpose, location, and time
detected algorithmically by using open source statistical
Trip profile computing and graphic software, namely R[10]. To do that, the
Transportation mode, data are first collected by using android application, TAD
Trip changes
location, and time
installed on the commuters' smartphone. The data are then sent
Location position GPS location, time to the central server through the data connection if Wi-Fi is not
Sensor data available. The program checks if the smartphone is connected
Speed, location to power source and Wi-Fi connection is available. If both
Accelerometer
change
conditions are met, the data are transmitted automatically to the
server. If not, the data are kept in the device and transmitted
Table 1 summarizes the data collected using smartphones. when the user initiates the data transmission. The collected
The information of subject profile is used to characterize the data are then analyzed to detect events like congestion from
sampled population in the experiment. The social demographic frequent braking activities, velocity, and travel-time variation.
data covers diverse group of subjects such as student, engineer, The information is useful to notify the drivers of the traffic

171
condition and send driving recommendation so that they will
be able to make well-informed decisions. The goal of this
research is to utilize sensor infrastructure to understand and Filter
Histogramize
Select
improve traffic condition and road safety via the Velocity from time
velocity data median
dataset period to time
implementation-adaptive traffic light timing in a motorbike- by link velocity
perid+1
dominant city like the Hanoi city.
(a)

A. Stop Count Velocity Level

Stop count is a measure of the number of occurrences of GPS full GPS data sub set Velocity Bin of Time Period
idling made by trip-makers. Idling is as an instance when a dataset of Link ID Time period
vehicle is not in motion but its engine is running. In this paper, Time Period +1
the frequency of these instances is regarded as critical in
GPS data sub set Velocity Bin of
congestion estimation. of Link ID +1 Time period +1
Time Period +2

Acceleration data in forward direction Time Period +3

GPS data sub set Velocity Bin of
of Link ID +2 Time period +2

Set CUSUM target

(b)
Fig. 4. Creating basic velocity profile
Compute CUSUM value
B. Velocity Variation
Another parameter that can assist in estimating traffic
condition is velocity variation. This section provides the
Set pre-determined control parameter
traffic-data analysis to derive velocity variation from GPS data.
Fig. 4 illustrates the basic steps to profile velocity variation. As
shown in Fig. 4, the GPS full dataset is used as the source for
Declare stop if CUSUM value is higher than control parameter the input velocity data. To be more specific, the velocity
information is derived from GPS data, which is calculated
Fig. 3. CUSUM change detection for stop count
based on the Doppler Effect. It is used in this report as it is
shown to be more accurate than the velocity derived from the
In this paper, a method is proposed to derive the number of measurements of two consecutive GPS data.
stop each trip maker is making during every leg of the trip. The The velocity dataset is filtered by link IDs. For each link
stop is characterized by a slow deceleration followed by a ID, the GPS data are ordered based on their timestamps. Then,
sudden acceleration as shown in Fig. 3. Therefore, CUSUM histograms of velocities are created for specified time period.
change detection is suitable for detecting such a phenomenon. For each set of bins in the histogram, a median velocity is
Rather than averaging the signal amplitude, CUSUM observes selected to represent the velocity level.
the accumulation of acceleration and determines if acceleration In representing the velocity level, a median value is
increases or decreases. First, forward acceleration data are selected over mean to eliminate data outliers. In the context of
retrieved. The data are then divided into smaller blocks. There traffic, these outliers may exist such as unnecessary breaking of
are three variables necessary in CUSUM method: CUSUM a motorbike, bus idling at the stop for alighting passengers, or
value, CUSUM target, and CUSUM control parameter. individual trip. These instances can be considered spurious to
CUSUM value is the cumulative sum of the difference between represent close-to-reality velocity levels, which are eliminated
the averaged mean and the CUSUM target. CUSUM target is by selecting only median value.
the median of acceleration data and determines if the data is in As mentioned, Fig. 4 provides a basic velocity profile. This
or out of bound. CUSUM control parameter is σN where N is a process can be further extended to create more targeted
desired number and σ is the standard deviation of averaged profiling of velocities such as:
mean. Any CUSUM value higher than CUSUM target or lower • Individual profile, which is realized by filtering trip-
than negative CUSUM target is regarded as a change. These maker’s identification
changes accumulate over time. When CUSUM value reaches • Population segment profile, which is made by filtering
the CUSUM control parameter, it is flagged as a stop. The only selected population samples by gender,
performance of CUSUM is governed by sensitivity and occupation, or age range
accuracy. Larger control parameter not only reduces false • Day-of-the-week profile, which is realized by filtering
detection but also reduces the sensitivity of detection, while GPS data only with timestamps associated with a
smaller one improves the sensitivity with the price of particular day of the week
increasing false positive.

172
 The individual and population velocity profile are
important to represent traffic perceptions because it targets on
travel experience made by an individual or a specific
population segment. The Day-of-the-Week velocity profile is
useful in monitoring traffic fluctuations on specific days such
as Mondays versus Fridays, or Weekdays against Weekends.
The process illustrated in Fig. 4 was carried out in all GPS
data for all applicable links of the network model of the Hanoi
city.

C. Travel-time variation 1
Logically, the traveled time can be separated into two parts:
the idling time and the actual running time when a vehicle is in 2
motion. Comparing their two values such as in the form of their
ratio can give a more accurate picture of traffic condition in © OpenStreetMap Contributors

contrast to monitoring only the traveled time. However, any Fig. 5. Selected street with high GPS concentration
traveler may not perceive in this manner. She/he may perceive
the idling time duration at one instance or other instances. For a Similarly, velocity-estimation method is applied on the
time-controlled traffic such as intersections with traffic-light Hanoi city dataset. Fig. 7 is the calculated velocity profiles for
control, the traffic-light cycle time is constant for a time period. the two traffic flows for a 1-day period. Fig. 7 corresponds to
In normal case, this cycle time is varied at specific peak-hour the traffic flow going into the city. The velocity levels dip at
periods. In the worst case, it is changed variably depending on two instances at 8-9 AM and 3-4 PM, which can be considered
the current traffic condition. For developing cities, where as rush hours in going to work or school and going back to
improvement of traffic-control infrastructure is needed, it can home from school, respectively, since most of the subjects are
be assumed that the cycle time is constant or is changed at students or school workers.
specific periods.
Assuming that the idling time duration is constant, the
frequency of idling instances (i.e., stop count) is a practical 70
measure of a traffic perception. The stop count is intuitive. 60
First, it can be proportional to the idling time based on the
assumption. Moreover, based on the above narrative, if a trip- 50
Stop Count

maker wants to avoid a perceived time loss due to idling or 40

queuing, she/he will try to lower the number of times it stops or
30
idles by avoiding the road with traffic congestion. Regardless
of the actual traveled distance, she/he will try to find routes 20
with minimal waiting times or minimal number of idling 10
instances. In this paper, the travel time of a particular road
0
segment is derived from GPS and time information collected 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
using TAD.
Hour
IV. APPLICATION OF THE PROPOSED METHODS ON THE HANOI Fig. 6. Stop count of Ton Duc Thang street in the Hanoi city
CITY DATASET
To illustrate, one area with high-GPS concentration and From 6 to 9 PM, the traffic is slow, moving at 20 KPH.
within the city area is shown in Fig. 5. In Fig. 5, the blue and This is the time when traveling public are returning home, or
red arrows depict the traffic flow direction of the street. going out for dining. Furthermore, it can be observed that the
Congestion indicators are calculated for link ID 2, which is traffic flow is lightest in both directions from 1 to 2 PM, which
indicated by red arrow. is early afternoon, and from 9 to 12 PM, when most have
The stop-counting method is applied on the data collected already returned home.
during the Hanoi city’s experiment. The sensor data Similarly, travel-time estimation method is applied on the
corresponding to the selected link were processed to detect the Hanoi city dataset. Fig. 8 shows the average time the commuter
stop count during different times of the day. Fig. 6 depicts the spent in the street throughout the day. The average time spent
stop count of motorists of that particular link. The observed during the peak hours is longer than that during off-peak hours.
stop count is between 50 and 60 per minute during morning Some of the times spent in early morning were zero because
peak hours and between 20 and 30 per minute during evening there was no trip made during those hours. This could be due to
peak hours within the city area and the frequency reduces error in time-tagged GPS data together with very small sample
during non-peak hours. set that makes the time spent close to zero.

173
 40 reference orientation. Therefore, if there is error in orientation
estimation and re-orientation, the accuracy of our algorithms
30
will be affected. Second, our experiment is carried out by
recruiting commuters in the Hanoi city and by installing TAD
Speed (KPH)

application on commuters’ smartphone. The commuters use

20 different brands and models of smartphones, which results in
inconsistency in the collected data. Sensor orientation issue can
10 be solved by detecting and virtually reorienting the device
during data collection, and inconsistency in sensing technology
0
may be solved by algorithmically calibrating the smartphone to
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 a reference mode. These issues will be resolved in our future
implementations.
Time of the Day (Hour)
ACKNOWLEDGMENTS
Fig. 7. Observed velocity profiles of Ton Duc Thang street in the Hanoi city
This research is part of the collaboration between Hitachi,
HAS, and Vietnam National University (University of
From the obtained results, it can be seen that the traffic Engineering and Technology) (VNU-UET). The authors thank
congestion of personalized vehicle dominant city can be Hitachi Ltd for funding collaborative research activities and
derived from the stop count, velocity, and travel time. VNU-UET for supporting data-collection experiments and
organizing workshops in Vietnam.
60

50
Travel Time (Sceond)

REFERENCES
40 [1] United Nations, Department of Economic and Social Affairs, Population
30 Division, "World Urbanization Prospects: The 2014 Revision," United
Nations, New York, 2015.
20
[2] D. A. Hensher and K. J. Button, Handbook of Transport Modelling (2nd
10 Edition), Emerald, Inc., 2008.
0 [3] R. W. Lyles, M. A. Siddiqui, N. Buch, W. C. Taylor, S. W. Haider, G.
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Dennis and B. W. Pigozzi, "Effective Experiment Design and Data
Time of the Day (Hour) Analysis in Transportation Research," NCHRP, Michigan, 2012.
[4] Cisco, "AGT and Cisco Traffic Incident Management Solution:
Fig. 8. Average time spent in Ton Duc Thang street in the Hanoi city Improving Traffic Safety and Efficiency," Cisco, [Online]. Available:
http://www.cisco.com/web/strategy/docs/agt-s-cc-city-tim-aag.pdf.
V. CONCLUSIONS AND FUTURE WORKS [Accessed 7 9 2015].
Our technology collects personal-level data and estimates [5] P. Mohan, V. N. Padmanabhan and R. Ramjee, "TrafficSense: Rich
the transportation necessity. Smartphone is the device that Monitoring of Road and Traffic Conditions using Mobile," Microsoft
connects people and their mobility pattern in multimodal Research, 2008.
transportation system. This research focuses on developing [6] P. Mohan, V. N. Padmanabhan and R. Ramjee, "Nericell: Rich
algorithms to mitigate the traffic congestion through the Monitoring of Road and Traffic Conditions using Mobile Smartphones,"
implementation of adaptive traffic light timing. The goal of in ACM Sensys, Raleigh, NC, USA, 2008.
these analyses is to assist the implementation of adaptive
[7] "AirSage," [Online]. Available: http://www.airsage.com/.
schedule by using traffic-congestion indicator from 3-axis
acceleration. These analyses include stop count, velocity [8] J. Aslam, S. Lim, X. Pan and D. Rus, "City-scale traffic estimation from
variation, and travel-time variation throughout the day. In this a roving sensor network," in 10th ACM Conference on Embedded
research, we applied signal-analysis method on smartphone’s Network Sensor Systems (SenSys '12), New York, NY, USA, 2012.
sensor data to extract features such as stop count, velocity [9] F. Calabrese, M. Diao, G. D. Lorenzo, J. Joseph Ferreira and C. Ratti,
variation and travel time variation of commuters riding "Understanding Individual Mobility Patterns from Urban Sensing Data:
motorbikes. The results show that smartphone-based data A Mobile Phone Trace Example," Transportation Research Part C:
collection and analysis is promising and is the basic method for Emerging Technologies, vol. 26, pp. 301-313, 2013.
the implementation of smart mobility in a developing country. [10] "The R Project for Statistical Computing," [Online]. Available:
Some of the issues that should be considered in future https://www.r-project.org/. [Accessed 7 9 2015].
implementations include sensor orientation and inconsistency
in sensing technology. Sensor orientation is an issue in
smartphone-based data-collection approach where sensors are
placed arbitrarily; the device has to be re-oriented to a

174