3140 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 68, NO.

4, APRIL 2019

Short-Term Traffic Prediction for Edge

Computing-Enhanced Autonomous
and Connected Cars
Shun-Ren Yang , Member, IEEE, Yu-Ju Su, Yao-Yuan Chang, and Hui-Nien Hung

Abstract—Autonomous and connected cars (ACCs), together

with edge computing (EC), have been recognized as a promis-
ing solution to achieve green intelligent transportation for smart
cities. This paper aims to address short-term traffic prediction,
a fundamental enabler for the success of ACC applications, un-
der the European Telecommunications Standards Institute multi-
access EC (MEC) architecture that exhibits constraints different
from conventional cloud computing. First, a data-centric exper-
iment platform is designed and implemented to facilitate traffic
prediction algorithm development. This paper further proposes a
novel short-term traffic prediction model that integrates a traffic
light model and a vehicle velocity model, considering limited com-
puting resources of MEC servers. We note that the effects of traffic
lights are complicated and have not been rigorously examined in
most, if not all, of the related work. This work models the queue-
ing time when a driver arrives at a road intersection and faces a
red light. Moreover, to forecast the vehicle velocity, we propose a
novel low-complexity semiparametric prediction model consider-
ing periodic features and spatial/temporal correlations of dynamic
road events. The experiment results demonstrate that our vehicle- Fig. 1. ETSI multi-access EC (MEC) for autonomous and connected cars.
velocity prediction model achieves almost equivalent accuracy to
the well-known Long Short-Term Memory Neural Network model, I. INTRODUCTION
requiring much lower computational complexity. Our experiment
APABLE of driving without or with little human inter-
results also confirm that since our integrated model pays attention
to the traffic-light effects and the individual driver behavior, it can
more effectively capture the real-time changes of traffic conditions.
C vention and communicating with each other and with the
infrastructure, autonomous and connected cars (ACCs), together
Index Terms—Autonomous and connected cars, multi-access with edge computing (EC) that provides cloud-computing ca-
edge computing (MEC), traffic light model, traffic prediction, pabilities at the edge of cellular networks and thus can support
vehicle velocity model. responsive processing, have been recognized as a promising so-
lution to achieve green intelligent transportation for smart cities
[1]. In contrast to conventional cars, the operation of ACCs
heavily (if not completely) relies on the instructions from the
traffic management centers deployed within the EC servers in
Manuscript received June 21, 2018; revised November 25, 2018 and Jan- the infrastructure. As illustrated in Fig. 1, typically, the ACCs
uary 15, 2019; accepted February 2, 2019. Date of publication February 13,
2019; date of current version April 16, 2019. This work was supported in part exchange information regarding their positions, destinations,
by the Ministry of Science and Technology (MOST), Taiwan, under Grants velocities, etc. and send the information to the EC servers (or,
MOST-106-2923-E-002-005-MY3, MOST-106-2221-E-007-013-, and MOST- more precisely, the traffic management centers therein) period-
107-2221-E-007-034-, and in part by the Ministry of Education Higher Edu-
cation Sprout Project. The review of this paper was coordinated by the Guest ically. The EC servers can then derive/update optimal driving
Editors of the Special Section on Fog/Edge Computing for Autonomous and strategies and recommended routes (in terms of travel delay
Connected Cars. (Corresponding author: Shun-Ren Yang.) and energy consumption) for the ACCs to follow during their
S.-R. Yang is with the Department of Computer Science, the Institute of
Information Systems and Applications, and the Institute of Communications journeys. However, road traffic conditions change dynamically
Engineering, National Tsing Hua University, Hsinchu 300, Taiwan (e-mail:, from time to time. If ACCs’ operation cannot adapt to such
[email protected]). road network dynamics in real time, the expected benefits from
Y.-J. Su is with the MediaTek, Inc., Hsinchu 300, Taiwan (e-mail:,
[email protected]). ACCs can be compromised, and even the overall resultant traffic
Y.-Y. Chang is with the Institute of Communications Engineering, National congestion can worsen. In this case, the ability to predict traffic
Tsing Hua University, Hsinchu 300, Taiwan (e-mail:,[email protected]). conditions [2] becomes essential to the above-mentioned strat-
H.-N. Hung is with the Institute of Statistics, National Chiao Tung University,
Hsinchu 300, Taiwan (e-mail:,[email protected]). egy and route derivations, and is thus a fundamental enabler for
Digital Object Identifier 10.1109/TVT.2019.2899125 the success of ACC applications. Particularly, the EC servers
0018-9545 © 2019 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.
YANG et al.: SHORT-TERM TRAFFIC PREDICTION FOR EDGE COMPUTING-ENHANCED AUTONOMOUS AND CONNECTED CARS 3141

should be able to predict the short-term road traffic conditions ETSI MEC architecture. In general, the development of a traf-
efficiently, and update the strategies/routes back to the ACCs fic prediction algorithm involves the following steps: (1) data
in a nearly road-segment by road-segment manner. This paper collection, (2) data organization, (3) model construction, (4) pa-
aims to address short-term traffic prediction for ACCs under rameters estimation, and (5) model verification. Each of these
the European Telecommunications Standards Institute (ETSI) steps requires a certain level of traffic data (current or historical)
multi-access EC (MEC) architecture, a solution to realize mobile and/or geographic data (e.g., road intersection coordinates). This
edge computing over cellular networks, that exhibits different paper aims to design and implement a data-centric experiment
constraints from conventional cloud computing. platform to facilitate traffic prediction algorithm development:
r specifically, the experiment platform incorporates the func-
A. Literature Review tionalities of OpenStreetMap, which is a collaborative
project that creates free editable geographic data of the
In the literature, short-term traffic prediction has attracted
worlds, and provides the interface with official traffic
extensive research efforts in the past three decades and still
databases (e.g., from Taipei-City Department of Trans-
remains a growingly active research topic nowadays under
portation) to automatically and periodically collect and
the umbrella of the “smart city”. At first, the researchers
parse the latest traffic data according to the geographic
have largely approached such traffic prediction problems
data of the specified areas;
from a classical statistical perspective (e.g., Autoregressive r our platform further transforms the geographic data into
Integrated Moving Average (ARIMA) [3], [4]). Then, with the
graph topological representation of the road networks, al-
dramatically increasing data availability and the advances of the
lowing a user-implemented route planning algorithm to be
computing capability to process such data and of the artificial
embedded into the platform more flexibly;
intelligence (AI) technologies, the recent research interests of r moreover, we also integrate the service of Google Maps
traffic forecasting have been gradually transferred to computa-
Directions API into the platform, by which the model de-
tional intelligence (CI) techniques [5]. Although lacking good
veloper can query the corresponding estimation by Google
theoretical formulation (and thus good model interpretability),
Maps for model performance comparison.
CI approaches, due to their better adaptability, allow to cope
The experiment platform is deployed within the MEC virtu-
with the problems of outliers, missing data and noise [6] and
alization environment as an MEC application.
describe non-linear relationships more effectively, instead.
In addition to the experiment platform design, this paper fur-
More specifically, according to the survey work in [7], the
ther proposes a novel short-term traffic prediction model that
existing traffic prediction methods can be generally divided into
integrates the following two sub-models, considering limited
two categories: parametric approaches and nonparametric ap-
computing resources of MEC servers:
proaches. In parametric approaches, justifiable analytical mod-
Traffic light model: We note that the waiting times at traffic
els (e.g., travel time functions and queueing models) or traffic
lights constitute a significant part of a driver’s total travel
simulation models (macroscopic, microscopic or mesoscopic)
time over a planned route. However, the effects of traffic
are applied, where the structures of the models are established
lights are complicated and have not been rigorously exam-
in advance and then real-time data are used to tune the inputs
ined in most, if not all, of the previous work for traffic fore-
and the parameters. On the other hand, neither the model struc-
casting. This work observes the behavior of traffic lights, an-
tures nor the values of the model parameters of nonparametric
alyzes the queueing time when a driver arrives at the entrance
approaches are predetermined, and are both decided from data.
intersection of a road segment and faces a red light, and de-
Consequently, compared with parametric approaches, nonpara-
rives formulas for the two times that constitute the whole
metric approaches are more flexible. Renowned nonparametric
red-light queueing time: (1) the remaining red-light time and
traffic prediction methods include: classical time series and re-
(2) the time that when the traffic light turns green, the driver
gression modeling and variations [8]–[10], Kalman filter theory
should wait at the traffic light until he/she, following all cars
[11]–[14], support vector regression [15]–[17], artificial neural
ahead, passes the traffic light. Compared with the existing
networks [18]–[20], fuzzy-neural networks [21]–[23], and hy-
models for the fixed-cycle traffic light queueing problem,
brid approaches that integrate the advantages of different types
our model considers real-time traffic signal timing informa-
of prediction methods [24]–[26].
tion from the government transportation department. More-
Although the CI approaches to short-term traffic prediction
over, we develop two computationally efficient approxima-
show good performance, due to their feature of high compu-
tions for the derivation of the red-light queueing time.
tational requirement, they may not be directly applicable to
Vehicle velocity model: To forecast the vehicle velocity
MEC environments that are equipped with limited computing
over a road segment, we propose a low-complexity semi-
resources, compared with a conventional cloud server. To the
parametric prediction model. As indicated in [23], road
best of our knowledge, none of the recent related works have
traffic typically exhibits obvious periodic patterns on a daily
investigated short-term traffic prediction for MEC taking into
basis. However, few studies take such periodic features into
account its characteristics and constraints.
account for traffic prediction (see the discussion in [23]
and related references therein). To address this, this paper
B. Motivations and New Contributions proposes using the mean of the traffic values observed at
This paper investigates traffic prediction systematically from the time point of a day to be estimated as the base of the
both the platform and model perspectives for ACCs under the prediction. Then, to capture the effects of dynamic road
3142 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 68, NO. 4, APRIL 2019

events, we exploit statistical techniques to extract spatial

and temporal correlations from the empirical traffic data.
These correlations are subsequently used to formulate a
linear model for estimating the discrepancies between the
(base) mean traffic values and the real traffic values. Note
that a common problem of ARMA and neural network
methods is their model complexity that requires estimating
a large amount of parameters. Moreover, because of such
complexity, these methods may typically over-fit, leading to
uncontrollable performance for predicting the future data.
Our new semi-parametric model retains the advantage of
ARMA models’ simple, linear structure while using a much
fewer amount of parameters, and thus is more suitable for
MEC with limited resources. Note also that the predicted
vehicle velocity over each road segment can be readily
applied to derive the desired end-to-end travel delays, and
geographical traffic flows, if required.
Fig. 2. ETSI MEC server deployment.
We conduct extensive experiments to investigate the perfor-
mance of each aspect of our novel prediction model. The ex-
periment results demonstrate that our semi-parametric vehicle- prediction algorithms.1 As illustrated in Fig. 2, following the
velocity prediction model achieves almost equivalent accuracy ETSI MEC architecture, our experiment platform is deployed
to the well-known Long Short-Term Memory Neural Network as an MEC application within an MEC server, which adopts
model, requiring much lower computational complexity by con- container-based virtualization that uses Docker for container
sidering periodic features. Moreover, the experiment results in- packaging with Kubernetes for multihost Docker container man-
dicate that the total queueing time at traffic lights accounts for a agement. More specifically, in this container-based environ-
large portion of the total travel time, which can be more than 50% ment, Kubernetes dynamically monitors application resource
in some cases. This justifies that the effects of traffic lights play requirements and/or usage, and then adjusts the resource provi-
a crucial role in road traffic prediction and cannot be neglected. sioned to the managed containers accordingly.
We also compare with the most popular route planner, Google
Maps, which is believed incorporating real-time drivers’ feed- A. System Architecture
back into its core planning strategy. We note that Google Maps
does not specifically consider the traffic-light effects, tending As illustrated in Fig. 3, the experiment platform consists
to under-estimate the travel times. Besides, since Google Maps of two parts: data sources and system components. The data
does not consider the behavior of each individual driver, it can- sources provide geographic data and traffic data for the opera-
not reflect the impacts of an individual driver’s driving habit on tion of the system components. On the other hand, two system
the required average travel time. In contrast, our model pays components-the Route Constructor (RC; Fig. 3 (1)) and Traffic
attention to the traffic-light effects and the driver behavior, and Data Collector (TDC; Fig. 3 (2)) retrieve and parse the geo-
hence can more effectively capture the real-time changes of traf- graphic data and traffic data, respectively. Specifically, the RC
fic conditions. These further justify our algorithm’s estimation transforms the geographic data into the corresponding graph
performance. topology over which the shortest path algorithm can then con-
struct the desired routes. The TDC periodically collects the
traffic data required by the Traffic Predictor (TP; Fig. 3 (3))
C. Paper Organization
from the government open data platform. Given such informa-
The paper is organized as follows: Section II presents the pro- tion, the prediction model developer only needs to design and
posed data-centric experiment platform. Section III describes deploy his/her model and shortest path algorithm in the TP and
the procedure our proposed traffic prediction model follows to RC modules. They can then conduct experiments to evaluate
estimate the travel time over a route candidate. Section IV ex- the performance of the proposed model using the platform. To
plains our model that estimates the queueing time at a traffic do this, the model developer prepares and passes the targeted
light. Section V presents the proposed low-complexity semi- travel origin-destination (OD) coordinates together with the de-
parametric vehicle velocity prediction model considering pe- parture times. The Experiment Controller (EC; Fig. 3 (4)) will
riodic features. Section VI investigates the performance of our send each OD pair to the RC, which will return several route
proposed travel-time prediction model. Finally, Section VII con- candidates back to the EC. The EC will in turn send a route
cludes the paper. candidate together with the departure time one at a time to the
TP to predict the routes travel time. The prediction in the TP
II. DATA-CENTRIC TRAFFIC PREDICTION can be based on the historical traffic data in the TDC. From
EXPERIMENT PLATFORM
This section presents our proposed data-centric experi- 1 The project for the developed traffic prediction platform can be found on
ment platform that facilitates the process of developing traffic GitHub at “https://github.com/vdfjq369/Experiment-platform”
YANG et al.: SHORT-TERM TRAFFIC PREDICTION FOR EDGE COMPUTING-ENHANCED AUTONOMOUS AND CONNECTED CARS 3143

Fig. 4. The XML file provided by the TCDOT.

Fig. 3. The system architecture of the data-centric experiment platform.
The tool SpatiaLite [28] can transform shapefile-format to an
SQL-like format that appears as more readable text rather than
the results predicted by the TP, the EC can generate statisti- a digital vector, so that it can support upper layers. After file
cal data allowing the model developer to evaluate the models format transformation, focused on road information, we can get
performance. Moreover, since more and more drivers rely on data contents including:
Google Maps for their route planning, our platform provides a r osmId: the identification of each road.
connection to Google Maps to enable the model developer to r name: the name of each road.
compare his/her model performance with that of Google Maps. r geometry: a set of latitude and longitude coordinates that
The Google Maps Querier (GMQ; Fig. 3 (5)) module in our constitute road osmId.
platform can automatically query the travel time estimated by For example, there is a road section with osmId 4860359;
Google Maps and return the result to the EC. its name is Sec. 1, ChongQing N Road, and its geometry
In the current implementation of the experiment platform, we is [(121.514199 25.053589), (121.514031 25.052797), (121.
have proposed a traffic prediction model and have deployed the 513851 25.051579), (121.513565 25.050617), (121.513451
model in the TP module. We have also implemented the A-star 25.050238), (121.513251 25.049562)].
shortest path algorithm [27] in the RC module. The details of 2) Traffic Data: Normally, the traffic data in a city can be
the prediction model will be elaborated on in Sections III-V. accessed from the administration. Take Taipei, Taiwan for ex-
ample: the related traffic data can be obtained from Taipei-City
B. Data Sources Department of Transportation (TCDOT). The traffic data are
uploaded to the government-maintained online platform, which
1) Geographic Data (GD): The GD are data about geo-
is open and free for common people. The platform records a
graphic information. In the platform, we obtain geographic
variety of open data sets about transportation; additionally, it
information from OpenStreetMap (OSM). OSM, inspired by
supports different services, like online browsing, file download,
the success of Wikipedia, is a collaborative project to create a
application programming interface (API), etc., in order to facil-
crowdsourced, free, editable map of the world and to provide
itate data access.
free geospatial data for anyone to use and share. Because OSM is
From the extensive open data set on the platform, we con-
multi-information, free, and easy for users to obtain geographic
centrate our attention on the vehicle velocity data, which are
information from it, up until now, OSM data have been widely
detected by sensors deployed on some roads of Taipei city by
applied to daily life, navigation, academic research and even
the government. For the API service, vehicle velocity data are
business. For example, OSM is one of the many cited sources
presented in XML files on the website as shown in Fig. 4.
for Apples custom maps in iOS 6.
The data that OSM provides for users are shapefile-format,
C. System Components
which uses digital vectors to describe points, polylines, and
polygons on a map. However, to support upper layer demand 1) Route Constructor (RC): The RC uses the shortest path
in a readable text format, file transformation is needed first. algorithm provided by the model developer to construct a route
3144 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 68, NO. 4, APRIL 2019

TABLE I
THE GRAPH TABLE

include the header file gragh.h. Second, call the func-

tion loadM ap() to load the GT. Third, use the func-
tion getN eighborN ode(Ni ) to get the coordinates of all
neighboring nodes of Ni . For example, if you want to
know the neighboring nodes of (X5 , Y5 ), call function
getN eighborN ode(X5 , Y5 ) and the result is the string
(X2 , Y2 ) (X6 , Y6 ) (X7 , Y7 ), which consists of several node
coordinates separated by the space character.
2) Traffic Data Collector (TDC): The TDC periodically col-
lects and stores the latest traffic data from the government. When
the system is running, it provides essential traffic information to
other system components. The TDC contains two entities: the
Periodic Traffic Data Parser (PTDP), and the Traffic Information
Fig. 5. The illustration of RC output. Base (TIB):
r The PTDP parses the actual vehicle velocity data that are
presented in XML-format file retrieved from TCDOT via
for the given OD pair. The inputs of the RC are the coordinates of the vehicle velocities API. The data in TCDOT are up-
the origin and destination while the output is a route that consists dated every five minutes and only show the latest vehicle
of all of the intersection coordinates on the route. For example, velocities for each road. However, in our traffic prediction
in Fig. 5 (a), for the input origin and destination coordinates model, long-term historical velocity data are necessary for
(X1 , Y1 ) and (X18 , Y18 ), the output is as shown in Fig. 5 (b). future prediction. As a result, we need an XML parser that
The RC contains three entities: the Geographic Data Analyzer can periodically execute in the system to gather the traffic
(GDA), the Graph Table (GT), and the Shortest Path Algorithm data into our database.
(SPA): r The TIB is a database that stores the historical traffic data
r The GDA is responsible for analyzing the geographic data from TCDOT. The items in the TIB include:
and constructing the corresponding graph topology. The r SectionId: One road is divided into many sections.
geographic data from OSM are presented with one road as SectionId indicates the different sections of the road.
a unit, and does not show the graph topology. However, an r SectionN ame: the name of the road section.
SPA typically requires the road-network graph in order to r AvgSpd: the average velocity of the road section.
derive the desired routes. For this, the GDA helps to ana- r StartW gsX: the longitude of the start of the road sec-
lyze and transform the geographic data into corresponding tion.
graph topology. In the current implementation, the GDA r StartW gsY : the latitude of the start of the road section.
can only analyze OSM format. However, the model devel- r EndW gsX: the longitude of the end of the road section.
oper can easily extend the GDA to analyze other formats. r EndW gsY : the latitude of the end of the road section.
r The GT is responsible for storing the graph topology de- r Date: the date of the record, sensed from the sensor.
rived by the GDA to support the operation of the SPA. The r T ime: the time of the record, sensed from the sensor.
GT is as shown in Table I, where the first column repre- For example, there is a road section with SectionId
sents an intersection P on the map, and the second column Z0111C0; its SectionN ame is Shih Kao Wang Tung,
consists of all of the points that connect to P . For example, Huan Ho Chung Hua; we can know that AvgSpd is
for the intersection (X5 , Y5 ), its neighboring points include 39.964706, StartW gsX is 121.506955, StartW gsY is
(X2 , Y2 ), (X6 , Y6 ), and (X7 , Y7 ). 25.049947, EndW gsX is 121.507057, and EndW gsY
r The SPA is responsible for planning a route from origin is 25.050408 at Date 02-03-2016 and T ime 16:00:00.
to destination. The SPA runs the A-star shortest path al- 3) Traffic Predictor (TP): The TP utilizes the deployed traf-
gorithm as default and allows users to deploy their own fic prediction model to estimate the travel time from origin to
path algorithm on it. If users want to deploy a path al- destination. The input is a route from the RC, and the output is
gorithm themselves, the GT can be used directly. First, the estimated travel time of the route. The TP will divide the
YANG et al.: SHORT-TERM TRAFFIC PREDICTION FOR EDGE COMPUTING-ENHANCED AUTONOMOUS AND CONNECTED CARS 3145

route into several road segments, and estimate the time needed
on each segment, respectively. Finally, all travel times of the seg-
ments will be summed up as the total travel time of the whole
route. Note that the TP can retrieve the related historical traffic
data from the TIB. In this paper, we propose a novel traffic pre-
diction model, which will be elaborated on in Section III. The
TP will run this traffic prediction model by default. However,
users can deploy and substitute their own prediction model for Fig. 6. The queueing and passing times of a road segment.
it. Moreover, because the proposed traffic prediction model is
based on empirical data, the related historical traffic data for
Algorithm 1: Travel Time Estimation.
prediction will be obtained from the TIB.
4) Experiment Controller (EC): The EC manages the exe- 1: Input:
cution of the platform. The input is the OD pair requested by the 2: Route r, which consists of K(r) road segments;
user, and the output is the accuracy ratio of the estimated travel 3: Departure time Tr ;
time between the OD pair. The EC contains two entities: the 4:
Experiment Manager (EM), and the Accuracy Ratio Calculator 5: Output:
(ARC): 6: Estimated travel time tv (r, Tr ) of route r;
r The EM is responsible for handling the interactions be- 7:
tween other system components. When the EM receives 8: Initialization:
the OD pair, it will call the RC to return a route. Next, the 9: τ1 := Tr ;
EM will send the route to the TP and the GMQ respectively 10:
in order to get the travel time of the route. 11: Procedure:
r The ARC is responsible for calculating the accuracy ratios 12: for k := 1 → K(r) do
of the travel times sent from the TP and the GMQ. Specif- 13: Estimate tq (k, τk );
ically, the result that will be returned to the user includes 14: Estimate tp (k, τk + tq (k, τk ));
two accuracy ratios: one is the accuracy ratio between the 15: tv (k, τk ) = tq (k, τk ) + tp (k, τk + tq (k, τk ));
traffic prediction model and the actual traffic conditions; 16: τk +1 = τk + tv (k, τk );
the other one is the accuracy ratio between Google Maps 17: end for K (r )
and the actual traffic conditions. 18: tv (r, Tr ) ← k =1 tv (k, τk );
5) Google Maps Querier (GMQ): The GMQ is responsible
for querying the travel time from origin to destination estimated
by Google Maps. The input is the route from the RC (the same 3) Initialization stage: This stage (line 1) initializes the
as the input of TP), and the output is the total travel time of entering time τ1 into road segment 1 as the departure
the route. The GMQ contains two entities: the Google Maps time Tr .
Directions API (GMD-API) and the Duration in Traffic Parser 4) P rocedure: After initialization, the main procedure is
(DITP): started from line 1 to line 1. Starting from road segment 1
r The GMD-API is the service that provides detailed infor- of route r with starting time Tr , the procedure iteratively
mation about different routings from places to places via estimates the travel time tv (k, τk ) of each segment k and
HTTP requests. the entering time τk +1 into the next road segment k + 1
r The DITP is responsible for parsing the travel time of the (line 12). The travel time tv (k, τk ) can be divided into the
route estimated by Google Maps. following two parts:
r The queueing time tq (k, τk ) at the traffic light of the
entrance intersection of road segment k: When driver
III. THE PROCEDURE FOR TRAVEL TIME ESTIMATION
d arrives at the entrance intersection of road segment k
As mentioned in Section II-C3, we propose a traffic prediction and encounters a green traffic light (e.g., the entrance
model and implement it in TP to estimate the total travel time intersection of segment 2 in Fig. 6), d simply passes
of a route. This section describes the procedure our proposed the traffic light and tq (k, τk ) will be zero. On the other
traffic prediction model follows to estimate the travel time over hand, if d encounters a red traffic light, d needs to spend
a route candidate r for a driver d who wants to pass through r time queueing in front of the traffic light before travel-
at time Tr . The overall operation of the prediction procedure is ing over road segment k. Line 1 estimates tq (k, τk ) by
depicted in Algorithm 1, which is elaborated on in the following. using our proposed traffic-light model, which will be
1) Input parameters: Algorithm 1 takes two input param- discussed in Section IV.
eters: route r and the corresponding departure time Tr r The passing time tp (k, τk + tq (k, τk )) through road
(lines 1-1), where r consists of K(r) road segments as segment k: After the queueing time tq (k, τk ), driver d
Fig. 6 illustrates. will leave the traffic light at τk + tq (k, τk ). Line 1 of the
2) Output result: The output of Algorithm 1 (line 1) is main procedure further estimates the time tp (k, τk +
the estimated total travel time tv (r, Tr ) over r starting at tq (k, τk )) required by d to pass through road segment
time Tr . k by applying our proposed vehicle velocity prediction
3146 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 68, NO. 4, APRIL 2019

model. First, the model predicts the vehicle velocity

over segment k at time point τk + tq (k, τk ). With this
estimated vehicle velocity and the distance information
of segment k from the geographic database, we can
then estimate the passing time tp (k, τk + tq (k, τk )).
The details of the vehicle velocity prediction will be
presented in Section V.
Notice that in line 1, the entering time τk +1 into seg-
ment k + 1 is derived as the sum of τk and the estimated
travel time tv (k, τk ). Finally, the total estimated travel
time tv (r, Tr ) over route r can be expressed as


K (r )
tv (r, Tr ) = tv (k, τk ), (1)
Fig. 7. A signal timing plan for a traffic light.
k =1

where derived as

tv (k, τk ) = tq (k, τk ) + tp (k, τk + tq (k, τk )). (2) tlf (k, τk ) = tc (k, τk ) − {[τk − τs (k, τk ) − to (k, τk )]
mod tc (k, τk )}. (4)
IV. THE TRAFFIC LIGHT MODEL
Besides, the elapsed red time tlb (k, τk ) (see Fig. 7) that will be
This section explains our model that estimates the queueing used to derive tw (k, τk ) in the next subsection can be obtained
time tq (k, τk ) at a traffic light. Note that when driver d arrives as
at the entrance intersection of road segment k at time point
τk , d can face a green light or a red light. Clearly, when the tlb (k, τk ) = tr (k, τk ) − tlf (k, τk ). (5)
traffic light is in the green phase, there is no queueing time, and
tq (k, τk ) = 0. However, if the traffic light is in the red phase, B. tw (k, τk ) Time
the time tq (k, τk ) needed for queueing in front of the traffic The tw (k, τk ) time depends on how many cars have arrived
light includes the remaining red time tlf (k, τk ), and the time and waited at the red light during the elapsed red time tlb (k, τk )
tw (k, τk ) that when the traffic light turns green, driver d should at time τk when driver d reaches the entrance intersection of
wait at the traffic light until he/she, following all cars ahead, segment k (see Fig. 7). Note that tlb (k, τk ) can be derived using
passes the traffic light. Therefore, tq (k, τk ) in this case can be Eq. (5) in the previous subsection. Assume that the cars arriving
expressed as at the entrance intersection of segment k follows a P oisson
process with arrival rate λ(k). In this case, the number of cars
tq (k, τk ) = tlf (k, τk ) + tw (k, τk ). (3)
Ncl (k, τk ) accumulating in front of driver d during tlb (k, τk )
In Eq. (3), the two times tlf (k, τk ) and tw (k, τk ) can be further is P oisson distributed with mean and variance both equal to
derived in the following two subsections. λ(k) tlb (k, τk ). Moreover, of these Ncl (k, τk ) + 1 cars (includ-
ing driver d’s car), let tI (υ, k) (υ ∈ [1, Ncl (k, τk )]) be, upon
the traffic light turns green, the interval between the two time
A. tlf (k, τk ) Time points when car υ and car υ + 1 pass through the traffic light,
To derive tlf (k, τk ), we first consider the operation of a traffic respectively. Then, tw (k, τk ) can be expressed as
light. A traffic light follows pre-defined signal timing plans N cl (k ,τ k )

to transit among the green, yellow and red phases, depending tw (k, τk ) = tI (υ, k). (6)
on the time of day (midnight, rush hours, or non-rush hours). υ =1
Assume that when driver d arrives at the entrance intersection of
segment k at τk , the traffic light of that intersection is executing If tI (υ, k) follows an exponential distribution with probabil-
a signal timing plan which starts from time point τs (k, τk ), as ity density function ft I (υ ,k ) (s) = ϕe−ϕs , mean 1/ϕ and vari-
illustrated in Fig. 7. Moreover, this signal timing plan specifies ance 1/ϕ2 , from (6), tw (k, τk )’s probability density function
that a green, a yellow and a red phases of lengths tg (k, τk ), ft w (k ,τ k ) (s) can be expressed as
ty (k, τk ) and tr (k, τk ), respectively, constitute a cycle of length ⎧ ∞
e−[λ(k ) t b (k ,τ k )]
l
⎪
⎪ ft w (k ,τ k ) (s) =
tc (k, τk ). ⎪
⎪  n =1
⎪
⎪ [λ(k) tlb (k, τk )]n
Note that Fig. 7 also shows an offset period to (k, τk ), which is ⎪
⎪
⎪
⎨ ×
used to coordinate the operation of consecutive traffic lights so
n!
that cars can pass through consecutive road segments without ϕ n (7)
⎪
⎪ × sn −1 e−ϕs for s > 0,
stopping. All the above traffic signal timing information can ⎪
⎪ (n − 1)!
⎪
⎪
be obtained from the government transportation department. ⎪
⎪
⎪
⎩
P (tw (k, τk ) = 0) = e−[λ(k ) t b (k ,τ k )]
l
With such information, the remaining red time tlf (k, τk ) can be for s = 0,
YANG et al.: SHORT-TERM TRAFFIC PREDICTION FOR EDGE COMPUTING-ENHANCED AUTONOMOUS AND CONNECTED CARS 3147

where the mean E[tw (k, τk )], variance V ar[tw (k, τk )], and
skewness κ can be further derived as follows [29]:
E[tw (k, τk )] = λ(k) tlb (k, τk )/ϕ,
V ar[tw (k, τk )] = 2λ(k) tlb (k, τk )/ϕ2 , and

κ= 3/ 2λ(k) tlb (k, τk ). (8)
Eq. (7) is not a closed form solution and can be approximated
by the following two approaches.
Normal approximation: We can use the normal distribu-
tion N (E[tw (k, τk )], V ar[tw (k, τk )]) with the same mean
and variance as the tw (k, τk ) time to approximate the dis-
tribution of tw (k, τk ), where the mean E[tw (k, τk )] and
variance V ar[tw (k, τk )] have been provided in (8).
Shifted gamma approximation: Another approach is to
use a shifted gamma distribution with probability density
βα α −1 −β (s−δ )
function ft w (k ,τ k ) (s) = Γ(α ) (s − δ) e for s ≥ δ,
2
√
mean δ + (α/β), variance α/β , and skewness 2/ α, to
approximate tw (k, τk )’s distribution. Similarly, we should
use the mean, variance and skewness of tw (k, τk ), provided
in (8), to set the corresponding values of the shifted gamma
distribution. Then, solving the following system of equa-
tions
⎧ α
⎪
⎪ E[tw (k, τk )] = δ + ,
⎪
⎨ β
α
V ar[tw (k, τk )] = 2 , (9)
⎪
⎪ β
⎪
⎩ κ = √2 ,
α

we obtain
Fig. 8. Velocity series.
4
α = 2,
κ
2 the traffic behaviors in Taipei City, Taiwan. Fig. 8 shows the
β= , and
V ar[tw (k, τk )] κ velocity Zk ,t series observed on a certain road segment during
different days, where weekdays and weekends are investigated
2 V ar[tw (k, τk )] separately.
δ = E[tw (k, τk )] − .
κ This figure indicates that the Zk ,t values at the same time-step
during the observed consecutive days (e.g., [7:10, 7:11] from
V. THE VEHICLE VELOCITY PREDICTION MODEL
2015-09-21 to 2015-09-25 for the weekday case, and [8:25,
To estimate the passing time tp (k, τk + tq (k, τk )) of driver 8:26] on 2015-09-19, 20, 26, and 27 for the weekend case) are
d through road segment k at line 14 of Algorithm 1, this sec- close. That is, the velocity series exhibits repetitive patterns on
tion proposes a low-complexity semi-parametric vehicle veloc- a daily basis. Note that such phenomenon are also observed
ity prediction model. First, through intensive investigation of on most of other road segments. However, due to the space
real traffic data, we observe that road traffic exhibits similar limitation, the corresponding details are omitted. Those readers
characteristics on a daily basis. Therefore, we propose using the who are interested in these statistical road traffic data are referred
mean of the traffic values observed at the time point of a day to to [31].
be estimated as the base of the prediction. Then, to capture the Inspired by this observation, we claim that when predicting
effects of dynamic road events, we exploit statistical techniques the velocity value at a given time-step t, the mean of the velocity
to extract spatial and temporal correlations from the empirical values at this t during the past few days can serve as the basis
traffic data. These correlations are subsequently used to formu- of an accurate prediction. To be more specific, let Zkd,t be the
late a linear model for estimating the discrepancies between the  observed on road k at time-step t of day d, and
velocity value
(base) mean traffic values and the real traffic values. X̄k ,t = ( li=1 Zkd−i
,t )/l be the long-term average velocity at this
t, which can be obtained directly from the historical data. Then,
A. Empirical Data-Based Vehicle Velocity Prediction Model we can express Zkd,t as
We collect realistic road traffic data from the Ministry of
Transportation and Communications, Taiwan [30] to analyze Zkd,t = X̄k ,t + Ykd,t , (10)
3148 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 68, NO. 4, APRIL 2019

road segment k. We explore the spatial correlation between Ykd,t

and Ykd( j ) ,t−s , for j ≥ 1. Fig. 9(b) shows the scatter dia-
k (j ) ,k
gram for the case that j = 1. A part of the scatter diagrams for
j ≥ 2 can be found in [31]. We observe an interesting result
that Ykd( j ) ,t−s is also positively and linearly correlated with
k (j ) ,k

Ykd,t . This result is consistent with the fact that if the traffic on
some road segments is congested, their immediately following
road segments will have a high probability to be congested later
as well.

C. A Linear Spatial-Temporal Model for Estimating Ykd,t

From the analysis of real road traffic data in SubSection V-B,
we see that Ykd,t is linearly and positively correlated in time
with Ykd,t−j and in space with Ykd( j ) ,t−s . Based on this
k (j ) ,k
observation, we then propose a linear spatial-temporal model,
STT(m, n), which is expected to accurately estimate Ykd,t . This
STT(m, n) model is expressed as follows

m 
n
Ŷkd,t = ak + bk ,i Ykd,t−i + ck ,j Ykd( j ) ,t−s . (11)
k (j ) ,k
i=1 j =1

In (11), the ak , bk ,i , and ck ,j parameters can be easily estimated

by using, for example, the well-known least square method [32].

D. Distance of Road Segment k

Fig. 9. Statistical correlations.
To estimate tp (k, τk + tq (k, τk )), besides the vehicle veloc-
ity over road segment k, the distance of road segment k is also
where Ykd,t represents the estimation error, which results from required. For this, rather than directly using the Euclidean dis-
the real-time traffic condition change. With the given X̄k ,t in tance, our platform uses the Vincenty’s formula [33] to calculate
(10), the problem of predicting Zkd,t can be transformed to the the distance of road segment k, due to the formula’s capability
of taking the curvature of the Earth into account.
problem of predicting Ykd,t .
VI. PERFORMANCE EVALUATION
B. Temporal and Spatial Correlations for Estimating Ykd,t This section investigates the performance of our proposed
This subsection utilizes scatter diagrams to extract temporal travel-time prediction model in Section III. We first show the
and spatial correlations from the empirical traffic data for Ykd,t performance of the low-complexity semi-parametric vehicle ve-
estimation. locity prediction model in Section V, specifically by comparing
1) Temporal Correlation: We explore the temporal correla- it with that of the well-cited Long Short-Term Memory (LSTM)
tion between Ykd,t and Ykd,t−j , j ≥ 1. Fig. 9(a) plots the scatter Neural Network model [20]. Next, we examine the effects of
diagram for the case that j = 1. This figure shows that Ykd,t is our traffic light model’s input parameters on the estimated total
systematically incremented by the increase of Ykd,t−1 , and all travel time. Lastly, we demonstrate the overall performance of
our travel-time prediction model and compare it with that of
(Ykd,t−1 , Ykd,t ) points tend to fall along a line. We also observe
Google Maps.
such correlation between Ykd,t and Ykd,t−j for other small j ≥ 2
[31]. This is reasonable because the traffic condition during a A. Vehicle Velocity Prediction Model
short consecutive period of time-steps does not change dramat-
ically but smoothly. 1) STT Model Configuration: Clearly, the larger the m and
2) Spatial Correlation: For each road segment k, we choose n values in the STT model, the higher the prediction accu-
a set of neighboring road segments which have strong spatial racy. Unfortunately, this requires to estimate more parameters,
correlation with k. These road segments are arranged in decreas- increasing the overall prediction complexity. The goal is thus
ing order of their correlation intensity with road segment k, and to select the smallest m and n while maintaining the desired
are denoted as k (1) , k (2) , k (3) , ...Note that different approaches prediction accuracy.
 
can be applied to derive this set, e.g., based on the distance Let Rk2 = d t (Ykd,t )2 , and denote Rk2 (m, n) as the
 
t (Yk ,t − Ŷk ,t ) when using the
d d 2
or the traffic-flow intensity. Let sk ( j ) ,k be the average amount sum of square errors d
d
of time-steps for vehicles to travel from road segment k (j ) to STT(m, n) model to predict Yk ,t . Further, define the
YANG et al.: SHORT-TERM TRAFFIC PREDICTION FOR EDGE COMPUTING-ENHANCED AUTONOMOUS AND CONNECTED CARS 3149

TABLE II
THE IMPROVEMENT RATIOS UNDER DIFFERENT (m, n) SETTINGS
IN THE STT MODEL

improvement ratio I(m, n) as

N 
1  Rk2 (m, n)
I(m, n) = 1 − ,
N Rk2
k =1

where N is the total number of road segments in the considered

vehicular network. Using the improvement ratio as an index,
we investigate how to select an appropriate (m, n) pair. Given a
(m, n) setting, we respectively use twenty-eight weekdays and
twelve weekends as training sets to derive the model parame-
ters in (11). With these parameters and the resulting models, we
then estimate Ykd,t for the considered training-set days and com-
pute the corresponding improvement ratios. Table II shows the
improvement ratios under different (m, n) settings in the STT
model. We observe that when (m, n) = (1, 1), the STT model
has already achieved a high improvement ratio near 85%. If we
continue to increase the m or n value, the corresponding im- Fig. 10. Average prediction accuracy.
provement becomes indistinct. This fact allows us to reduce our
model complexity by simplifying the general model in (11) as
periments of Fig. 10. For the LSTM model, the program is
Ŷkd,t = ak + bk Ykd,t−1 + ck Ykd( 1) ,t−s . (12) implemented in Python, using Keras library with TensorFlow
k ( 1) , k
backend. In particular, we follow the settings used in [20], and
In (12), there are only in total 3RN parameters in our semi- consider two cases for demonstration purposes: epoch = 3 and
parametric prediction model, where
r R: the number of classes (e.g., weekdays or weekends); epoch = 5, which mean that the LSTM model will work through
r N : the number of road segments in the considered vehic- the entire training dataset for three and five times, respectively.
Figs. 10(a) and 10(b) exhibit that our semi-parametric predic-
ular network. tion model achieves almost equivalent accuracy to LSTM in
2) Prediction Performance: Utilizing the simplified model both the short-term (i.e., near the current time) and long-term
in (12) together with the trained parameters, we next examine (i.e., far in the future) predictions. This justifies that X̄k ,t in (10)
the accuracy of our proposed semi-parametric prediction model. provides an accurate estimation foundation for Zkd,t . Moreover,
Specifically, given the traffic data in the past and at the current based on the identified spatial and temporal correlations, our
time t, we iteratively predict the traffic conditions at t + 1, t + STT model can estimate Ykd,t effectively. On the other hand, to
2, . . .We evaluate the prediction accuracy at time-step t in terms
predict a Zkd,t , LSTM requires an execution time about 259s
of At which is defined as
  for epoch = 3 and 369 s for epoch = 5 (mainly for training),
1    Zkd,t − Ẑkd,t 
N D
while our model requires an execution time only 0.04s (under
At = 1 −  , (13) the same computing equipment: Intel Core i7-6500U 2.5 GHz,
ND  Zkd,t 
k =1 d=1 8GB RAM, and Windows 10 64-Bit). Therefore, our model is
where D is the number of observed days. Fig. 10 shows the more suitable for MEC with limited resources. Notice that in
results for the following two cases: our platform, the historical velocity data have been stored in
Case 1: Now at 7AM on a weekday, and to predict for the the local TIB, and the TP can retrieve the related data from the
rush-hour period from 7AM to 8AM (see Fig. 10(a)); TIB directly. Therefore, no extra communication and accessing
Case 2: Now at 2PM on a weekday, and to predict for the times are required to obtain drivers’ behaviors from a remote
non-rush hour period from 2PM to 3PM (see Fig. 10(b)). government database in real time when the TP executes a traffic
For more experiment results, the readers are referred to [31]. prediction model.
Note that we collect speed data from September 19, 2015 to In addition to accuracy, we also conduct an ns-3 experiment
July 29, 2016 with the updating frequency of five minutes, from to simulate the mobility and packet transmission operation of
which fourteen weekdays are observed as test sets for the ex- an ACC within the radio coverage of a 4G LTE base station
3150 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 68, NO. 4, APRIL 2019

Fig. 11. Timeliness for route update transactions.

deployed at a road intersection. We compute each required

time component during an HTTP transaction between the Fig. 12. Effects of traffic light model parameters.
ACC and an MEC server that updates a new best route for
the ACC, and check if the ACC can complete successfully the use the vehicle velocity prediction model in Section V to derive
route update transaction within the passing time through the the predicted vehicle velocity Zkd,t on each road segment k that
coverage of the 4G LTE base station. The parameter setting the driver will pass, and (2) then select a uniform random variate
of the simulation experiment and the experiment results are from the range [Zkd,t × 0.8, Zkd,t × 1.2] as the driver’s revised
provided in Figs. 11(a) and 11(b), respectively. As Fig. 11 estimated velocity on road segment k. With the estimated veloc-
indicates, given the coverage passing time 500 m*2/(50 km/h) ity on each road segment and the queueing time at each traffic
= 72 sec., due to its low computational complexity, our model light (derived using the traffic light model in Section IV) over
can support dynamic route updates for ACCs in a road-segment the considered route OD1, we can finally obtain the estimated
by road-segment manner, while LSTM cannot. total travel time for each of the 100 drivers.
Fig. 12(b) shows the average total queueing and estimated
B. Traffic Light Model travel times of the 100 drivers. The figure indicates an intuitive
result that, as the car arrival rate λ(k) and the mean of tI (υ, k)
In our traffic light model developed in Section IV, we assume
increase, both the average total queueing and travel times exhibit
that:
r the cars arriving at the entrance intersection of segment k linear increase accordingly. This is because the higher the λ(k)
and E[tI (υ, k)] values, the heavier the traffic condition. More-
follows a P oisson process with arrival rate λ(k);
r tI (υ, k), i.e., the time interval between car υ and car υ+1’s over, the experiment result indicates that the total queueing time
accounts for a large portion of the total travel time. For example,
passes through the traffic light at the entrance intersection
the average total travel time under E[tI (υ, k)] = 0.5 (sec.) and
of segment k when the traffic light turns green, follows an
λ(k) = 0.2 (vehicles/sec.) is estimated as about 25.8 minutes,
exponential distribution.
within which the average total queueing time is responsible for
These P oisson and exponential assumptions are for illustra-
about 11.3 minutes. This justifies that the effects of traffic lights
tion purposes, which allow us to perform mean-value analysis.
play a crucial role in road traffic prediction, and the accuracy of
This subsection conducts an experiment to investigate the ef-
the traffic light model merits further investigation (note: traffic
fects of the car arrival rate λ(k) and the mean of tI (υ, k) on the
light statistical data have not yet been made available in Taiwan
total queueing time at traffic lights and the estimated total travel
and many other countries).
time. Specifically, the experiment considers the travel times of
100 drivers over a route in Taipei City with distance about 11
kilometers (referred to as origin-destination pair one (OD1)), as C. Performance of Proposed Travel Time
Estimation Algorithm
illustrated in Fig. 12(a). Moreover, we note that different drivers
have different driving habits. To take driving habits into account, This subsection examines the performance of our proposed
during the prediction of a driver’s total travel time, we (1) first travel time estimation algorithm, i.e., Algorithm 1 in Section III.
YANG et al.: SHORT-TERM TRAFFIC PREDICTION FOR EDGE COMPUTING-ENHANCED AUTONOMOUS AND CONNECTED CARS 3151

Fig. 13. The OD pairs for experiment.

Following the experiment in Section VI-B, the present exper-

iment estimates the travel times of 100 drivers in Taipei City,
taking into account their different driving habits. In addition to
OD1 in Fig. 12(a), we also consider another two, more com-
plicated routes, OD2 and OD3 in Fig. 13. Furthermore, our
experiment investigates both the non-rush hour and rush-hour
cases. Specifically, the car arrival rate λ(k) is set as 0.15 (vehi-
cles/sec.) and 0.3 (vehicles/sec.) for the non-rush hour and the
rush-hour cases, respectively. On the other hand, since the inter-
passing time tI (υ, k) is less relevant to the traffic congestion
level, we assume that tI (υ, k) is exponentially distributed with
mean 1.5 (sec.) for both the non-rush hour and the rush-hour
cases. Fig. 14. Comparison of estimation performance.
In this experiment, we compare the performance of our pro-
posed algorithm with that of
r PUAS (standing for our Proposed algorithm while Using rush hours are longer than those in non-rush hours under each
Actual vehicle Speeds): which also applies the traffic light individual algorithm and a given route. Moreover, as Fig. 14(a)
model of our proposed travel time estimation algorithm, exhibits, for the shortest route OD1 which requires a low estima-
while the passing time tp (k, τk + tq (k, τk )) through each tion complexity, the estimation differences among PUAS, GM
road segment k is derived by using the actual vehicle speeds and our proposed algorithm are insignificant. However, as the
from the government traffic database (instead of using the route length increases (OD1→OD3), the estimation difference
proposed vehicle velocity prediction model in Section V); between GM and PUAS increases accordingly, while that be-
r GM: which refers to the Google Maps traffic prediction tween PUAS and our algorithm still remains almost negligible,
algorithm, whose estimated travel times can be queried as shown in Figs. 14(b) and 14(c). This is due to the fact that
and retrieved via the GMQ module in our experiment GM does not specifically consider the traffic-light effects and
platform. thus may not be able to effectively capture the real-time changes
Given that traffic light statistical data have not yet been made of traffic conditions. In this case, GM tends to under-estimate
widely available, PUAS here serves as a reasonable bound for the travel times.
performance comparison. We also compute the standard deviation (SD) of the 100
Fig. 14 shows the average travel times estimated by the three drivers’ estimated travel times under each algorithm, and plot
algorithms. It makes sense that the estimated travel times in the respective range between [the mean ± the SD] in Fig. 14.
3152 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 68, NO. 4, APRIL 2019

According to the theory of statistics, we have that: 1) around ACKNOWLEDGMENT

70% of the travel times fall within this range (i.e., one SD of
The authors would like to thank the anonymous reviewers.
the mean), and 2) a slower driver could likely take a travel time Their valuable comments have significantly enhanced the qual-
within [mean + SD, mean + 2SD], while a faster driver within
ity of this paper.
[mean - 2SD, mean - SD]. We note that since GM does not
consider the behavior of each individual driver, it cannot reflect
the impacts of an individual driver’s driving habit on the required REFERENCES
average travel time. It turns out that in this experiment, the SD is
equal to 0 under GM. Moreover, besides the average estimated [1] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile
edge computing—A key technology towards 5G,” ETSI, Sophia Antipolis,
travel time, our proposed algorithm also leads to an SD similar to France, White Paper no. 11, Sep. 2015.
that of PUAS, which further justifies our algorithm’s estimation [2] A. Abadi, T. Rajabioun, and P. A. Ioannou, “Traffic flow prediction for
performance. road transportation networks with limited traffic data,” IEEE Trans. Intell.
Transp. Syst., vol. 16, no. 2, pp. 653–662, Apr. 2015.
[3] B. L. Smith, B. M. Williams, and R. K. Oswald, “Comparison of paramet-
ric and nonparametric models for traffic flow forecasting,” Transp. Res.
VII. CONCLUSION Part C: Emerg. Technol., vol. 10, no. 4, pp. 303–321, Aug. 2002.
[4] W. Min and L. Wynter, “Real-time road traffic prediction with spatio-
This paper addressed short-term traffic prediction for ACCs temporal correlations,” Transp. Res. Part C: Emerg. Technol., vol. 19,
under the ETSI MEC architecture that exhibits different con- no. 4, pp. 606–616, Aug. 2011.
straints from conventional cloud computing. First, a data-centric [5] E. I. Vlahogianni, M. G. Karlaftis, and J. C. Golias, “Short-term traffic
forecasting: Where we are and where we’re going,” Transp. Res. Part C:
experiment platform was designed and implemented to facili- Emerg. Technol., vol. 43, pp. 3–19, Jun. 2014.
tate traffic prediction algorithm development. Specifically, the [6] M. G. Karlaftis and E. I. Vlahogianni, “Statistical methods versus neural
experiment platform (1) incorporates the free editable Open- networks in transportation research: Differences, similarities and some
insights,” Transp. Res. Part C: Emerg. Technol., vol. 19, no. 3, pp. 387–
StreetMap geographic data, and provides the interface with of- 399, Sep. 2011.
ficial traffic databases; (2) transforms the geographic data into [7] J. W. C. van Lint and C. P. I. J. van Hinsbergen, “Short term traffic and
graph topological representation of the road networks, allowing travel time prediction models,” in Transportation Research Circular, S.
Kikuchi, Ed. Washington, DC, USA: Transp. Res. Board, Nov. 2012,
a user-implemented route planning algorithm to be embedded pp. 22–41.
into the platform more flexibly. This paper further proposed a [8] P. Shang, X. Li, and S. Kamae, “Chaotic analysis of traffic time series,”
novel short-term traffic prediction model that integrates a traffic Chaos, Solitons Fractals, vol. 25, no. 1, pp. 121–128, Jul. 2005.
[9] Y. Zhang, A. Haghani, and X. Zeng, “Component GARCH models to
light model and a vehicle velocity model, considering limited account for seasonal patterns and uncertainties in travel-time prediction,”
computing resources of MEC servers. We noted that the effects IEEE Trans. Intell. Transp. Syst., vol. 16, no. 2, pp. 719–729, Apr. 2015.
of traffic lights are complicated and had not been rigorously ex- [10] L. Li, X. Su, Y. Wang, Y. Lin, Z. Li, and Y. Li, “Robust causal dependence
mining in big data network and its application to traffic flow predictions,”
amined in most, if not all, of the related work. This work modeled Transp. Res. Part C: Emerg. Technol., vol. 58, pp. 292–307, Sep. 2015.
the queueing time when a driver arrives at a road intersection [11] H. Chen and S. Grant-Muller, “Use of sequential learning for short-term
and faces a red light. Moreover, to forecast the vehicle velocity, traffic flow forecasting,” Transp. Res. Part C: Emerg. Technol., vol. 9,
no. 5, pp. 319–336, Oct. 2001.
we proposed a novel low-complexity semi-parametric predic- [12] Y. Wang, M. Papageorgiou, and A. Messmer, “Renaissance—A unified
tion model considering periodic features and spatial/temporal macroscopic model-based approach to real-time freeway network traffic
correlations of dynamic road events. Our vehicle-velocity pre- surveillance,” Transp. Res. Part C: Emerg. Technol., vol. 14, no. 3, pp. 190–
212, Jun. 2006.
diction model retains the advantage of ARMA models’ simple, [13] J. W. C. van Lint, “Online learning solutions for freeway travel time
linear structure while achieving higher accuracy by using much prediction,” IEEE Trans. Intell. Transp. Syst., vol. 9, no. 1, pp. 38–47,
fewer parameters, and, thus, avoiding unnecessary over-fitting. Mar. 2008.
[14] Y.-C. Zhao, Y.-C. Liu, L.-Q. Shan, and B. Zhou, “Dynamic analysis of
Our experiment results indicated the following:
r Our semi-parametric vehicle-velocity prediction model Kalman filter for traffic flow forecasting in sensornets,” Inf. Technol. J.,
vol. 11, no. 10, pp. 1508–1512, 2012.
achieves almost equivalent accuracy to the well-known [15] C.-H. Wu, J.-M. Ho, and D.-T. Lee, “Travel-time prediction with support
vector regression,” IEEE Trans. Intell. Transp. Syst., vol. 5, no. 4, pp. 276–
Long Short-Term Memory Neural Network model, 281, Dec. 2004.
requiring much lower computational complexity by [16] Y. Zhang and Y. Xie, “Forecasting of short-term freeway volume with
considering periodic features. v-support vector machines,” Transp. Res. Rec.: J. Transp. Res. Board,
r The total queueing time at traffic lights accounts for a vol. 2024, pp. 92–99, Aug. 2007.
[17] M. T. Asif et al., “Spatiotemporal patterns in large-scale traffic speed
large portion of the total travel time, which can be more prediction,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 2, pp. 794–804,
than 50% in some cases. This justifies that the effects of Apr. 2014.
[18] J. W. C. van Lint, S. P. Hoogendoorn, and H. J. van Zuylen, “Accurate
traffic lights play a crucial role in road traffic prediction freeway travel time prediction with state-space neural networks under
and cannot be neglected. missing data,” Transp. Res. Part C: Emerg. Technol., vol. 13, no. 5/6,
r We also compared with the most popular route planner, pp. 347–369, Oct.–Dec. 2005.
[19] Q. Ye, W. Y. Szeto, and S. C. Wong, “Short-term traffic speed forecasting
Google Maps. In contrast to Google Maps, our model pays based on data recorded at irregular intervals,” IEEE Trans. Intell. Transp.
attention to the traffic-light effects and the driver behav- Syst., vol. 13, no. 4, pp. 1727–1737, Dec. 2012.
ior, and hence can more effectively capture the real-time [20] X. Ma, Z. Tao, Y. Wang, H. Yu, and Y. Wang, “Long short-term memory
neural network for traffic speed prediction using remote microwave sensor
changes of traffic conditions. This further justifies our al- data,” Transp. Res. Part C: Emerg. Technol., vol. 54, pp. 187–197, May
gorithm’s estimation performance. 2015.
YANG et al.: SHORT-TERM TRAFFIC PREDICTION FOR EDGE COMPUTING-ENHANCED AUTONOMOUS AND CONNECTED CARS 3153

[21] C. Quek, M. Pasquier, and B. B. S. Lim, “Pop-traffic: A novel fuzzy Yu-Ju Su received the B.S. degree in computer sci-
neural approach to road traffic analysis and prediction,” IEEE Trans. Intell. ence and information engineering from the Chang
Transp. Syst., vol. 7, no. 2, pp. 133–146, Jun. 2006. Gung University, Taoyuan, Taiwan, in 2014, and the
[22] L. Dimitriou, T. Tsekeris, and A. Stathopoulos, “Adaptive hybrid fuzzy M.Sc. degree in communications engineering from
rule-based system approach for modeling and predicting urban traffic National Tsing Hua University, Hsinchu, Taiwan, in
flow,” Transp. Res. Part C: Emerg. Technol., vol. 16, no. 5, pp. 554–573, 2016. Since 2017, she has been an Android Media
Oct. 2008. Framework Engineer with MediaTek, Inc., Hsinchu,
[23] J. Tang, F. Liu, Y. Zou, W. Zhang, and Y. Wang, “An improved fuzzy neural Taiwan. Her current research interests include image
network for traffic speed prediction considering periodic characteristic,” compression and framework performance optimiza-
IEEE Trans. Intell. Transp. Syst., vol. 18, no. 9, pp. 2340–2350, Sep. 2017. tion.
[24] W. Zheng, D.-H. Lee, and Q. Shi, “Short-term freeway traffic flow pre-
diction: Bayesian combined neural network approach,” J. Transp. Eng.,
vol. 132, no. 2, pp. 114–121, Feb. 2006.
[25] K. Hamad, M. T. Shourijeh, E. Lee, and A. Faghri, “Near-term travel
speed prediction utilizing Hilbert-Huang transform,” Comput.-Aided Civil
Infrastructure Eng., vol. 24, no. 8, pp. 551–576, Nov. 2009.
[26] J. Tang, H. Wang, Y. Wang, X. Liu, and F. Liu, “Hybrid prediction approach
based on weekly similarities of traffic flow for different temporal scales,”
Transp. Res. Record: J. Transp. Res. Board, vol. 2443, pp. 21–31, Oct.
2014.
[27] W. Zeng and R. L. Church, “Finding shortest paths on real road networks:
The case for a*,” Int. J. Geographical Inf. Sci., vol. 23, no. 4, pp. 531–543, Yao-Yuan Chang received the B.S. degree in com-
2009. munications, navigation, and control engineering
[28] A. Furieri, “SpatiaLite,” 2012. [Online]. https://www.gaia-gis.it/fossil/ from the National Taiwan Ocean University, Keelung,
libspatialite/index. Accessed on: Aug. 13, 2012. Taiwan, in 2017. He is currently working toward
[29] S. M. Ross, Stochastic Processes, 2nd ed. New York, NY, USA: Wiley, the M.Sc. degree with the Institute of Communica-
1996. tions Engineering, National Tsing Hua University,
[30] “Real-time traffic database,” Institute of Transportation, Ministry of Hsinchu, Taiwan. In 2017, he joined the Wireless and
Transportation and Communications, Taiwan. 2018. [Online]. Available: Mobile Network Laboratory, Department of Com-
http://www.iot.gov.tw/ puter Science, Institute of Information Systems and
[31] H.-C. Chang and S.-R. Yang, “Temporal and spatial correlations Applications, and Institute of Communications Engi-
of road traffic data in Taipei city, Taiwan,” National Tsing Hua neering, National Tsing Hua University. His current
Univ., Hsinchu, Taiwan, Tech. Rep., Mar. 2018. [Online]. Available: research interests include multi-access edge computing, machine learning, and
http://www.cs.nthu.edu.tw/ sryang/submission/TSCRTDiTCT.pdf performance modeling.
[32] A. Bjorck, Numerical Methods for Least Squares Problems. Philadelphia,
PA, USA: SIAM, 1996.
[33] T. Vincenty, “Direct and inverse solutions of geodesics on the ellipsoid
with application of nested equations,” Survey Rev., vol. 23, no. 176, pp. 88–
93, 2013.

Shun-Ren Yang (M’13) received the B.S. and M.Sc.

degrees in computer science and information engi-
neering and the Ph.D. degree from the National Chiao
Tung University, Hsinchu, Taiwan, in 1998, 1999, and Hui-Nien Hung received the B.S. degree in mathe-
2004, respectively. From April 2004 to July 2004, he matics from the National Taiwan University, Taipei,
was a Research Assistant with the Department of In- Taiwan, in 1989, the M.S. degree in mathematics from
formation Engineering, The Chinese University of the National Tsing Hua University, Hsinchu, Taiwan,
Hong Kong. Since August 2004, he has been with in 1991, and the Ph.D. degree in statistics from the
the Department of Computer Science, Institute of In- University of Chicago, Chicago, IL, USA, in 1996.
formation Systems and Applications, and Institute of He is currently a Professor with the Institute of Statis-
Communications Engineering, National Tsing Hua tics, National Chiao Tung University. His research in-
University, Hsinchu, Taiwan, where he is currently an Associate Professor. His terests include applied probability, biostatistics, sta-
current research interests include design and analysis of mobile telecommuni- tistical inference, statistical computing, and industrial
cations networks, mobile computing, and performance modeling. statistics.