CALIFORNIA PATH PROGRAM
INSTITUTE OF TRANSPORTATION STUDIES
UNIVERSITY OF CALIFORNIA, BERKELEY
This work was performed as part of the California PATH Program of
the University of California, in cooperation with the State of California
Business, Transportation, and Housing Agency, Department of Trans-portation,
and the United States Department Transportation, Federal
Highway Administration.
The contents of this report reflect the views of the authors who are
responsible for the facts and the accuracy of the data presented herein.
The contents do not necessarily reflect the official views or policies of
the State of California. This report does not constitute a standard, spec-ification,
or regulation.
ISSN 1055- 1417
December 2008
The Naturalistic Driver Model: Development,
Integration, and Verification of Lane Change
Maneuver, Driver Emergency and Impairment
Modules
California PATH Working Paper
UCB- ITS- PWP- 2008- 8
CALIFORNIA PARTNERS FOR ADVANCED TRANSIT AND HIGHWAYS
Delphine Cody, Swekuang Tan,
Jeff K. Caird, M. Lees, C. Edwards
Final Report for Task Order 5500
1
The Naturalistic Driver Model: Development, Integration, and Verification of Lane
Change Maneuver, Driver Emergency and Impairment Modules
Task Order 5500
June 2008
Delphine Cody, Swekuang Tan ( California PATH Program)
Jeff K. Caird, M. Lees & C. Edwards ( University of Calgary)
2
Abstract
This report documents work conducted in order to support the development of a driver
model. This work consisted of ( i) a review of driver models for identifying the possibility
to add some functionalities to the current model based on existing models and ( ii) the
data collection and analysis in order to describe the mechanism of distraction and
potentially offer some quantification of its effect on drivers’ behavior and performance.
Key- words: driver model, driver’s distraction, driving simulator study
3
Executive Summary
This intermediary report on the project of developing a driver cognitive model addresses
three main issues:
• the development of the model from a theoretical perspective, in terms of
association of cognitive processes into a cognitive architecture
• the implementation choices of the model,
• the gathering of data supporting the development of the model.
In order to address the two first points, the report provides a review of current driver
models with the goal of identifying how these models fit the needs for the development
of a simulation environment which would allow driver behavior modeling, an Intelligent
Transportation System, and observe a simulation of driver behavior using such a system.
The challenges for this type of research are highlighted, especially the implicit work
needed to develop a simulation environment integrating the necessary level of details for
each component of the model, as well as the need to organize research from different
fields, which bring different perspectives for describing the same phenomenon with
different levels of abstraction.
The gathering of data provides quantitative results describing driver behavior that can be
integrated to the model development. An experiment was conducted in order to
investigate of the effect of repeated exposure to a non driving device ( Ipod) on the
driving performance in terms of vehicle control and hazard detection. This study was
carried out on a driving simulator at the University of Calgary. The results provide
insight regarding the effect of repeated exposure on driving performance, although it
could not be established that the effect had reached its plateau. It also provides a new set
of data to use for describing glances ( duration and numbers) while using such a device, as
well as the time required for completing several tasks involving a manipulation of the
Ipod. This section also comprises a method for categorizing lane changes with data
collected during a previous and related project and illustrates how the lane change
decision can be related to the strategic level involved in driving.
4
List of Acronyms
ACME A Common Mental Environment
ACT- R Adaptive Control of Thought-- Rational
ADAS Advanced Driver Assistance System
ANOVA Analysis Of Variance
AOI areas of interest
ASL Applied Science Laboratory
CD Compact- Disc
FLOWSIM Fuzzy LOgic based motorWay SIMulation
HF Human Factors
HMO Head Mounted Optics
ITS Intelligent Transportation Systems
IVIS In- Vehicle Information System
MIXIC Microscopic model for Simulation of Intelligent Cruise Control
MOU Memorandum of Understanding
MP3 MPEG Audio Layer III
MUTCD Manual on Uniform Traffic Control Devices
PATH Partners for Advanced Transit and Highways
PRT Perception Reaction Time
RT Response Time
SDLP standard deviation of lane position
TCT Task Completion Time
TNO Nederlandse Organisatie voor toegepast- natuurwetenschappelijk onderzoek
TO Task Order
TRB Transportation Board Research
UCDS University of Calgary Driving Simulator
VTTI Virginia Tech Transportation Institute
5
Table of contents
Abstract_______________________________________________________________ 2
Executive Summary _____________________________________________________ 3
List of Acronyms _______________________________________________________ 4
1 Introduction _______________________________________________________ 6
2 Review of driver models ______________________________________________ 7
2.1 Driver models architecture and implementation___________________________ 8
2.2 Models calibration and validation______________________________________ 14
2.3 Synthesis of presented models _________________________________________ 16
3 Data supporting the development of the model___________________________ 20
3.1 Investigation of drivers’ distraction ____________________________________ 20
3.1.1 Background______________________________________________________________ 20
3.1.2 Methods_________________________________________________________________ 22
3.1.2.1 Participants _________________________________________________________ 22
3.1.3 Apparatus and Materials ____________________________________________________ 23
3.1.3.1 The University of Calgary Driving Simulator ( UCDS) _______________________ 23
3.1.3.2 ASL- 501 Eye Tracking System. _________________________________________ 23
3.1.4 Procedure _______________________________________________________________ 24
3.1.4.1 Secondary Tasks _____________________________________________________ 24
3.1.4.2 Experimental Sessions_________________________________________________ 25
3.1.4.3 Experimental Drive Descriptions ________________________________________ 26
3.1.5 Results and results discussion________________________________________________ 26
3.1.5.1 Participants _________________________________________________________ 26
3.1.5.2 Experimental Design __________________________________________________ 27
3.1.5.3 Discussion __________________________________________________________ 27
3.1.5.3.1 Hazard Detection __________________________________________________ 28
3.1.5.3.2 Prolonged Experience_______________________________________________ 29
3.1.5.3.3 Adaptation _______________________________________________________ 30
3.1.5.3.4 Distraction Metrics _________________________________________________ 30
3.2 Naturalistic driving data mining _______________________________________ 32
4 Conclusion _______________________________________________________ 37
5 References _______________________________________________________ 39
Appendix – Detailed Results from driving simulator experiment ________________ 44
6
1 Introduction
From a general standpoint, generating driver models can be seen as equivalent to
developing a comprehensive description of scientific knowledge about drivers. In making
a hierarchy of scientific methods to approach a phenomenon such as the driving activity,
a first step is to conduct specific studies to observe and understand the phenomena ( e. g.
drivers’ glances, perception of speed, decision making process, drivers’ attention control)
and the second is to reconcile the results of the different studies into a comprehensive
description, in other words, a model dynamically linking these results of these different
studies to reproduce the driving activity. Hence, one of the interests of building driver
models is to validate the results of studies looking into specific aspects of driving and
identify the aspects that need to be addressed to further the understanding of driving.
From a practical standpoint, the development of the model consists of the following
phases:
• identification of a relevant architecture or theoretical framework explaining
human information processing,
• identification of relevant models of performance developed about driver behavior,
identification of a relevant world simulation model, i. e., description of
infrastructures ( e. g. highway with entrances, merges, split, intersections, traffic
lights), traffic and/ or vehicles characteristics, and integration of these framework
onto a simulation environment.
Until now, the work conducted at PATH ( Partners for Advanced Transit and Highways)
on the development of a cognitive driver model ( Memorandum of Understanding ( MOU)
369, Task Order ( TO) 4222 and TO 4238) had identified a relevant architecture, some
models of performance and perception, and relied on an in- house simulation tool,
SmartAHS, as a simulation environment, this environment came with a vehicle model,
sensors and communication models described at a relatively low level of details. As
SmartAHS was developed in order to simulate automated highway, its representation of
the world corresponds to what an automaton would see, and as such, is quite different
from what a human would see. As the world’s description is deeply rooted into
SmartAHS, we faced the choice of either redeveloping a new world using SHIFT, the
environment used for developing SmartAHS, or use other environment.
In order to make this decision, we set to explore several other driver models that are
implemented or in the process of being implemented, in order to identify some potential
redundancy with our effort. This will represent the first part of this report.
In a second part, we will present the continuation of our work on gathering data about
driver behavior, with data collection distraction, research led at the University of Calgary
by Professor Caird and his team, and a method of analysis for describing lane change
behavior.
7
2 Review of driver models
This review is mostly a presentation of the models presented and debated at a workshop
for the Transportation Board Research ( TRB) conference in January 2005 on the theme
of driver models, co- organized by D. Cody and T. Gordon1. The aim of the workshop
was to bring together members from the various disciplines - Vehicle and Traffic
Engineering, Psychology, Human Factors ( HF), Artificial Intelligence to mention the
most common – involving driver models. The definition of the term varies betweens
disciplines, and even between different authors within any given discipline. Recent
efforts in Applied Psychology and Human factors emphasize the interest of developing
models that can be implemented and used in computer simulation, hence representing a
possible link between these disciplines, and also a chance to consider the broader picture
of driver model within a transportation/ traffic system.
The need for developing such models resides in their application. We focus on two types
of applications: i) application to safety and driver assistance system design and traffic
assessment, and ii) prediction and evaluation of Intelligent Traffic Systems ( ITS) on
traffic flow. A major difference between these two types of application is mostly a matter
of scale, most of the work conducted on the first one is developed at the “ driver” level,
i. e., the model focus on simulating one given driver, while for the second type of
application, the model focus on modeling flows of drivers. This difference of scale leads
to differences in requirements and levels of details for each type of models.
This section focuses on the models that were presented and debated during the workshop.
Three models fall under the application to safety and driver assistance systems, namely,
the “ In- Vehicle Information System ( IVIS) model”, the ACT- R2 driver model and the
“ workload in driver modelling”. Two other Models, ACME3 – Driver model, and the
Fuzzy LOgic based motorWay SIMulation ( FLOWSIM), address the traffic assessment
category.
The purpose of IVIS is to provide a proof of concept design tool to assess and compare
in- vehicle system concepts, prototypes and products for system designers and to provide
insight into potential design improvements. The approach for the development of this
model consists of compiling, analyzing and using data from actual empirical research for
creating “ micro- models” of performance prediction. The model predicts the visual and
cognitive resources needed at a basic level. These predictions are made at the subtask
level, because subtasks often require different resources, and subtasks can be combined
into tasks which can themselves be combined into system assessment. Therefore, the
model accounts for a variety of information input, information processing and response
output combinations.
1 This workshop led to a presentation ( Cody and Gordon 2005) at the workshop on “ Modelling Driver
Behaviour in Automotive Environments” organized by Network of Excellence HUMANIST, sponsored
by the European Commission, General Directory INFSO, within the 6th Framework Program of R& D.
2 Adaptive Control of Thought— Rational
3 A Common Mental Environment
8
The theoretical endeavour of the ACT- R driver model is to provide a psychologically
plausible model of an individual driver and combine cognitive, perceptual and motor
dimensions. The approach consists of applying a cognitive architecture, ACT- R
associated to a computational framework. The advantage of this approach is the
possibility to re- use theories and mechanisms already integrated within the cognitive
architecture. This approach currently focuses on highway driving involving moderate
traffic. The two current practical applications are: i) the prediction of driver distraction,
for which models of typical secondary tasks ( e. g. phone dialling) are integrated to the
model in order to predict real- world observables measures and ii) the recognition of
driver intentions, where many models are run simultaneously, each trying to accomplish a
different goal and the method consist of tracking which model best matches the observed
data.
The driver modelling effort at TNO aims at evaluating driver behaviour, performance and
workload and at providing inputs for traffic flow model, such as MIXIC driver model and
Human- Kinetic traffic flow model. This approach focuses on individual driver and
vehicle units and is based on experimental research. The framework applied for this
modelling effort is the one of “ optimal control”, i. e. the use of a linear system theory,
where the assumption is that the driver is well trained and well motivated to behave
optimally. The model also integrates inherent limitations and constraints. The resulting
model is a realistic description of driver behaviour in terms of driver performance,
workload and total system performance measures.
The purpose of the ACME driver model it to develop a combined view on car- following
and lane- changing as well as evaluate the mental processes needed for driving and
describe them. It is a man in the loop simulation of a “ car- driver” unit composed of three
sub- models: model of human sensors, model for information processing and a model for
action execution. The simulation integrates models of the dynamics of the vehicle a
driver uses and models of the simulated area.
FLOWSIM development started in 1997 and was initially focused on highway traffic. It
now covers all type of roadway. This framework was first used for Advanced Driver
Assistance System ( ADAS) and infrastructure speed control applications. Recent urban
applications include network travel time prediction. It is now the object of an intensive
enhancement program with China and has been chosen as the traffic simulation tool for
supporting the city of Tianjin transportation planning and management.
2.1 Driver models architecture and implementation
Model architecture can be seen as the blue print of the model. In that sense, a simple
description of a model ( see Fig. 1) consists of three elements: inputs, information
processing or behaviour and outputs. This simple architecture allows generating
parameters of factors that can be used to derive measures of effectiveness or evaluation.
In this section, we will expose briefly the architectures that were presented for each of the
9
model and refer the reader to the speakers’ publications on their model for more thorough
description. Also, each model will be described in terms of: simulation scale, along two
dimensions: the unit simulated: one driver, groups of drivers or vehicle fleets and the
number of unit that can be simulated at once; visualization, in terms of drivers’ states,
parameters that can be observed in simulation and traffic; and finally the ease of use of
their model by other researchers will be addressed.
Figure 1: Modular high level representation of driver model structure
The overview of the IVIS demand model is presented in Fig. 2 below. During the
workshop, the accent was put on the software development of the tool and how user
would manipulate the tool. Hence, the component the most closely related to the
information processing from Fig. 1 is in the top block with the list of Driver Resources
Involved, i. e.: visual demand, auditory demand, manual demand and speech demand. The
assumption of the model is that the driver is a finite capacity has a single channel
processor of visual information. He/ she can “ share” cognitive resources to differing
degrees and the resources needed to perform IVIS tasks compete with the resources
needed to drive the vehicle. In order to avoid reducing the driving performance to
unacceptable levels, only limited resources can be required by IVIS. It is considered that
different resource components and/ or magnitudes can be used to perform the same task
and are required for different tasks. The combination of the demand of these specific
components determines the required resources to perform a task. The required resource
demand can be used to estimate the potential of a decrement in driving performance. The
principle of IVIS is that nominal values for measures are derived that can be modified to
match a task or design specification, for example, a subtask modifier is the message
length and a task modifier is the roadway complexity.
10
Figure 2: Overview of IVIS Demand Model
The scale of simulation is one driver within one of three age groups. The software allows
visualizing the behaviour change outcome and the modifiers that can be selected to
adjust. Finally, the source code and a user manual are available from FHWA. The
software is still in a proof of concept stage.
Fig. 3 and 4 contain diagrams describing ACT- R and the ACT- R driver model. The
diagram on Fig. 3 is an ACT- R architecture and details the overlap between brain
structures and information processing steps. Fig. 4 describes the component involved in
the ACT- R driver model. The approach used here is that ACT- R provides a framework
and specific rules are developed for the driving model. The driver model also integrates
cognitive models for other tasks, such as using a cell phone or a navigation system. The
current scale of simulation for the ACT- R driver model is focused on one driver in an
environment that provides interactions with other vehicles. This effort includes also the
modelling of younger versus older drivers. The visualization allows observing driver’s
eye view and mirror and in a new system currently under development, graphs will
provide measure of behaviour. On the ease of use topic, the current version of the ACT- R
driver model cannot be used easily by other researchers, but here also, a new version is
under development that should allow easier use of the model ( Salvucci in press; Salvucci
and Lee, 2003).
11
Figure 3: ACT- R architecture
Figure 4: ACT- R driver model representation
12
In the work on incorporating workload in driver modelling for ITS ( Intelligent
Transportation Systems) applications, two figures are also used for describing the scope
of the model. In Fig. 5, the concept of task covers elements such as lane keeping, car
following or speed control, while the behaviour is described as measures of performance
and workload. Fig. 6 illustrates a 3x3x3 description of driver behaviour ( Theeuwes,
2001) relative to the task hierarchy, task performance and information processing. In dark
is the part of the behaviour that is currently modelled, i. e. the control of the vehicle at a
skilled based level. For this part of the behaviour, all of the information process is
integrated, from perception to action. The presentation of the right hand side figure
provoked a lot of interest from the audience, as it is commonly considered that strategic
equals knowledge based, manoeuvring equals rule based level and control equals skill
based level. The questions that was raised was about how to transition within this 3x3x3
representation, for one driver, for example how can a driver be at the skill based control
and then go into the knowledge based level, is it only by learning and can it also be due to
other factors? The example of degraded driving conditions was given to explain that even
an experience driver could be in the knowledge based levels.
Figure 5: Block diagram driver vehicle system
Figure 6: TNO’s driver model
Fig. 1.
13
Figure 7: ACME driver model
The ACME driver model is a very modular architecture ( Fig. 7) and one of the most
driver vehicle infrastructure system oriented among the presentations given. The model
can be decomposed in three major substructures: Senses, information processing and
actions execution. The senses that are implemented are vision and hearing, the perception
of acceleration is used to influence speed decision and the haptic input is not integrated.
The information processor consists of an internal world representation storing, a planning
instance and an execution instance. In order to carry out the execution of action, the
extremities are simulated to move in- vehicles devices to certain position. These devices
in turn determine the vehicles dynamic. Regarding the scale of simulation, approximately
20 vehicles can be simulated around an intersection. The simulation step ranges from 10
to 100 ms. The visualization represents the states within the driver’s cognition and
includes timelines of measures. In order to use the model, it would still require time from
the user to become familiar with the model and its implementation.
14
Figure 8: FLOWSIM driver model
For the FLOWSIM driver model ( Brackstone 2000; Wu et al. 2000; Bracstone et al.
1997), the representation of the model ( Fig. 8) presents the three basic steps of watching,
thinking and responding and how these basic steps can be associated to fuzzy logic. The
example used for the description of the model focused on speed and gap control. It
showed how the relative speed and distance divergence influence the drivers’ action in
terms of acceleration rate. This model is the one that allows the most units to be
simulated during one simulation, with the possibility of simulating up to a 1000 vehicles.
Different behaviours are simulated using distributions. The speaker found little empirical
evidence about the existence of groups such as young or old or aggressive drivers and
therefore uses the distribution without labels tying the driver/ vehicle unit to specific
groups. The visualization displays mainly the individual vehicles and the road
geometries. In order to use the system, it would be necessary to receive training.
2.2 Models calibration and validation
The model calibration and validation provoked interesting conversation and a notable
question from the audience was on the difference between calibration and validation.
This question was also treated differently based on the different approaches applied by
the speakers. The issue of calibration is interesting in regard of the aspect of the model
that is considered. For example, in a model such as ACT- R, the question is not as much
calibration of the timing of information processing steps as inferences built on ACT- R on
the global time it takes to process information. Therefore, the calibration aspect that was
covered during the workshop was based on behaviour that can be identified, observed and
measured.
We distinguished two main methods for calibrating models. The first one consists of first
using data and results available in the scientific community. The second method consists
15
of running specific data collection. This method can be subdivided in three more
categories, depending if the data collected is driver centric ( gaze, attention) or vehicle
centric ( steering, braking, speed control) or traffic centric ( lane keeping, range, range
rate). The use of the first method brings up the issue of standardization in method to
gather and measure data and the definition of parameters that are derived.
On the topic of model validation, there is one main method of comparing the model
output with data sets. The variation in the application of the method depends on the type
of data used. For instance, the model output can be compared with collected data, with
on- going behaviour, or with more general data. The point of comparison of model output
with on- going behaviour brought up the very interesting question of the nature of the
model output. Is a model providing a trend or should we expect to be able to predict real
time behaviour? The question was not answered directly during the workshop, although
some researchers present in the audience thought that a prediction of driver behaviour in
real time would be a huge achievement for driver models, a lot of doubt about the
possibility to reach such a level of accuracy was expressed.
Clearly, there is fundamental issue resulting from the inherently stochastic nature of
driving. On the one hand, a ‘ real- time’ model should be capable of making precise and
detailed predictions of driver actions as a function of time. On the other hand, real drivers
display a range of decisions, driving styles, levels of attention etc., and so should any
realistic driving model; resolving these apparently conflicting viewpoints requires a
somewhat deeper concept of what is required for model validation. At the very least,
validation of real- time models must take account for a wide range of stochastic
influences.
Out of the five models that were presented, four are proceeding with more or less
extensive validation and one is still in the process of calibration. The IVIS model is at the
level of a proof of concept. It is also important to note that the model is used and
designed to be used in the industry which can make the feedback to the scientific
community a slower process. For a review of the validation carried out at the Virginia
Tech Transportation Institute ( VTTI), the reader can consult Jackson and Bhise ( 2002).
The approach applied for the ACT- R driver model is the comparison of model and human
data sets recorded similarly in terms of real- world measures. The data sets that have been
collected so far are highway driving, phone dialling and distraction, radio tuning and
distraction and the age effects. Examples of measures are lane change steering profiles,
gaze distribution on highway, lateral deviation during phone dialling. The outputs
generated by the OCM are time simulation results, performance and workload measures
and a workload index. The example of validation provided during the workshop was a
comparison of data collected on a driving simulator and of model prediction on the lane
keeping task. The agreement between the model and experimental results indicates a
useful predictive capability. The ACME model is still on the process of development
hence has not been validated yet. The validation of FLOWSIM ( Wu et al. 2003)
concentrated on the comparison of simulated and measured traffic flows and showed a
very satisfying fit.
16
2.3 Synthesis of presented models
We consider that the basic architecture of a simulation involving a driver model is
constituted of at least three main components: a driver model, a vehicle model and an
infrastructure model. This hierarchy ( Fig. 9) below can be further resolved into increasing
levels of detail. For example, the driver model can be expanded in more modules or
dimensions, such as driving tasks classification ( strategic, tactical and operational),
psycho- motor dimension, which is constituted of perception, cognition and motor control.
For the purpose of the workshop, we defined a number of dimensions for each of these
three main components and ask the speakers to rate their models on these dimensions.
Figure 9: Hierarchy of model components
For the Driver model component, the proposed list of dimensions is the following:
psycho- motor
driving tasks
control level
model capacities
simulated phenomena
simulation control
For the vehicle model component, the dimensions were:
control variables
vehicle motions
vehicle subsystems
vehicle response
17
Finally, for the infrastructure, the two main categories that were proposed are:
level of details ( lane, signalization),
environment type ( highway, urban ….).
As the workshop was limited in time, we focused the scope of the description to the three
main components and on the details of the psycho- motor dimensions. The rating was as
follows:
0: not represented
1: indirectly represented ( there is a related parameter)
2: basic inclusion ( there is a specific parameter in the model that indicates relevant
trends)
3: included ( the parameter or state is directly represented via a simple sub- model)
4: modelled ( a sub- model represents the process)
5: represented in detail ( the process is a core aspect of the model, and be related in some
detail to experiments or theories)
This rating was then integrated in a “ radar web” graph. Each speaker presented for a
graph for each of the three components and the psycho motor level. Then, all of the
models ratings were integrated on a same graph as illustrated in Fig. 10 below.
Figure 10: Representation of driver model dimensions
18
The value of representing the models presented at the workshop via a common graph is to
identify synergies and limitations. In order to categorise synergistic dimensions, it was
proposed that where a dimension was ranked three or above by at least three models, then
a synergy between the three models is possible. In Fig. 10, the dimensions that fit this
category are: perception, cognition, operational, rule, reaction and static input. A second
category is one where dimensions could clearly be further developed, where at least one
model ranks 3 or above. The dimensions in this category are: motor control, strategic,
tactical, skill, non driving activities, knowledge, anticipation, distraction and driver
characteristics. Another interesting category is when none of the model ranked three or
above, with three dimensions: learning, variable and dynamic inputs.
Figure 11: Representation of psycho- motor dimensions
The same method was applied to the description of the psycho- motor dimensions ( Fig.
11). Here, the areas for which we can identify synergies are: vision, decision making,
memories, steering control and velocity control. The areas to develop are: haptic,
auditory, recognition, anticipation, rules/ knowledge and attentional resources. Finally, the
areas to cover are tactile and other vehicle control.
A question about the high and low ranking dimensions covered by the models is what
dictates the choice of what to implement first? Is it because these dimensions are easier to
implement than the low ranking or because they are more important to the concept of
driver model and its current applications?
19
20
3 Data supporting the development of the model
The sets of data presented below illustrate the complementarities between inputs brought
by controlled experiment results and behavior observation in naturalistic settings. This
section is composed of two main parts. First, a study on drivers’ distraction is presented,
with a discussion of the metrics that can be extracted from this experiment. Secondly, a
description of driver behavior is provided, as well as how these observations can support
the development of the driver model.
3.1 Investigation of drivers’ distraction
3.1.1 Background
Driver distraction from a variety of in- vehicle sources, including car radios, has been
cited since the 1930’ s as potential crash contributors ( Caird & Dewar, in press; Goodman
et al., 1997). In particular, the progression of in- vehicle audio entertainment systems has
included radio, 8- track, cassette, CD, and now MP3 players, which is the focus of this
study. The overwhelming majority of drivers use music systems while driving. Almost 92
percent of drivers were observed, in a naturalistic study by Stutts et al. ( 2005), using
audio/ music devices while driving their own vehicles. Interaction with these systems has
resulted in crashes. Music system use ( i. e., adjusting the radio/ cassette/ CD) was a
contributor to about 11 percent of all distraction crashes compared to 1.7 percent of
crashes when using a cell phone ( i. e., talking/ listening/ dialing) ( Stutts et al., 2001).
Interacting specifically with CD players was associated with a several fold increase in
crash risk ( Klauer et al. 2006). Cell phone use while driving, which has been the focus of
the majority of research on driver distraction, has been associated with approximately a
four- fold increase in crash risk ( Laberge– Nadeau et al., 2003; McEvoy et al., 2006;
Redelmeier & Tibshirani, 1997).
Relatively few studies have systematically examined the impact of more recent
technologies on driver performance, including MP3 players. For example, the effects of
email ( Jamson et al., 2004; Lee et al., 2001), text messaging ( Hosking et al., 2005) and
MP3 players ( Donmez et al., 2005) on driver performance are less common than studies
of cell phones and driving ( Caird et al., 2004; Horrey & Wickens, 2006). Performance
decrements associated with distractive tasks have been found in significant increases in
reaction time during cell phone tasks ( Ålm and Nilsson, 1995; Ålm & Nilsson, 1994;
Brookhuis et al., 1991; Strayer et al, 2004) and speech based e- mail interactions ( Jamson
et al, 2004). Lateral vehicle control, as indicated by measures such as lane positioning,
and steering, has not consistently been affected by cell phone conversation ( Brown et al.,
1969; Parkes and Hööijmeijer, 2001) or speech based email ( Jamson et. al., 2004). In
contrast, text messaging while driving has been found to significantly increase the
number of lane excursions observed ( Hosking, et al., 2005). Additional research on more
recent technologies that are brought into the vehicle by the driver is needed.
21
To determine the allocation of attention to specific sources within and outside the vehicle
requires the use of eye movement measures. Previous research has shown that drivers
reduce or constrain the breadth of eye movements in the presence of a distractor
( Chisholm et al., 2006; Green, 1999a, Recarte & Nunes, 2000; 2003) such as a cell
phone. However, while interacting with a visual secondary device, increased visual
sampling into the vehicle is required and many eye movement behaviors. Hoskings et al.,
( 2005) found a 400% increase in eyes off road time while sending or receiving text
messages in a simulator. From a practical point of view, eyes off the roadway and the
frequency of glances to a device have been used to suggest that eye movements that are
too long or too frequent are unsafe ( Green & Shah, 2004). When the eyes are focused into
the vehicle the probability of missing critical external events while driving increases,
thereby increasing the potential for collisions ( Klauer et al., 2006).
Most studies that have addressed the impact of in- vehicle tasks have used a cross
sectional design. With repeated exposures to a task, a variety of adaptations may occur
that are not appreciated in cross sectional designs ( Caird & Dewar, in press). For
example, task sharing may become more efficient and less likely to produce common
errors ( e. g., missed detections). By allowing drivers to become accustomed to both the
vehicle and the 5 secondary tasks, as well as having sufficient time to acquire multi-tasking
skills, a more accurate picture of distraction effects will result. Only one other
study has chosen this approach. Shinar et al., ( 2005) studied cellular telephone distraction
over five simulator sessions, to see if multiple sessions improved vehicle control while
drivers were engaged in a conversation task. The results of their study indicated that
some improvement over sessions occurs, lessening the impact of cell phone conversation
on driver performance.
To address this limitation and the lack of empirical studies on MP3 player interaction, the
present study examined the effects of iPod interactions over multiple sessions to
determine the effects on driver performance to both event- based responses and vehicle
control measures. Participant practice over multiple sessions was used to determine
whether the driver became more efficient, less error prone, and more aware of the risks
associated with using a device. The addition of the distraction task was expected to cause
decrements on several measures of driving performance relative to baseline. Specifically,
degrading perception response times to critical hazards, increasing the number of
collisions observed, more so for the difficult iPod interactions. Whereas the easy iPod
interactions would be executed with little trouble and be less distracting than the difficult
interactions while driving. The visual complexity of the difficult iPod task was expected
to increase the frequency and duration of fixations made into the vehicle to complete the
task. Over the sessions participants were expected to become more efficient in their
interactions with the iPod, leading to a decrease in task completion time. Some increased
efficiency in task sharing was expected over sessions, which will lead to decreases in
perception response time and the number of collisions as performance improved.
Although improvement equal to or beyond that of the baseline measures was not
expected to occur.
22
3.1.2 Methods
3.1.2.1 Participants
Nineteen participants ( 10 Females, 9 Males) between 18 and 22 years of age were
recruited from the University of Calgary and surrounding community to participate in the
study. Participants were asked to volunteer for a preliminary screening session plus six
simulator sessions conducted weekly over a two- month period. Remuneration for the
sessions increased incrementally throughout the six sessions, each participant received a
total of $ 200.00 ($ CAN) for the successful completion all seven sessions. A prize draw
was also conducted for those who completed all of the sessions.
All participants were required to hold a valid class 5 driver’s license, not be in the
graduated licensing program and drive a minimum of 10 000 kilometers per year. All
participants were also required to be in good physical and mental health, free of any
visual and hearing disorders, and not be under the influence of medications or drugs that
would affect their driving or cognitive performance. Demographic information and
driving experience information was collected via the Driver Experience Questionnaire.
Due to the difficulties of calibrating the eye movement system, those who wore glasses
while driving were not permitted to participate in the study. Those with contact lenses
were allowed into the study.
Visual testing was performed using a number of tests to ensure participants met the
minimum acuity requirements mandated by law for licensure, which is 20/ 40 acuity in
Alberta ( Casson & Recette, 2000). Corrected visual acuity was tested for both long and
short distances. Specifically, long distance acuity was tested monocularly at 6.1 m ( 20ft)
with the Snellen Visual Acuity eye chart. A short distance acuity test was performed
binocularly at a distance of 70 cm with the Landolt C test to determine near distance
visual problems that could affect viewing of the iPod device during testing. Contrast
sensitivity was measured at 3 m with the Vistech Contrast Sensitivity Chart. The Ishihara
Test for Color Blindness ( plates no. 3 and 27) was used to test colour vision ( Ishihara,
1993). Those who did not meet the minimum requirements for visual acuity ( 20/ 40),
contrast sensitivity, or had colour vision deficits were not allowed to participate in the
study ( N = 2). All others were then screened for simulator sickness susceptibility with the
Simulator Sickness Questionnaire ( Kennedy et al., 1993) and provided information
regarding their in- vehicle device use information was collected with the Device Use
Questionnaire.
During the initial screening and testing session, participants also did a practice drive,
which lasted about 10 minutes, and included stoplights, turns, curves, hills and speed
changes. The purpose of the drive was to allow drivers to become familiar with the
simulator and to screen out participants who were susceptible to simulator sickness ( N =
3).
23
3.1.3 Apparatus and Materials
3.1.3.1 The University of Calgary Driving Simulator ( UCDS)
The following brief description of the UCDS and eye movement system is abridged from
Caird et al. ( 2007). The UCDS is composed of 5 networked Dell computers. A D- Link
KVM ( Keyboard Video Mouse) switch handles the network communication between the
three 3.6 GHz visual channels with a NVIDIA GeForce6600GT graphics cards linked to
the three projectors. Each Epson 703C projector displays ( 1064 x 768 resolution and
maximum lumens 1000) onto a WrapAround Clarion Screen by Draper each of which
measures 86.5” wide by 65” in height and are approximately 230 cm from the drivers
head position. The display refresh rate of the system is 60 Hz. The total projected forward
field- of- view from the drivers seated position is 150 degrees.
Traffic environments and experimental scenarios for the driving simulator are developed
and run in HyperDrive ™ ( v. 1.9.25). Tiles can be selected from an extensive pallet of
intersections, freeway sections, streets, and so forth, all of which adhere to the Manual on
Uniform Traffic Control Devices ( MUTCD). The placement of dynamic objects, such as
vehicles and pedestrians, require iterative testing and development using a variety of
Tcl/ Tk scripts. HyperDrive ™ also manages the data collection of a number of driving
variables such as perception response time ( PRT), velocity variability, minimum
headway, and lane positioning.
The purpose of the SimObserver and microphone audio systems is to record participant
and experimenter activities and to facilitate communication among the same by
integrating multiple visual and auditory inputs into a single display. A system of three
black and white “ lipstick” cameras ( Model No. KPC 500) are mounted inside the Saturn
and provide views of the driver’s face, hands on the steering wheel, and feet on the brake
and accelerator. A fourth colour camera ( Panasonic Model No. WV- CP460 with a
WVLA 9C3A 9mm Panasonic lens) shows the centre screen of the simulated traffic
environment. Data was transferred to a 3.0Ghz Dell XPS 600 Intel Pentium D desktop
computer, with two 232GB removable hard drives. Video analysis is then performed
using Data Distillery, which is an offline data review and reduction analysis program.
3.1.3.2 ASL- 501 Eye Tracking System.
Eye movements were captured during half of the experimental sessions ( i. e., sessions 2,
4, and 6) using an Applied Science Laboratory ( ASL) 501 eye tracking system. The ASL-
501 uses a lightweight, head- mounted, infrared corneal reflection system that allows data
collection while head and body movements occur. The illumination beam and image of
the eye is reflected off of a headband- mounted monocle located just below the
participants’ left eye. The illuminator, eye camera, scene camera, and monocle are
integrated by the Head Mounted Optics Module or HMO. Eye position is sampled at a
rate of 60Hz ( Applied Sciences Laboratory, 2001).
24
3.1.4 Procedure
3.1.4.1 Secondary Tasks
During each session, the participants had the opportunity to interact with the 20GB Apple
iPod with a 2.5- inch colour LCD display screen while driving ( see Figure 12). The iPod
was placed in a holder and mounted on the center console of the Saturn. Music was
played through a portable speaker system ( i. e., JBL on tour), which was placed on the
centre of the dashboard. All iPod task instructions were presented in green writing on the
center screen. Both easy and difficult iPod tasks were performed during the sessions.
Easy tasks were defined as having one or two steps, represent common tasks ( i. e.,
achieved frequent goals), and took less than five seconds to accomplish when tested
alone. These included turning off the iPod, pausing it, and skipping ahead a couple of
songs. Difficult tasks required five to seven steps, are used to accomplish more complex
or specific tasks, and took about 20 to 30 seconds to complete when tested alone.
Difficult interactions required participants to turn on the iPod and find a specific song in
the song titles menu. A total of 900 songs were programmed into the iPod for use in the
study. The position of selected songs in the song list was varied. Time to complete the
task as a function of song list location was recorded. In total three pairs of iPod
interactions were performed in each session ( 3 easy and 3 difficult tasks).
Figure 12: The 20 GB iPod shuffle mounted on the center console in the simulator
25
3.1.4.2 Experimental Sessions
The first experimental session included training and benchmark testing on the iPod
requiring participants to perform both easy and difficult tasks. The benchmark testing
was performed at the beginning of both the first and sixth experimental sessions to
determine whether changes in secondary task performance occurred as a function of
practice over sessions. Time to complete each task was recorded in seconds using a
stopwatch. During each of the six experimental sessions, participants drove a total of
three drives. The first was a practice drive to familiarize participants each week with the
handling characteristics of the simulator. The second and third drives were
counterbalanced for presentation and included a drive interacting with the iPod and a
baseline drive with no secondary task. Each drive lasted approximately 12 minutes.
Participants were scheduled to return each week at the same time of day.
Over the course of the sessions participants were required to respond to four different
events during iPod interactions as well as baseline drives ( see Figure 13). The first event
involved a pedestrian who emerges from between two parked cars and “ walks” into the
path of the driver on the road. The second event involved a parked vehicle that pulled out
from the side of the roadway into the path of the participant. The third event was a lead
vehicle traveling 1.5 s ( approx. 40 m) in front of the participant. At a specified location, it
decelerated at a rate of 8 ms2 for 75 meters before accelerating at a rate of 4 ms2 and
resuming its original speed. The first three events required braking, steering, or a
combination therein to avoid a collision and have been developed and used previously
( Chisholm et al., 2006). The last event required participants to respond to a late yellow
light that was initiated 2.69 seconds prior to the intersection stop line ( Caird et al., in
press). A total of three occurrences of each event type were encountered within each of
the iPod easy, difficult iPod, and baseline secondary task conditions. All events were
counterbalanced randomly across the six experimental sessions to reduce anticipation on
the part of the driver.
26
Figure 13: Pedestrian event in residential roadways ( top left) and pullout event in urban roadways
( top right). Lead vehicle braking event ( bottom left) and light changing intersection ( bottom right)
3.1.4.3 Experimental Drive Descriptions
Eighteen experimental drives were created and used during the six experimental sessions.
For the pedestrian and pullout events, 50 km/ hr urban or residential roads that had two
lanes of traffic were used. Parked cars and commercial buildings lined the urban route
and parked cars and single- family homes with attached garages lined the residential
routes. The late yellow light event occurred on 70 km/ hr suburban roadways with two
lanes in each direction, commercial buildings, and parked cars. Last, the lead vehicle
braking events occurred on the 100 km/ hr six- lane freeways, with three lanes in each
direction, separated by a grassy median. The volume of ambient traffic encountered in the
scenarios varied depending on road type, and consisted of a mix of cars, trucks, and
SUV’s.
3.1.5 Results and results discussion
3.1.5.1 Participants
All 19 young drivers successfully completed the seven sessions of the study ( i. e.,
screening + six experimental sessions). A description of the demographic, driving
experience and visual acuity test measures for the sample of participants is shown in
Table 1. In our sample, males drove further per year than females. Females reported more
crashes and more moving violations than the males. Eight males and seven females
owned MP3 players. Eight participants ( 5 males, 3 females) reported owning a model of
the Apple iPod player. Six of the 15 owners reported having used their MP3 players
while driving ( 4 males, 2 females).
27
Table 1: Participants characteristics
N Mean
Age
( SD)
Avg.
Km/ yr
Crashes Moving
violation
Visual
Acuity
Left
Visual
Acuity
Right
Short
Distance
VA
Male 9 19.33
( 0.87)
22,222
( 10663)
0.11
( 0.33)
0.44
( 0.73)
1.0
( 0.18)
0.99
( 0.20)
0.81
( 0.15)
Female 10 19.4
( 1.35)
18,040
( 12,246)
0.60
( 0.70)
0.90
( 1.29)
1.13
( 0.38)
1.17
( 0.50)
0.98
( 0.10)
Total 19 19.37
( 1.12)
20,021
( 11,407)
0.37
( 0.60)
0.68
( 1.06)
1.07
( 0.30)
1.08
( 0.39)
0.90
( 0.15)
3.1.5.2 Experimental Design
Data were analyzed using an adjusted split plot ANOVA with gender as the between
subjects factor; secondary task ( iPod easy, iPod difficult, Baseline), event type
( pedestrian, lead vehicle braking, pullout vehicle, and light change), and occurrence of
the event ( 3 per condition) were the within- subjects variables. Occurrence is defined as
the order of each event type within each secondary task condition. Overall analyses
examined the larger design with all sessions, events, and secondary tasks included to
determine the pattern of results. Significant interactions were followed- up by examining
each event type separately because event type was considered contextually different a
priori. Multiple comparisons were made using the Sidak adjustment ( Tabachnick &
Fidell, 2006).
Analyses of a number of dependent variables are not included here due to manuscript
length restrictions, but can be found in Chisholm ( 2006). These include perception time,
response time ( i. e., fractionated from PRT), minimum headway distance, velocity change
during iPod interactions, standard deviation of lane position ( SDLP), standard deviation
of velocity, horizontal and vertical gaze variability, and percentage of time to complete a
task ( i. e., % of total glance duration to a task). The selection of dependent variables for
inclusion here was based on the importance of the result, removal of redundant results
and completeness of interpretation ( i. e., one variable from hazard response, lateral
vehicle control, eye movements and secondary task performance). Thus, the results are
organized by hazard response ( i. e., PRT and collisions), lateral control ( SD of steering
wheel angle), eye movements ( glance frequency, glance duration), and secondary task
performance ( task completion time). Definitions of each of these dependent variables
precede the results of each analysis and can be found in the appendix.
3.1.5.3 Discussion
This study examined the effects of repeated iPod interactions on driver performance to
determine if performance decrements decreased with practice. A multi- measure approach
was used to understand the range of driver performance dimensions including hazard
detection and response, lateral vehicle control, eye movements, and secondary task
28
performance. A comprehensive and convergent view of the effects of distraction on
driver performance with practice is evident.
3.1.5.3.1 Hazard Detection
iPod interactions impaired drivers’ ability to respond to hazards on the roadway and
maintain safe vehicle control. Difficult iPod interactions resulted in decrements to PRT.
Over the events and occurrence, PRT increased by 0.13 s or 13% over the baseline when
performing the iPod difficult task, depending on event type. Similar to Donmez et al.,
( 2006) who found a 0.33 s increase in perception time to a lead vehicle braking event,
during iPod difficult interactions a 0.42 s increase in perception time was found for the
first occurrence over that of the baseline for the braking events.
In Strayer and Drews ( 2004), younger and older drivers response times ( RT) to a lead
vehicle braking while naturalistically conversing on a cell phone, had an 18 percent
increase in braking response time, which is analogous to PRT. Overall, a 16 percent
increase was found for braking events in the present study. However, a 26% increase in
PRT during iPod difficult interactions over the baseline was found for the first occurrence
of the lead vehicle braking events, which would indicate that the difficult iPod task might
be more difficult than cell phone conversations. Although similar distraction decrements
between these studies are evident, cell phone conversations and iPod interactions differ in
a number of important ways. First, a cell phone conversation is mainly cognitively
absorbing but does not require deviation of gaze from the roadway. iPod interactions are
both cognitively and visually absorbing, requiring attention to be directed away from the
roadway and to the interface. Second, when an event occurs, attention must be
disengaged from the conversation or returned from the iPod to the roadway. Many would
argue that prolonged glances away from the road pose greater crash risk ( Dingus et al.,
1989; Green, in press). This argument is substantiated in this study by the higher
frequency of collisions while interacting with the iPod difficult tasks ( 53) than during
either the iPod easy ( 34) or the baseline drives ( 28).
During the easy iPod tasks, no consistent detrimental effects were found. The iPod easy
task took very little time to complete ( M = 4.3 s). The average time to complete iPod
difficult tasks was approximately 30 seconds. This difference in task completion time had
varying effects on driving performance. Specifically, the longer iPod difficult tasks
exhibited consistent detrimental distraction effects, whereas the detrimental effects of the
iPod easy task were brief and transient.
While iPod interactions had a consistent detrimental effect on hazard detection, however,
vehicle control findings were less concise. Contrary to previous studies on cell phones
( Shinar et al., 2005) and speech- based e- mail ( Jamson et al., 2004) that found a
significant decrease in steering wheel variation while drivers were engaged in
conversations or email, respectively, participants had greater amounts of steering angle
variability in the iPod difficult condition than in the baseline. Cell phone conversations
and speech based e- mail tasks do not require a driver to physically manipulate something,
29
and thus may not affect steering per se. Therefore, the driver is able to focus on the road
and control the vehicle. Completion of the iPod difficult task, however, required attention
to be directed into the vehicle and physical manipulations to be made. In particular, iPod
interactions required attention to be focused, in a serial fashion, between the iPod and the
roadway to accomplish both tasks. On average, it took approximately 15 glances into the
vehicle to complete an iPod difficult task.
The largest eye movement effects occurred between the iPod difficult and baseline
conditions. Glance durations toward the roadway during the iPod difficult task were 0.24
s shorter. Furthermore, average duration of glances into the vehicle during iPod difficult
tasks was 1.35 s compared to 0.56 s in the baseline. This increase in glance duration is
similar to previous findings ( Green & Shah, 2004). Normal glances into the vehicle are
shorter in duration than glances towards the center of the road, 0.41 s and 0.73 s,
respectively ( Olson et al., 1989). Attention to in- vehicle tasks caused other sources of
driving information to be dropped from scan patterns ( i. e. off road objects and rearview
mirror). A serial sampling between in- vehicle task and immediate forward roadway
resulted ( Wierwille, 1993).
The demands of the driving environment also affected eye movements. In the higher
velocity freeway environment ( 100 km/ h), more time was spent looking at the road, less
time off road, and slightly less time into the vehicle. In- vehicle task and roadway
demands pull and push attention, and thus the eyes, while driving. If the eyes are inside
the vehicle when a critical event occurs, the success of responding adequately is less
likely. The added eye movement demands into the vehicle thus would degrade hazard
detection to the degree that eyes are in the vehicle at event onset ( also see Horrey et al.,
2006).
3.1.5.3.2 Prolonged Experience
The purpose of a multiple- session approach was to determine if repeated practice of the
secondary task while driving in demanding contexts would lessen the detrimental impact
of the distraction on driver performance. Shinar et al. ( 2005) found repeated trials of
conversation with a cell phone lessened the performance decrement on the vehicle control
measures of speed maintenance, lane positioning, and steering wheel deviations. The
present study used an event- based paradigm to examine the impact of a common MP3
player on distraction. However, Shinar et al. ( 2005) did not measure this variable
sufficiently in their study. Thus, their conclusions that cell phones are less of a safety
threat than previously thought may be premature.
Single session or cross- sectional studies may not provide an accurate picture of
cumulative distraction effects. Results are likely to differ with driver practice. These
results do not support the inference or conclusion that single session testing is invalid
because with practice, the detrimental impact goes away. Although decreases in PRT
were found with practice, performance with the iPod difficult task never achieved the
same level of performance as in the baseline condition. Even after additional practice,
30
drivers were still unable to improve their dual- task performance to a safe level. Multiple
session studies do provide information with which to determine the practical effects of in-vehicle
distractions over time compared to single session designs. While there are costs
and benefits with each method, the added array of information provided over several
months of testing sheds light on certain aspects of driver behavior such as adaptation.
3.1.5.3.3 Adaptation
With the addition of secondary tasks, drivers were hypothesized to adapt to the additional
task demands by changing their behavior. Decreasing speed, task shedding, and stopping
at the light change intersections, would indicate changes in behavior that compensate for
these demands ( Caird & Dewar, in press). One adaptive strategy would be to shed the
secondary task in favor of the primary task of driving. There were no consistent
differences in the time that it took to complete the secondary task that would seem to
indicate adaptation. Instead, results indicate that participants actually sped up the
movement time to the device itself from session 1 to session 6, and did not put off
beginning the task as an adaptive behavior. A consistent adaptive behavior that was found
involved stopping at the light changing intersections. Drivers often used the light change
as an opportunity to complete iPod tasks safely and were more likely to stop at the light
to do so.
Multiple session studies provide information with which to determine the practical effects
of in- vehicle distractions over time compared to single session cross- sectional designs.
This study found improvements in performance over the six sessions. However, it did not
determine the extent of experience over which participants might continue to improve.
Presumably a plateau or “ ceiling” would be reached and no additional practice would
affect performance. The learning curve associated with dual- task distraction performance
is an area of potential future research.
3.1.5.3.4 Distraction Metrics
Two metrics have been suggested to quantify distraction potential of in- vehicle devices.
These include number of glances to the device and task completion time ( Blanco et al.,
2005; Green, 1999b). Tasks that require more than nine glances or greater than 15
seconds to complete statically represent problematic tasks that should not be engaged
while the vehicle is in motion. As can be seen in Figure 17, using the suggested metrics
of task completion time and glance frequency, the iPod easy tasks ( green) conform to the
suggested criteria of less than nine glances and 15 second completion time. However, the
iPod difficult results ( red) took, on average, 15 glances and 35 seconds to complete the
task. The iPod difficult interactions are clearly not appropriate to perform while the
vehicle is in motion.
31
Figure 14: Frequency of glances and task completion time in seconds
32
3.2 Naturalistic driving data mining
Data collected during TO 4222 have been mined in order to quantify the triggers for lane
changing due to reaching a slower vehicle. As a reminder, lane changes can be
categorized based on the reason why they occur. For example, we distinguish between a
lane change for entering the highway, a lane change for reaching a preferred lane, and a
lane change for overtaking a slower vehicle. The work presented below is the description
of the method. This method requires a fair amount of automatic processing that no PATH
personal was available to conduct; therefore, all of the work presented below is the result
of manual processing. This process is extremely time consuming and therefore could not
be applied to the entire set of data available.
Data was collected for a two weeks period and the data analysis was carried out on
commutes, where most of the variability in terms of driver behavior can be attributed to
traffic. One of the assumptions of the model is that when commuting, drivers will first try
to reach a preferred lane, then stay on this lane and start to move back toward the exit
lane. These two moments, reaching the preferred lane and moving toward the exit, are
called transition zone.
The figure below illustrates a method for assessing the length of the transition zone. The
observed commute was composed of three distinct “ cruising” zones, in other words, once
the driver entered the highway, he would “ cruise” for several miles, then exit on a
highway split and reenter onto another highway. In this figure, all of the lane changes
observed on the second section of cruising have been plotted. The x axis indicates the
direction of the lane change, i. e. if it was toward the right or the left. The y axis indicates
the position where the lane change was initiated on that section.
33
Figure 15: Lane changes categorization
From this graph, we can deduce that the driver’s preferred lane was two lane to the left of
the entrance and that he reached this lane within 2,5 km of the entrance, at the exception
of Trip 30 where the driver returned to the right lane shortly after reaching his preferred
lane and stayed there for a quarter of the commute. This difference is likely explained by
the difference of traffic, with a denser traffic during that commute. It also shows that the
driver prepares its exit in two phases, within 2 to 4 km from the exit, the driver start
moving toward the exit lane. Within 1 km from the exit, he is taking the lane that will
exit. This information can be used in order to determine when to trigger lane changes that
correspond to the driver’s plan on how to follow his itinerary and in order to give an
order of criticality to the behavior.
The figure below illustrates the successive following events during one commute. For
each of the following event, the end/ beginning cause of the event is specified. For
example, the first two following events are not considered following, as the driver
reached slower vehicles and then changed lane, these two lane changes are also part of
the transition when the driver intends to reach his preferred lane for this section. Another
data contained in this graph is the brake activation, which occurred three times.
34
Figure 16: following events during one cruising section
Two interesting events are observable. The first one is when another vehicle ( Principal
Other Vehicle) cuts in front of the driver as a very short time- gap, initially measured at
less than 0.4 seconds, which is a very short distance, however, the driver does not brake.
On the other hand, when following the last target, the driver initiated a braking three
times. This difference in behavior can be better understood when looking also at the
vehicle speed, such as displayed in the figure below.
35
Figure 17: Vehicles speed and time gap
In this figure, the vehicles speed and time- gap between the host vehicle driven by our
participants and the lead vehicle are plotted against time. The time gap is multiplied by
10 to accommodate the speed scale. For the cut- in event, the lead vehicle speed was
superior to the one of the host vehicle, while for the second event, the speed of both
vehicles was similar, and they both considerably reduced speed at one point, therefore the
first two braking are a response to the lead vehicle braking.
Another way to look at this data is to plot the relative velocity on the x axis and the time
gap on the y axis, as done on the figure below. The traces of each of the targets from the
figure above have been plotted below.
36
Figure 18: Gap regulation
We propose that in order to quantify the different triggers intervening during car
following behavior, we should categorize the following events based on where they
happen during the commute, how they are initiated ( catch up with a slower lead or a
vehicle from the adjacent lane pulls in front) and ended ( lead vehicle changes lane, open
the gaps, or host vehicle changes lane). For each of these categories, distribution can be
made and then used in simulation. The biggest challenge with the dataset at hand is the
difficulties with which to treat the radar traces, as the current sensor provide up to 7
different target, and that the interpretation of the target location depends of the road
geometry, and distinguish between the lead vehicle and the other vehicles on the adjacent
lanes. In other words, if the target ahead is on a curve, it becomes difficult to determine
which vehicle is the lead vehicle. This type of processing is made for systems such as
adaptive cruise control, but they require years of algorithm fine tuning and incur a cost
that is beyond the scope of such a project. In future data collection, this problem could be
overcome with the use of a different sensor on the host vehicle, and possibly the use of
digital map providing information about the road geometry.
37
4 Conclusion
The goal of this report was to present two key steps on the development of a cognitive
driver model that can be implemented and interfaced with micro traffic simulation tools:
• The support of the development of a cognitive architecture or the adoption of an
existing architecture supporting the development of the driver model.
• The gathering of data supporting the development of the model in terms of
calibration of parameters and a specific description of behavior.
Regarding the adoption of a cognitive architecture for a driver model, we first depicted
several different meanings the term “ driver model” can have among the various research
fields involving driver models, and to the different approaches it produces for
implementation. We then presented several models implemented in different simulation
environments, and synthesized how these models cover the different components needed
for such a simulation; for example, which type of behaviors they allow to generate. Our
review concludes that most of the existing models are difficult to apply in our context.
The main difficulty results from the state of development of these models. All of these
driver models implementation are still in relatively early development stage, and cannot
be exchanged with other platforms that already have some well developed components,
such as roadway or vehicle models. Also, the in- house development of these models and
simulation suffer from a lack of available documentation, making it also difficult for
other developers to use. Although the level of detail provided regarding these models
gives a mean to assess the models’ benefits and relevance, it does not allow for a direct
application. Nevertheless, this review lead to a fruitful reflection about the challenges
linked to the design and implementation of a model, where we identified the elements of
the system/ framework that the driver model is a part of, as well as how the level of details
of these elements can vary in terms of their representation in the modeling effort.
Regarding the gathering of data, we presented to two different sets. In the first set, the
goal of the data collection was to assess if practice with the system mitigates the decrease
in driving performance in terms of vehicle control and hazard detection. The results
indicate that under complex manipulation of the iPod ( requiring 5 to 7 steps, for a task
such as finding a specific song), the reaction time for the detection hazard increases, lane
keeping control is also affected, most likely due to the manipulation of the iPod. In terms
of effect over time, practice with the system improved the driving performance to some
extent, but it depended of how complex the secondary task. This new set of data also
provides quantitative descriptions of the time required to interact with this type of device
is terms of glances duration and number of glances that can be directly integrated into the
model.
The second set of data presented the results of the mining of a naturalistic data collection
and addressed lane changes at the strategic, tactical and operational level. Most research
focuses on the operational level involved during a lane change, and describes it in terms
of time- gap with a lead vehicle prior to the action, gap with the side traffic. Here, we
illustrated that lane changes can be categorized based on whether they are accomplished
in order to respond to a navigation goal or overtaking of a slower vehicle. The application
38
of this method will allow the creation of distributions describing the range of values to
expect for the parameters involved in lateral control and decisions for lane changes.
The information presented in this report will be used in order to continue the
development of the model, and further progress will be presented in a subsequent report.
39
5 References
Ålm, H., Nilsson, L., 1994. Changes in driver behaviour as a function of handsfree
mobile phones -- a simulator study. Accident Analysis and Prevention, 26( 4), 441- 451.
Ålm, H., Nilsson, L., 1995. The effects of a mobile telephone task on driver behaviour in
a car following situation. Accident Analysis and Prevention, 27( 5), 707- 715.
Applied Sciences Laboratory, 2001. Eyenal Manual ( v. 2.0). Bedford, MA: Author.
Brackstone, M. ( 2000) An Examination of the Use of Fuzzy Sets to Describe Relative
Speed Perception Ergonomics, 43 ( 4), pp 528- 42.
Brackstone, M., McDonald, M. and Wu, J. ( 1997) Development of a Fuzzy Logic Based
Microscopic Motorway Simulation Model. Proceedings of the IEEE Conference on
Intelligent Transportation Systems ( ITSC97).
Blanco, M., Hankey, J. M., Chestnut, J. A., 2005. A taxonomy for secondary in- vehicle
tasks based on eye glance and task completion time. Proceedings of the 49th Annual
Meeting of the Human Factors and Ergonomics Society ( pp. 1975- 1979). Santa Monica,
CA: Human Factors and Ergonomics Society.
Brookhuis, K. A., De Vries, G., De Waard, D., 1991. The effects of mobile telephoning
on driving performance. Accident Analysis and Prevention, 23( 4), 309- 316.
Brown, I. D., Tickner, A. H., Simmonds, D. C. V., 1969. Interference between concurrent
tasks of driving and telephoning. Journal of Applied Psychology, 53( 5), 419- 424.
Caird, J. K., Chisholm, S., Lochhart, J., in press. The effect of in- vehicle advanced signs
on older and younger drivers’ intersection performance. International Journal of Human
Computer Studies.
Caird, J. K., Chisholm, S. C., Lockhart, J., Vacha, N., Creaser, J. I., Edwards, C. J., et al.,
2007. In- vehicle intelligent transportation system ( ITS) countermeasures to improve older
driver intersection performance ( Rep. No. 14610e). Montreal: Transportation
Development Centre, Transport Canada.
Caird, J. K., Dewar, R. E., in press. Driver distraction. In R. E. Dewar & R. Olsen ( Eds.),
Human Factors in Traffic Safety ( 2nd Ed.). Tucson, AZ: Lawyers & Judges Publishing.
Caird, J. K., Scialfa, C. T., Ho, G., Smiley, A., 2004. Effects of cellular telephones on
driving behaviour and crash risk: Results of meta- analysis. Edmonton, AB: CAA
Foundation for Traffic Safety.
40
Carsten, O., Brookhuis, K., 2005. The relationship between distraction and driving
performance: Towards a test regime for in- vehicle information systems. Transportation
Research: Part F, 8, 75– 77.
Casson, E. J., Recette, L., 2000. Vision standards for driving in Canada and the United
States. A review for the Canadian Opthalmological Society. Canadian Journal of
Ophthalmology, 35, 192- 203.
Chisholm, S., 2006. Practice effects on driver performance during MP3 player
interaction. Unpublished M. Sc. Thesis. Calgary, Alberta: University of Calgary.
Chisholm, S., Caird, J. K., Teteris, E., Lockhart, J., Smiley, A., 2006. Novice and
experienced driver performance with cell phones. Proceedings of the 50th Annual Human
Factors and Ergonomics Meeting ( pp. 2354– 2358). Santa Monica, CA: Human Factors
and Ergonomics Society.
Cody D. Gordon T. 2005 TRB Workshop on driver models: a step towards a
comprehensive model of driving Workshop on “ Modelling Driver Behaviour in
Automotive Environments” sponsored by the European Commission, General Directory
INFSO, within the 6th Framework Programme of R& D Ispra Italy
Dingus, T. A., Antin, J. F., Hulse, M. C., Wierwille, W. W., 1989. Attentional demand
requirements of an automobile moving- map navigation system. Transportation Research:
A, 23A( 4), 301– 315.
Donmez, B., Boyle, L., Lee, J. D., Scott, G., 2006. Assessing differences in young
drivers’ engagement in distractions. Papers of the 80th Annual Transportation Research
Boardmeeting [ CD- ROM], Washington, D. C.: Transportation Research Board.
Goodman, M. J., Bents, F. D., Tijerina, L., Wierwille, W., Lerner, N., Benel, D., 1997.
An investigation of the safety implications of wireless communications in vehicles ( DOT
HS 806- 635). Washington, D. C.: National Highway Transportation Safety
Administration. http:// www. nhtsa. dot. gove/ people/ injury/ research/ wireless/
Green, P., in press. Where do drivers look while driving ( and for how long)? In R.
E. Dewar & R. Olson ( Eds.), Human factors in traffic safety ( 2nd Ed.) Tucson, AZ:
Lawyers & Judges Publishing.
Green, P. 1999a., Visual and task demands on driver information systems. ( UMTRI Tech.
Rep. No. 98- 16). Ann Arbor: University of Michigan, Transportation Research Institute.
Green, P. 1999b., The 15- second rule for driver information systems. In Proceedings of
theITS America 9th Annual Meeting. Washington, DC: ITS America.
41
Green, P., Shah, R., 2004. Safety Vehicles using adaptive Interface Technology ( Task 6):
Task time and glance measures of the use of telematics: A tabular summary of the
literature. Ann Arbor, MI: University of Michigan, Transportation Research Institute.
Horberry, R., Anderson, J., Regan, M. A., Triggs, Y. J., Brown, J., 2006. Driver
distraction: the effects of concurrent in- vehicle tasks, road environment complexity and
age on driving performance. Accident Analysis and Prevention, 38, 1, 185– 191.
Horrey, W. J., Wickens, C. D., 2006. The impact of cell phone conversations on driving:
A meta- analytic approach. Human Factors, 48( 1), 196– 205.
Horrey, W., J., Wickens, C. D., Consalus, K. P., 2006. Modeling drivers’ visual attention
allocation while interacting with in- vehicle technologies. Journal of Experimental
Psychology: Applied, 12( 2), 67- 78.
Hosking, S. G., Young, K. L., Regan, M. A., 2005. The effects of text messaging on
young novice driver performance. ( Final Rep.). Melbourne, Australia: The National
Roads and Motorists’ Association Motoring Services.
Ishihara, S., 1993. Ishihara’s test for colour- blindness. Tokyo: Kanehara. International
Standards Organization, 2002. Road vehicles – Measurement of Driver Visual behaviour
with respect to transport information and control systems – Part 1: Definitions and
parameters, ISO committee Standard 15007- 1, Geneva, Switzerland: International
Standard Organization.
Jackson, D. L.; Bhise, V. D. ( 2002). An evaluation of the IVIS- DEMAnD driver attention
demand model. Michigan University, Dearborn. 10 p. Human Factors in Seating and
Automotive Telematics, Warrendale, SAE, 2002, p. 61- 70. Report No. SAE 2002- 01-
0092. UMTRI- 95608 A08
Jamson, A. H., Westerman, S. J., Hockey, G. R., Carsten, M. J., 2004. Speech based e-mail
and driver behaviour: Effects of an in- vehicle message system interface. Human
Factors, 46( 4), 625- 639.
Kennedy, R. S., Lane, N. E., Berbaum, K. S., Lilenthal, M. G., 1993. Simulator sickness
questionnaire: An enhanced method for quantifying simulator sickness. International
Journal of Aviation Psychology, 3( 3), 203- 220.
Klauer, S. G., Dingus, D. R., Neale, T. A., Sudweeks, J., Ramsey, D. J., 2006. The impact
of driver inattention on near- crash/ crash risk: An analysis using the 100- car naturalistic
study data ( Rep. No. DOT HS 810 594). Washington, D. C.: National Highway Traffic
Safety Administration.
Laberge- Nadeau, C., Maag, U., Bellavance, F., Lapierre, S. D., Desjardins, D., Messier,
S., Saïdi, A., 2003. Wireless telephones and the risk of road crashes. Accident Analysis
and Prevention, 35, 649- 660.
42
Lee, J. D., Caven, B., Haake, S., Brown, T. L., 2001. Speech based interaction with in-vehicle
computers: The effect of speech based e- mail on drivers’ attention to the
roadway. Human Factors, 43, 631- 640.
McEvoy, S. P., Stevenson, M. R., McCartt, A. T., Woodward, M., Haworth, C., Palamara,
P., Cercarelli, R., 2005. Role of mobile phones in motor vehicle crashes resulting in
hospital attendance: A case- crossover study. British Medical Journal, bmj. com.
Olson, P. L., Battle, D. S., Aoiki, T., 1989. Driver eye fixations under different operating
conditions. ( Tech Rep. No. UMTRI- 89- 3). Ann Arbor: University of Michigan
Transportation Research Institute.
Olson, P. L., Farber, E., 2003. Forensic aspects of perception and response ( 2nd Ed.).
Tuscon, A. Z.: Lawyers and Judges Publishing.
Parkes, A., Hööijmeijer, V., 2001. Driver situational awareness and car phone use.
Proceedings of the 1st Human- Centered Transportation Simulation Conference ( ISSN
1538- 3288). Iowa City, IA: University of Iowa.
Recarte, M. A., Nunes, L. M., 2000. Effects of verbal and spatial imagery tasks on eye
fixations while driving. Journal of Experimental Psychology: Applied, 6, 31- 43.
Recarte, M. A., Nunes, L. M., 2003. Mental workload while driving: Effects on visual
search, discrimination, and decision making. Journal of Experimental Psychology:
Applied, 9( 2), 119- 137.
Redelmeier, D. A., Tibshirani, R. J., 1997. Association between cellular- telephone calls
and motor vehicle collisions. The New England Journal of Medicine, 336( 7), 453- 458.
Shinar, D., Tractinsky, N., Compton, R., 2005. Effects of practice, age, and task demands
on interference from a phone task while driving. Accident Analysis and Prevention,
37( 3), 315- 326
Strayer, D. L., Drews, F. A., 2004. Profiles in driver distraction: Effects of cell phone
conversation on younger and older drivers. Human Factors, 46( 4), 640- 649.
Stutts, J., Feaganes, J., Reinfurt, D., Rodgman, E., Hamlett, C., Gish, K., et al., 2005.
Driver’s exposure to distractions in their natural driving environment. Accident Analysis
and Prevention, 37, 1093- 1101.
Stutts, J. C., Reinfurt, D. W., Staplin, L., Rodgman, E. A., 2001. The role of driver
distraction in traffic crashes. Washington, D. C.: AAA Foundation for Traffic Safety.
http:// www. aaafoundation. org/ projects/ index. cfm? button= distraction
43
Salvucci, D. D. ( in press). Integrated driver modeling in the ACT- R cognitive
architecture. To appear in T. Jürgensohn ( Ed.), Driver Modelling.
Salvucci, D. D., & Lee, F. J. ( 2003). Simple cognitive modeling in a complex cognitive
architecture. In Human Factors in Computing Systems: CHI 2003 Conference
Proceedings ( pp. 265- 272). New York: ACM Press.
Tabachnick, B. G., Fidell, L. S., 2006. Using Multivariate Statistics ( 5th ed.). Allyn and
Bacon. Boston, MA.
Theeuwes, J. ( 2001). Chapter 13: The Effects of Road Design on Driving. In: Barjonet,
Pierre- Emmanuel ( ed). Traffic Psychology Today. Boston: Kluwer Academic Publishers,
241- 263.
Wierwille, W. W., 1993. An initial model of visual sampling of in- car displays and
controls. In A. G. Gale et al. ( Eds), Vision in Vehicles IV ( pp. 271- 279). Amsterdam:
Elsevier Science.
Wu, J., Brackstone, M., McDonald, M. ( 2000) " Fuzzy Sets and Systems for a Motorway
Microscopic Simulation Model". Fuzzy Sets and Systems 116( 1), pp 65- 76.
Wu, J., Brackstone, M. and McDonald, M. ( 2003) The Validation of A Microscopic
Simulation Model: A Methodological Case Study Transportation Research C, 11( 6), pp
463- 479. 2003.
44
Appendix – Detailed Results from driving simulator experiment
Hazard Detection and Response
* Perception Response Time
Perception response time ( PRT) was calculated in seconds from the onset of an event to a
braking response ( Olson & Farber, 2003). Overall, male drivers had significantly shorter
PRTs ( M = 1.06, SE = .019) than females ( M = 1.12, SE = .017), F( 1, 492) = 4.35, p =
.037. The longest PRT was found while drivers were performing the difficult iPod tasks
( M = 1.17, SE = .02), than the baseline ( M = 1.05, SE = .02), and iPod easy ( M = 1.06,
SE = .02) conditions, F( 2, 35) = 6.21, p = .005, which is illustrated in Figure 14. Post hoc
analyses showed significant differences between the iPod difficult and baseline
conditions, p < .05, and between the iPod difficult and iPod easy conditions, p < .05.
Figure 19: Perception Response Time ( PRT) by secondary task and occurrence
Drivers’ perception response time improved significantly over occurrences with the
longest PRT in the 1st occurrence ( M = 1.17, SE = .02) followed by the 2nd ( M = 1.09,
SE =. 02) and the shortest PRT in the 3rd occurrence ( M = 1.01, SE = .02), F( 2, 35) =
13.32, p <. 001. All of which differed significantly in post hoc comparisons, p < .05.
Event type also significantly affected perception response times to hazards, F( 3, 475) =
92.38, p <. 001. Lead vehicle braking events had the longest mean PRT ( M = 1.36, SE =
.02) followed by the pullout vehicle ( M = 1.19, SE = .02), the pedestrian ( M = 1.04, SE =
.03) and finally the light changing ( M = 0.78, SE = .03) events. All event types differed
significantly from each other, p < .001.
45
The two- way interactions between occurrence and event type, F( 6, 475) = 3.16, p = .005,
and between secondary task and event type, F( 6, 475) = 4.12, p < .001, were also
significant. Both interactions however, were embedded in a significant three- way
interaction among occurrence, secondary task, and event type, F( 11, 475) = 5.53, p <
.001. Follow- up analyses examined each event type separately.
Pedestrian Event. Secondary task had a significant effect on PRT to the pedestrian event,
F( 2, 34) = 5.97, p = .006. The fastest PRT to the pedestrian event was observed with the
easy iPod tasks ( M = 0.90, SE = .05), which was significantly faster than the iPod
difficult task ( M = 1.15, SE = .04), p < .05. However, both iPod tasks did not differ
significantly from the baseline ( M = 1.03, SE = .04), p > .05.
The interaction between occurrence and secondary task type was significant, F( 3, 50) =
11.07, p < .001. For the 1st occurrence of the pedestrian event, there were no differences
in PRT between the baseline or iPod difficult tasks, F( 1, 35) = 2.31, p > .05. In the 2nd
occurrence of the pedestrian the fastest PRT was found in the iPod easy task ( M = 0.78,
SE = .07) followed by the baseline ( M = 1.05, SE = .07), and the longest PRT times were
found in the iPod difficult condition ( M = 1.46, SE = .07), F( 2, 54) = 21.07, p < .001.
All of which significantly differed from each other, p < .05. By the 3rd occurrence of the
pedestrian event, no significant differences were found between the secondary task
conditions, F( 2, 54) = 0.58, p > .05.
Lead Vehicle Braking Event. Perception response time ( PRT) for the braking event
indicated steady improvement in performance between the 1st occurrence ( M = 1.54, SE
= .06), and decreasing on the 2nd ( M = 1.34, SE = .06), and 3rd occurrences ( M = 1.20,
SE = .06), F( 2, 39) = 8.61, p = .001. The difference between the 1st and 3rd occurrence
was significant, p < .05.
Secondary task also significantly affected drivers perception response time to the braking
event, F( 2, 39) = 3.30, p = .048. The fastest PRT to the lead vehicle braking was observed
in the baseline condition ( M = 1.24, SE = .06), followed by the iPod easy task ( M = 1.40,
SE = .06), and finally the iPod difficult condition ( M = 1.44, SE = .06). Only the
difference between the baseline and difficult iPod conditions was significant, p < .05.
Pullout Vehicle Event. During the pullout vehicle event perception response times were
significantly longer in the iPod difficult task ( M = 1.36, SE = .04), than the iPod easy ( M
= 1.15, SE = .04), and the baseline conditions ( M = 1.08, SE = .04), F( 2, 36) = 12.95, p<
.001. Significant differences were found between the iPod difficult and baseline, p <
.001, and between the iPod difficult and the iPod easy conditions, p < .001. The two- way
interaction between occurrence and secondary task was significant, F( 4, 70) = 4.37, p =
.003.
As can be seen in Figure 20, for the pullout event over occurrences, baseline means
remained relatively constant whereas in the difficult iPod condition, a decrease in PRT
was found over occurrence. In the 1st occurrence of the pullout event results, as expected,
the shortest PRT times occurred in the baseline ( M = 1.10, SE = .05), followed by the
46
iPod easy task ( M = 1.14, SE = .05), and the longest PRT found in the iPod difficult
condition ( M = 1.41, SE = .05), F( 2, 54) = 10.08, p < .001. No significant differences
were found between the baseline and iPod easy conditions, but both differed significantly
from the iPod difficult conditions, p < .05. During the 2nd occurrence of the pullout
event, the shortest PRT was observed during the baseline ( M = 0.998, SE = .07),
followed by the iPod difficult ( M = 1.33, SE = .07), and iPod easy ( M = 1.34, SE = .07)
conditions, F( 2, 53) = 7.73, p = .001. Comparisons show that PRT for both iPod tasks
were significantly longer than the baseline, p < .05. In the 3rd occurrence of the pullout
event, PRT was significantly affected by secondary task, F( 2, 53) = 4.72, p = .013.
However, the iPod easy condition resulted in the shortest PRT ( M = 0.96, SE = .07),
followed by the baseline ( M = 1.13, SE = .07), and the iPod difficult condition ( M = 1.28,
SE = .08). PRT in the iPod difficult condition was significantly longer than the iPod easy
condition, p < .05.
Figure 20: Perception Response Time to the pullout vehicle by the secondary task and occurrence
Light Changes. As hypothesized, the longest PRT was found for the 1st occurrence of the
light change event ( M = 0.86, SE = .03) and decreased for the 2nd ( M = 0.76, SE = .03)
and 3rd occurrences ( M = 0.74, SE = .03), F( 2, 40) = 3.33, p = .046. The latter two means
significantly differed from the 1st occurrence, p < .05, but not from each other, p > .05.
PRT to the light changes was not found to differ significantly depending on the secondary
task being performed, F( 2, 47) = 2.21, p > .05. Nor was the two- way interaction between
occurrence and secondary task significant, F( 4, 50) = 0.73, p > .05.
Collisions
The number of collisions that participants experienced during the various events was
determined by examining the minimum headway distance data. A total of 513 event
occurrences were included in this analysis, which represents all the experimental
combinations of the independent variables, with the exception of the light changing
events ( i. e., pedestrian, lead vehicle braking, and pullout vehicle), for each of the
47
secondary tasks ( i. e., baseline, iPod easy, and iPod difficult). Of the 513 events
encountered by participants over the course of the six sessions, a total of 115 collisions
resulted. Secondary task had a significant effect on collision frequency, χ2 ( 2) = 11.67, p
= .003. Twenty- eight collisions occurred during the baseline drives, 34 during the iPod
easy condition, and 53 in the iPod difficult interactions. Significant differences in
collision frequency were found between the iPod difficult and baseline conditions, χ2 ( 1)
= 10.35, p = .001, and between the iPod difficult and iPod easy condition, χ2 ( 1) = 5.60, p
= .018. Frequency of collisions also decreased significantly from the 1st occurrence ( 52)
to the 2nd occurrence ( 39) and finally the 3rd occurrence ( 24), χ2 ( 2) = 8.98, p = .011.
Standard Deviation of Steering Wheel Angle
Steering angle variation was used to determine steering corrections made while
interacting with the iPod and comparable baseline measures. Due to the short time needed
to complete the iPod easy task, only the iPod difficult and baseline analyses are
presented.
The iPod difficult tasks had larger variation in steering wheel adjustments ( M = 2.17, SE
= .04) than during the baseline ( M = 1.21, SE = .04), F( 1, 19) = 58.35, p < .001. Roadway
geometry showed the greatest effect on deviation of steering wheel angle, F( 2, 39) =
103.82, p < .001. The largest variation occurred on the straight freeway roads ( M = 2.32,
SE = .05) followed by the residential/ urban ( M = 1.39, SE = .04), and suburban ( M =
1.44, SE = .06) roadways. No significant differences in means were found between the
residential/ urban and suburban roads, p > .05, but both of these means differed
significantly from that of the freeway roads, p < .05.
The two- way interaction between secondary task and road type, F( 2, 477) = 6.85, p =
.001 was significant. Follow- up analyses examined each roadway type separately to
determine differences between the baseline and iPod difficult conditions. On the
residential/ urban roads drivers had significantly higher variation in steering position
during the iPod difficult tasks ( M = 1.93, SE = .05), compared to the baseline condition
( M = 0.86, SE = .05), F( 1, 19) = 85.03, p < .001. Significantly larger amounts of steering
wheel deviation were also found during iPod difficult secondary tasks on suburban roads
( M = 2.11, SE = .08) than the baseline ( M = 0.76, SE = .07), F( 1, 18) = 66.50, p < .001.
Finally, on freeways again greater steering deviation occurred during the iPod difficult
interactions ( M = 2.56, SE = .09) than in the baseline ( M = 2.08, SE = .09), F( 1, 18) =
7.52, p = .013.
Eye Movement Variables
Eye movements were collected on the even numbered sessions ( sessions 2, 4, and 6) in
both the iPod and baseline conditions. Video data analysis of eye movements was
analyzed using SimObserver and Data Distillery hardware and software. A glance was
defined as consecutive fixations to an area of interest ( i. e., in the vehicle, on road) not
including saccade transition time and blinking behavior ( International Standards
Organization, 2002). In- vehicle, on road, off road ( which included any signs, buildings,
48
parked cars that are not in the central roadway), rearview mirror, and on- screen text
instructions were the areas of interest ( AOI) that were extracted.
Mean Glance Frequency
Glance frequency is defined as the number of glances to a target during the task where
each glance is separated by at least one glance to a different target ( ISO, 2002). The
numbers of glances made during iPod interactions into the vehicle were examined to
determine the average number of glances needed to complete the required tasks.
Obviously the iPod difficult task required significantly more glances into the vehicle ( M
= 15.30, SE = .37) than the iPod easy interactions ( M = 1.93, SE = .37), F( 1, 17) =
223.44, p < .001. The number of glances to the rearview mirror could not be compared
using parametric statistics because too few participants in the iPod tasks glanced at the
rearview mirror. Chi- square analyses revealed that there was a significantly higher
frequency of glances to the rearview mirror in the baseline condition ( 76/ 95) compared to
the iPod easy condition ( 11/ 171), χ2 ( 1) = 150.18, p < .001; and iPod difficult condition
( 27/ 171), χ2 ( 1) = 106.12, p < .001.
Mean Glance Duration
Glance duration was calculated as the time ( in seconds) from first looking at an AOI until
gaze was moved off that area. The mean duration of each glance in seconds was extracted
and categorized into various AOIs ( i. e., in- vehicle, on- road, off- road, etc.) for each
secondary task ( i. e., baseline, iPod easy, hard). The duration of glances significantly
differed depending on the AOI, F( 6, 156) = 128.34, p < .001. Specifically, longer glances
were made into the vehicle ( M = 0.87, SE = .02) than on the road ( M = 0.69, SE = .03),
off the road ( M = 0.45, SE = .03), and to the rearview mirror ( M = 0.30, SE = .03), all of
which significantly differed from one another, p < .05.
Mean glance durations differed by secondary task, F( 2, 80) = 12.92, p < .001. Shorter
glance durations were found in the iPod easy condition ( M = 0.62, SE = .03) compared to
both the iPod difficult ( M = 0.81, SE = .02) and baseline ( M = 0.81, SE = .02) conditions,
p < .05. The two- way interactions between AOI and secondary task, ( F( 11, 1524) =
27.69, p < .001) and AOI and road type, ( F( 23, 1524) = 5.57, p < .001) were significant.
The three- way interaction among AOI, secondary task, and road type was also
significant, F( 28, 1524) = 4.10, p < .001. Follow- up analyses examined each AOI ( i. e., on
road, in- vehicle, off road, and rearview mirror) separately to determine the effects
ofsecondary task on glance duration.
On- Road. Glance duration to the roadway differed depending on secondary task, F( 2, 42)
= 4.67, p = .015. Longer glances to the roadway were found in the baseline condition ( M
= 0.83, SE = .05) followed by the iPod easy ( M = 0.62, SE = .05) and iPod difficult ( M =
0.59, SE = .05) conditions. No difference in means were found between the iPod easy and
iPod difficult tasks, p > .05, but both of these conditions had significantly shorter glances
to the roadway compared to the baseline, p < .05.
49
Mean glance duration to the roadway significantly differed depending on the two- way
interaction between secondary task and road type, F( 7, 278) = 2.92, p = .006. On the
residential roadways, significantly longer glances were made during the baseline
condition ( M = 0.81, SE = .11) than the iPod easy ( M = 0.52, SE = .08), and iPod
difficult ( M = 0.54, SE = .06) conditions, F( 2, 37) = 6.61, p = .004. No difference was
found between the iPod conditions, p > .05, but both means differed significantly from
the baseline, p < .05. No other significant interactions were found between road type and
secondary task conditions for the urban ( F( 2, 27) = 3.24, p > .05), suburban ( F( 2, 35) =
1.38, p > .05), or freeway ( F( 2, 5) = 2.08, p > .05) roads.
In- Vehicle. The in- vehicle AOI was defined as any glances that were made into the
vehicle, whether at the iPod device, center console, or speedometer. Mean glance
durations made into the vehicle, ostensibly at the iPod or speedometer differed depending
on secondary task, F( 2, 37) = 100.02, p < .001. Longer glances were made in the iPod
difficult condition ( M = 1.35, SE = .02) compared to the baseline ( M = 0.56, SE = .03)
and iPod easy ( M = 0.64, SE = .03) conditions. No differences were found for glance
duration between the iPod easy and baseline conditions, p > .05, but both were
significantly shorter than those in the iPod difficult condition, p < .05.
Mean glance duration differed depending on the road type, F( 4, 330) = 19.90, p < .001.
Significantly longer glances were made on the suburban roadways ( M = 1.13, SE = .04)
compared to all other road types: residential ( M = 0.83, SE = .03), urban ( M = 0.75, SE =
.04), freeway ( M = 0.77, SE = .03) and curved freeway ( M = 0.84, SE = .05), p < .05.
This increase in glance duration on suburban roads could be due to the fact that 88% of
drivers stopped at the lights and had more time without control or visual demands to
complete the iPod tasks. The two- way interaction between secondary task and road type
was also significant, F( 7, 330) = 20.85, p < .001, see Figure 21.
Figure 21: Average duration of glances ( s) made into the vehicle by secondary task and road type
Glance durations made into the vehicle were significantly longer in the iPod difficult
condition regardless of road type, but due to changing intersection lights on the suburban
50
roadways, increasingly longer mean glance durations were found compared to the other
roadways. For example, on the suburban roads significantly longer glances were made in
the iPod difficult task ( M = 2.11, SE = .07) compared to the iPod easy ( M = 0.58, SE =
.04) and baseline ( M = 0.71, SE = .08) conditions, F( 2, 85) = 109.45, p < .001. However,
on freeway roads, shorter glances into the vehicle were made in the baseline ( M = 0.44,
SE = .07) followed by the iPod easy ( M = 0.65, SE = .05), and the iPod difficult ( M =
1.20, SE = .07) conditions, F( 2, 83) = 45.58, p < .001. All of which significantly differed
from one another, p < .05. Residential roads showed similar results, with significantly
longer glances into the vehicle being made in the iPod difficult condition ( M = 1.22, SE =
.04) compared to both the iPod easy ( M = 0.66, SE = .04) and baseline ( M = 0.62, SE =
.07), F( 2, 120) = 25.36, p < .001. Finally, on the urban roads again glance duration in the
iPod difficult condition were significantly longer ( M = 1.03, SE = .05) than either the
iPod easy ( M = 0.67, SE = .07), or baseline ( M = 0.55, SE = .07) conditions, p < 0.5. Off
Road. The off- road AOI was defined as any area surrounding the roadway; including
signs, buildings, parked cars, and grass medians. Task difficulty did not have a significant
effect mean glance durations made off road, F( 2, 52) = 2.05, p > .05. Off road glance
durations were significantly affected by road type, F( 4, 238) = 4.62, p = .001. On the
freeway roads, glance durations off road were significantly shorter ( M = 0.26, SE = .05)
than all other road types of residential ( M = 0.43, SE = .03), urban ( M = 0.51, SE = .04),
and suburban ( M = 0.51, SE = .04), p < .05. No other significant differences were found
among the other road types, p > .05. On freeway sections, fewer objects such as parked
cars and signs are available for fixation, as well more ambient traffic is present, so less
glances are made off road.
The interaction between secondary task and road type was significant, F( 7, 238) = 2.47, p
= .018. No differences were found among the various road types for duration of glances
off the road during the iPod difficult conditions, F( 4, 136) = 0.67, p > .05. Similarly, no
difference among the road types were found in the iPod easy interactions, F( 3, 79) = 0.70,
p > .05. However, mean duration of glances in the baseline condition differed
significantly by road type, F( 4, 88) = 6.38, p < .001. Glance durations off road were
significantly higher in the suburban road ( M = 0.62, SE = .07) than in the residential ( M
= 0.47, SE = .06) and freeway roads ( M = 0.39, SE = .06), p < .05.
Secondary Task Performance
Task Completion Time ( TCT)
Time needed to complete each task was analyzed using SimObserver and Data Distillery.
Data were categorized into segments of actions that include: starting the task, movement
towards the iPod, touching the iPod, and selecting the correct song, these segments were
then aggregated to gather total task completion time in seconds. Because various songs
were required for each iPod difficult interaction, they differed in their position in the
menu system. Therefore, song position in the menu system was used as a covariate to
account for differences in task time due to distance in the menu system. All analyses were
performed on each of the secondary tasks separately. Adjustments for violations of
51
sphericity using the Greenhouse- Geisser correction are indicated with a GG next to
certain results ( Tabachnick & Fidell, 2006).
Difficult iPod Interactions. Males were significantly faster at completing the iPod
difficult tasks ( M = 31.75, SE = 1.16) than were females ( M = 39.16, SE = 1.11), F( 1,
320) = 21.28, p < .001. Task completion time for the iPod difficult task also differed by
session, F( 5, 320) = 13.09, p < .001. The longest time taken to complete the task was in
Session 2 ( M = 45.21, SE = 2.11), and Session 3 ( M = 44.31, SE = 1.99), which did not
differ between each other, p > .05, but were significantly higher than TCT in Session 1( M
= 34.43, SE = 1.98), 4 ( M = 30.38, SE = 2.05), 5 ( M = 27.90, SE = 2.02), and 6 ( M =
30.48, SE = 1.94), p < .05. No other significant differences were found between the
sessions, p > .05. The two- way interaction between gender and session was significant,
F( 5, 320) = 2.19, p = .05. In Session 1, TCT was significantly longer for the females ( M
= 38.57, SE = 2.70) than the males ( M = 30.29, SE = 2.84), F( 1, 55) = 5.65, p = .021.
Similar results for the females ( M = 52.08, SE = 2.90) and males ( M = 38.33, SE = 2.90)
were found in Session 2, F( 1, 52) = 6.30, p = .015, and Session 3 with the females ( M =
51.50, SE = 2.82) taking longer than the males, ( M = 37.13, SE = 2.82), F( 1, 52) = 6.30,
p = .015. Meanwhile, in sessions 4, F( 1, 55) = 2.63, p > .05, 5 F( 1, 52) = 0.11, p > .05,
and 6 F( 1, 55) = 0.19, p > .05, no significant gender differences were found.
Movement time to the device also decreased slightly but progressively by session, F( 5,
70) = 5.84, p = .003 ( GG). Longest movement times occurred in Sessions 1 ( M = 0.92,
SE = .04), and 2 ( M = 0.95, SE = .06), but decreased for Sessions 3 ( M = 0.83, SE = .03),
4 ( M = 0.72, SE = .03), 5 ( M = 0.79, SE = .03), and 6 ( M = 0.76, SE = .03). Significant
differences in means were found between Session 1 and both Session 4 and 6, p < .05,
and between Sessions 2 and 4, p < .05.
Easy iPod Interactions. No gender differences in task completion time were found in the
iPod easy tasks, F( 1, 14) = 0.511, p > .05. Time to complete task did differ depending on
the session, F( 5, 70) = 3.48, p = .03 ( GG). A significant decrease in completion time was
found between Session 2 ( M = 5.55, SE = .47) and Session 5 ( M = 3.81, SE = .11), p <
.05. No other significant differences were found amongst Session 1 ( M = 4.10, SE = .39),
Session 3 ( M = 4.25, SE = .49), Session 4 ( M = 4.55, SE = .21), and Session 6 ( M = 4.21,
SE = .24), p > .05. Movement time during the iPod easy task did not differ by gender,
F( 1, 13) = 0.067, p > .05, or by session, F( 5, 65) = 0.565, p > .05. Nor was the two- way
interaction between gender and session significant, F( 5, 65) = 0.417, p > .05.