TorkkolaZhangLiZhangSchreinerGardner(MIRW2007)

Traffic Advisories Based on Route Prediction
Kari Torkkola, Keshu Zhang, Haifeng Li, Harry Zhang, Christopher Schreiner, Mike Gardner
Motorola, Intelligent Systems Lab, Tempe, AZ, USA
Phone: (602)-438-3467
harryzhang@motorola.com
ABSTRACT
Traffic advisories can be delivered to assist travelers in avoiding
congested areas and reaching their destinations in a timely
manner. Instead of relying on a user to manually enter travel
routes and times, we propose a prototype application that learns
from the user's travel history and automatically predicts the user's
destinations and routes. Traffic flow information is then accessed
to determine whether congestions will occur along the predicted
routes to warrant the delivery of a traffic advisory alert. User
acceptance and trust are evaluated in a field study.
Categories and Subject Descriptors
H.5.2 User Interfaces: User-centered design
General Terms
Algorithms, Design, Human Factors.
Keywords
Context aware, user modeling, mobile assistance.
1. INTRODUCTION
Highway congestion has become a way of life in many parts of the
world. This problem stems from several factors. First, the limited
highway capacity simply cannot keep up with the population
growth in many metropolitan areas. In U.S.A., driving is still the
dominant method of transportation, despite the availability of
public transportation. Second, urban sprawling has resulted in a
rush hour nightmare when driving to metropolitan areas from the
suburban areas during the morning commute and when driving in
an opposite direction during the afternoon commute. The use of
the high occupancy vehicle (HOV) lane can mitigate the
congestion problem, but carpooling is not always available or
convenient to everyone. Finally, highway constructions often lead
to lane restrictions, and highway crashes can bring the traffic to a
standstill. In U.S.A., there are over 6.4 million police-reported
highway crashes every year.
Several commercial companies provide traffic flow information
that is gathered by the department of transportation at the state
level, the local police department, and their own helicopter fleet.
Roadway sensors and cameras have been the prevalent method of
acquiring the traffic flow information. The traffic flow
information is made available to the public via radio or TV
broadcast, Internet portals (e.g., www.traffic.com,
maps.yahoo.com, maps.google.com), RSS feeds, email messages,
or SMS messages on mobile phones. Radio or TV broadcast
appeals to a large-scale audience, whereas a user can manually set
up a traffic monitoring service on Internet portals that will send a
custom report (via an email, an SMS message, or an RSS feed)
when the traffic congestion exceeds a threshold for a particular
route during a time window.
A traffic report that is customized for a user is desirable because it
tends to provide information that is more pertinent to the user.
However, the initial manual set-up using Internet portals is
tedious. If travel routes and times are changed, they will need to
be modified manually on the Internet portals. The manual set-up
and modification process is difficult to manage when the number
of travel routes and times grows to accommodate mobile
lifestyles. This process can be simplified or largely bypassed by
using an automatic algorithm that predicts the routes that a user
will take at a particular time. This paper describes one such
algorithm that learns from a user's travel history, predicts likely
routes and destinations using context information such as the
current location and time, determines traffic conditions along the
predicted routes, and delivers timely advisories and alerts to the
user when there are unexpected slowdowns.
Krumm and Horvitz and their colleagues [3,4] have attempted to
predict a user's routes from the user travel history. We adopt a
similar approach to learn a user's usual locations and routes from
the usage history and predict likely routes based on the location
and time information. Unlike Krumm and Horvitz, we do not
attempt to handle destinations that the user has not previously
traveled to. The resulting traffic advisory system is a functional
prototype which uses time and location information in
conjunction with learned patterns of user behavior to determine
when a user is leaving a location for a learned destination. The
application reviews past routes that the user has taken to traverse
from the current location to the predicted destination and checks
available traffic monitoring services to determine if there are any
unusual alerts.
2. DATA COLLECTION
Location and route data was collected from fourteen participants,
representing 6 undergraduate students, 5 office employees, and 3
independents (stay-at-home parents or businessmen such as
realtors who did not work consistently in an office environment).
Each participant was provided with a commercially available
mobile phone equipped with Bluetooth, along with a separate
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
Mobile HCI'07, September 9-12, 2007, Singapore.
Copyright 2007 ACM 978-1-59593-862-6….…$5.00
88

Bluetooth GPS receiver to acquire location data. Data logging
software was installed on the mobile phone. Each participant used
the phone and the GPS receiver for a two to three-month period.
Context data logging software is “ContextPhone”, public domain
software from University of Helsinki, Finland [6]. The logged
context variables consist of the following:
• GPS data from a Bluetooth GPS receiver,
• Current GSM cell ID,
• Bluetooth devices around the phone,
• Phone profile,
• Active phone application,
• Phone idle/active time,
• Battery status and charger status,
• Incoming/Outgoing calls, SMS or multimedia-messages,
• User interaction with the Phonebook and recent call log,
• Media captured with the device (e.g., photo, video).
The data is uploaded over the air to a web server, transparent to
the participants. Logged data consists of events, a time stamp
followed by the event. GPS and Bluetooth environment are
scanned every 30 seconds. Other variables are logged as their
status changes.
Figure 1. GPS coordinates are plotted together with selected
phone application events for a typical participant.
As our focus in this paper is on a user's locations and routes, GPS
coordinates can be plotted together with the use of phone
applications to visualize the roads to which the participant has
traveled and where the mobile phone is used (Figure 1). Figure 1
is cropped to cover only the local metropolitan area. It indicates
that there are a small number of seemingly important locations
where the phone has been used. It also indicates that the user’s
daily commute routes are regular and predictable.
3. ADAPTIVE LOCATION LEARNING
3.1 Adaptation of location clusters
Having alluded to the existence of regularities in the user
behavior data, we describe our experimentation in discovering
those regularities. Based on the collected context data, we employ
machine learning techniques to determine the important locations
where the user spends most of his/her time or most often is
engaged in some activities. Location learning is mainly based on
the collected GPS data. The GPS signal may be uncertain,
however, when the user enters a building, a concrete canyon in an
urban area, or when the GPS device is powered off [7]. To
mitigate this problem, we use cell IDs as a secondary source to
resample GPS data when GPS data loss occurs. The void GPS
data is made up by repeating the last valid GPS reading, if the area
in which a GPS signal disappeared and reappeared, is covered by
the same group of cell IDs.
Although a batch (non-adaptive) clustering method can
automatically identify important locations, it is computationally
expensive and requires processing available data as a whole [3,4].
It cannot adjust to changes in a user's lifestyle pattern. In order to
track changes, adaptive (or incremental/sequential) algorithms are
adopted. An adaptive algorithm is trained with a small amount of
data initially, and adapts continuously as new data is entered.
There are several advantages of using adaptive algorithms over
batch algorithms. Adaptive algorithms are computationally
efficient. They require less storage because a large set of
previously used training data do not need to be kept in memory.
The algorithm parameters can vary continuously to accommodate
lifestyle changes.
For our traffic advisory system, a Gaussian mixture model is used
to identify important locations. Assume for the moment that the
true (but unknown) density of the GPS value tg is of the form
=
=
tN
tttttt ggp
1
),;();( μφϖθ
Where the number of components ∞<tN is known, parameter
vector is },,,,,,{ 111 ttt N
t
N
t
N
ttttt ΣΣ= μϖμϖθ , the nonnegative
mixing coefficients tϖ sum to unity, and ),;( tttg μφ is the
normal probability density function with 2 dimensional mean
tμ and standard deviation t . The goal is to estimate the
parameter vector tθ , which consists of tN3 components.
At time t , the number of Gaussians tN in the mixture model
represents the number of significant location candidates, the
weight of Gaussians tϖ denotes the significance of the learned
locations, and the mean tμ and variance value t in each
Gaussian specifies the important location center and size.
In order to make model parameters adaptive, Titterington [8] has
provided a recursive update formula based on maximum
likelihood estimation. With Priebe’s adaptive mixture estimator
[5], we have designed an adaptive location clustering method
which includes the update rule, creation rule, pruning rule and a
model group switch rule. The details of the adaptation procedure
will be described in a forthcoming paper. The adaptive location
clustering method can recursively adjust not only the location
parameters, such as the center and size, but also the number of
location clusters needed to fit the data. This adaptive mixtures
approach does not require intensive computational power and
storage capacity and produces a consistent small sample estimator.
9th Intl. Conf. on Human Computer Interaction with Mobile Devices and Services (MobileHCI’07) 89

3.2 Evaluation of adaptive location learning
To evaluate the location learning feature in the traffic advisory
system, we compare our adaptive location learning algorithm to a
batch learning method. As the batch method we employ the
DBSCAN (density-based clustering) algorithm, because it can
automatically find the number of clusters, identify outliers, work
well for arbitrary-shaped clusters and efficiently for large datasets
[1]. We first evaluate the algorithms on one office employee who
worked in regular business days at a fixed location. Both batch
and adaptive algorithms are applied to the re-sampled GPS data,
the comparison results for important location discovery are
generated for the first day, the first week, and the first month.
Figure 2 presents the visualization results of the discovered
significant locations after one week.
Figure 2. Location identification after one week
After the first day, the discovered locations by both batch and
adaptive methods match very well. After the first week, compared
with the batch method, the adaptive method discovered location
L0 instead of L1 (Figure 2). Since the adaptive method adjusts the
parameters online, location L1 which was visited early in the
week is already discarded from the significant location candidates
due to the short visit time and low visit frequency (only one visit
for that week). Instead, location L0 which has been visited most
recently had higher initial weight and then becomes more
important than location L1. Meanwhile, both algorithms can
discover several nearby locations in certain large areas, i.e. the
“University” campus and the “Gym”. This comparison shows that
the proposed location learning algorithm is effective in location
identification and can adapt to the user’s daily life.
4. ROUTE LEARNING AND PREDICTION
After learning the important locations, the next step is to learn
routes between a pair of locations. Currently, most transportation
route learning methods are based on geographic information
systems (GIS). Because our system runs on cell phones with
limited computing and storage resources, we instead use raw GPS
data sequences to represent routes [3,4]. A GPS string between
two locations is denoted as a route. The starting point and ending
point of the route involves detecting of the user leaving one
important location boundary and entering into another important
location boundary.
Because travelers often take different routes from location A to
location B, GPS data strings between these two locations may not
represent one single route. In order to determine if two GPS data
strings represent the same route, the data strings must be
compared. If the difference is small (with a MinMax criterion),
then the two GPS strings belong to the same route. Here is the
mathematical formula. Suppose two GPS data strings are
{ }11
2
1
11 ,,, ngggS = and { }22
2
2
12 ,,, mgggS =
The best match of 1S in terms of 2S is
{ }11
2
1
11 ˆ,,ˆ,ˆˆ
ngggS = , where ( )||||minargˆ 121
2
ij
g
i ggg
j
−= .
Similarly, { }22
2
2
12 ˆ,,ˆ,ˆˆ
mgggS = is the best match of 1S in
terms of 1S , where ( )||||minargˆ 212
1
ji
g
j ggg
i
−= .
Two GPS strings are declared to belong to the same route if
( ) 1
11
ˆ||max dgg ii <− and ( ) 2
22
ˆ||max dgg jj <− ,
where 1d and 2d are the largest sampling intervals of R1 and
R2, respectively.
With the adaptive location clustering, the significant locations are
gradually discovered and updated. Once an important location
pair is identified, the commute routes between them are recorded
and learned. Using the preceding comparison process, unique
routes are identified and saved in the database. The unique routes
are non-directional. That is to say, the same route may be traveled
from location A to location B, or from location B to location A.
Frequencies of the unique routes are tallied in an hourly basis.
After two month learning, 47 unique routes are obtained for one
user. The route frequency is derived by considering all different
routes starting from Location A to Location B. Table 1 presents
the frequency of routes from "home" to "office" for one user. Note
that routes 4, 8, and 26 are the most frequently traveled routes.
Table 1. Frequency of travel routes from home to office
Route ID 4 8 25 26 27 28 29 30
Frequency 0.6 0.1429 0.0286 0.1143 0.0286 0.0286 0.0286 0.0286
5. ROUTE PREDICTION-BASED TRAFFIC
ADVISORY SYSTEM
Once routes are predicted, a traffic advisory system can be built.
We designed the system to ensure that the user experience was
unobtrusive and well accepted by the users. Ideally, the system is
designed to assist users during transition times – as the user is
transitioning out of one location and likely into a vehicle
environment. Previous research has shown that messages from
context aware devices are better received during transitions [2],
and in our case it is during transitions that the information
provided by the application is of the greatest relevance.
The application is designed to automatically detect when the user
is leaving a location and predict to where they are likely headed,
and the most likely routes the user will take. When this occurs, a
pop-up window will appear on the visual display informing the
user that that the application has detected the transition and is
searching for traffic information. This will usually occur before
driving. Soft keys will be allocated to allow the user to edit or
cancel this function. No audio or haptic feedback will be provided
90 9th Intl. Conf. on Human Computer Interaction with Mobile Devices and Services (MobileHCI’07)

to alert the user in case the user is not interacting with the mobile
device during this time. To adopt a user-centric approach and
minimize the obtrusiveness, traffic history may be studied to
derive traffic averages and variances and traffic alerts are only
provided when current or predicted traffic flow exceeds the
average by a certain margin. If a traffic alert is detected that is out
of the ordinary for that route and time of day, then the user will
receive an audio or haptic alert, depending upon the phone’s
current profile settings. In addition, a pop up window such as
Figure 3 will appear alerting the user that there is an alert, and
offering the user the choice of ignoring the alert or obtaining more
information. If the user selects the latter, more detailed
information on the alert (e.g. type of incident and location) will be
provided with an option for the application to recommend a new
route to the predicted location. Because of driver distraction
concerns, alerts are issued only when it is necessary to avoid
traffic delays.
Figure 3. Screen shots of the application. The left side depicts a
textual pop-up alert, and the right side a map view of the alert.
The developed prototype is a client-sever system. The client
software is running on a cell phone. The client collects GPS data
and cell IDs and learns semantic locations and routes between
them with our adaptive algorithm. An easy-to-use user interface is
developed to alert users about heavy traffic situations. The server
side includes a set of machines to provide map, real-time traffic,
and other information to clients based on Java EE. The
communication between clients and the server side is based on
standard HTTP protocol.
All specific aspects of the implementation of the design of the
traffic advisory system, as well as the overall acceptance of the
application are validated by real users from specific demographic
areas of interest in a large-scale field evaluation of the application.
6. CONCLUSION
Context knowledge such as the user's needs, locations, and travel
routes is a major enabler for providing the user with unobtrusive
assistance for mobile device. In this paper, the focus has been on a
location-based traffic advisory system that is based on learning
and predicting the user’s patterns using data available from the
mobile devices. We used a data-driven approach in the
development of the traffic advisory system. We collected usage
data from multiple users on phones equipped with sensors over an
extended period of time, mined the data to identify various
lifestyle patterns across different forms of context, and attempted
to predict the patterns. The predicted patterns were used to design
applications that satisfy user needs.
We have demonstrated a successful implementation of a traffic
condition assistant that learns user’s important locations, routes
between those, and learns to predict those routes. In order to
maximize the system effectiveness and user acceptance, advisories
are provided to the user only when there is a real need (e.g., a
traffic problem on the predicted route). We envision that context
aware features like this including a user modeling component will
become essential in supporting new applications and concepts that
connect consumers with media and services through multiple
devices in the office, car, or home.
7. REFERENCES
[1]. Ester, M. et al. A Density-Based Algorithm for Discovering
Clusters in Large Spatial Databases with Noise. KDD, 1996.
[2]. Ho, J., Intille, S. Using Context-Aware Computing to Reduce
the Perceived Burden of Interruptions from Mobile Devices.
CHI 2005. April 2-7, 2005. Portland, OR.
[3]. Krumm, J., Horvitz, E., Predestination: Inferring Destinations
from Partial Trajectories. UbiComp 2006: The Eighth
International Conference on Ubiquitous Computing,
September 17-21 (2006), Orange County, CA.
[4]. Letchner, J., Krumm, J., Horvitz, E. Trip Router with
Individualized Preferences (TRIP): Incorporating
Personalization into Route Planning. Eighteenth Conference
on Innovative Applications of Artificial Intelligence (IAAI-
06), July 2006.
[5]. Priebe, C. E. Adaptive mixtures. Journal of American
Statistical Association, 1994, pp. 796-806.
[6]. Raento, M., Oulasvirta, A., Petit, R., Toivonen, H.,
ContextPhone: A prototyping platform for context-aware
mobile applications. IEEE Pervasive Computing 4:51-59,
2005.
[7]. Rukzio, E., Hamard, J., Noda, C., De Luca, A. Visualization
of uncertainty in context aware mobile applications.
MobileHCI'06, September 12-15, 2006, Helsinki, Finland.
[8]. Titterington, D. M. Recursive parameter estimation using
incomplete data. Journal of Royal Statistical Society, 1984, pp.
257-267.
9th Intl. Conf. on Human Computer Interaction with Mobile Devices and Services (MobileHCI’07) 91

TorkkolaZhangLiZhangSchreinerGardner(MIRW2007)

More Related Content

TorkkolaZhangLiZhangSchreinerGardner(MIRW2007)