Conference PaperPDF Available

A Cognitive Model of Drivers Attention

Authors:

Abstract and Figures

Cognitive architectures can account for highly complex tasks. One of the greatest challenges is understanding and modeling human driving behavior. This paper describes an integrated cognitive model of human attention during the performance of car driving. In this task, the attention process can be divided into at least three basic components: the control process, the monitoring process, and finally, the decision making process. Of these basic tasks, the first has the highest priority. All three phases are implemented in a cognitive model in the cognitive Architecture ACT-R 6.0. The model is able to keep a traffic lane, overtake another vehicle by lane change, identifies traffic signs and different situations emerging at crossroads.
Content may be subject to copyright.
A Cognitive Model of Drivers Attention
Kerstin Sophie Haring (ksharing@fennel.rcast.u-tokyo.ac.jp)
Research Center for Advanced Science and Technology, The University of Tokyo
4-6-1, Komaba, Meguro-ku, Tokyo, 153-8904, Japan
Marco Ragni (ragni@cognition.uni-freiburg.de)
Lars Konieczny (lars@cognition.uni-freiburg.de)
Center for Cognitive Science, University of Freiburg
Friedrichstr. 50, 79098 Freiburg, Germany
Abstract
Cognitive architectures can account for highly complex tasks.
One of the greatest challenges is understanding and modeling
human driving behavior. This paper describes an integrated
cognitive model of human attention during the performance
of car driving. In this task, the attention process can be
divided into at least three basic components: the control
process, the monitoring process, and finally, the decision
making process. Of these basic tasks, the first has the highest
priority. All three phases are implemented in a cognitive
model in the cognitive Architecture ACT-R 6.0. The model is
able to keep a traffic lane, overtake another vehicle by lane
change, identifies traffic signs and different situations
emerging at crossroads.
Keywords: Driver behavior model; cognitive architecture;
ACT-R; Attention
Introduction
Even for long-time practitioners driving a car is a highly
complex task. This becomes evident by the still high
number of accidents. E.g., in 2010 in Germany nearly
375.000 persons were injured in approximately 290.000
automobile accidents (Statistisches Bundesamt, 2011). In
about 84% of all cases the cause of an accident could be
traced back to driver errors (cp. Fig. 1). Nowadays passive
safety systems like the airbag are reaching their
technological limits and the focus shifts more towards active
safety systems. Active systems, however, require exact
knowledge about the driver, the vehicle, and the
environment. To increase the acceptance of active
intervention through the safety systems in cars, these
systems should act in accordance to the driver. The driver
and the human driving behavior must be considered for the
future development of safety systems. Consequently, one
focus of research is to analyze human behavior and predict
possible errors.
We present the implementation of a cognitive driver model,
simulating human attention and driving behavior. A driver
model can be a powerful instrument with several possible
fields of application, such as the development of intelligent
driver assistant systems. The model is an adaption of
Salvuccis`s (2006) driver model developed in the Cognitive
Architecture ACT-R 5. Our model is implemented the
newer version ACT-R 6 (Anderson, 2007) and using the
standard ACT-R development environment running on an
open source LISP, which not only guarantees support and
accountability, but also enables the research community to
use the developed model for further research. It is able to
keep a traffic lane, initiate and decide about a change of the
lane in case of upfront traffic, identify prevalent situations at
crossroads and react to traffic signs.
Fig. 1: Driver errors in automobile accidents with person
injury (Statistisches Bundesamt, 2011).
Previous work
Most developed approaches can be distinguished into two
classes: task specific and generic approaches. Task specific
approaches such as Cosmodrive (Bellet et al., 2007) and
Pelops (Benmimoun, 2004) reproduce the cognitive
functions of a car driver. In contrast to task specific
approaches, generic approaches can model various aspects
of human behavior. Therefore, it is necessary for these
architectures to include a theory of human information
processing. Examples for such architectures in which driver
models have been implemented are ACT-R (Anderson,
1993; Salvucci, 2006), SOAR (Aasman,1995) and QN-
MHP (Liu et al., 2006).
Previous models can be divided into three categories:
First, early models concentrated mainly on steering and lane
keeping. These models focus on the control process and are
able to detect some cognitive aspects, but according to Boer
(1999) they are highly dependent on difficult perceivable
inputs from the environments. Second category comprises
perception-action models which are through the perceptual
constraints oriented closer on human behavior (e.g. Rushton
275
et al., 1998; Salvucci & Gray, 2004; Wilkie & Wann, 2003).
Yet, these models do not allow for movement dynamics.
Finally, the third category includes models that are trying
to unify the various aspects of a driving task and are
therefore the most closely associated to the here presented
work. These models not only explore and unify the various
aspects of driving behavior, they also explore the generality
of the cognitive architectures used for their development.
Driver models were described by Aasman (1995) in the
cognitive architecture SOAR and by Liu (1996) in Queuing
Network-Model Human Processor (QN-MHP). Although
these models already exist in other cognitive architectures
and the central ideas remain the same in any architecture,
the ACT-R model of a driver shows a broader spectrum of
application (Salvucci 2001; 2006).
Salvucci (2006) developed a first integrated cognitive
model of human driving behavior in ACT-R. He showed in
his work the generality and the applicability using the
cognitive architecture ACT-R for the specific task of
driving. His model is designed to keep a standard vehicle on
a multi-lane highway with moderate traffic. The model is
also able to recognize the distance to a vehicle ahead and to
make the decision for overtaking. As driving is a highly
complex task and not readily implementable, this model has
some limitations. The model solely was meant to interact
with a highway environment without recognition of traffic
signs, crossings or slip roads. An implementation limitation
was the use of the previous version ACT-R 5.0 and its
incompatibility to newer versions. It was also not possible to
make the ACT-R model interact directly with a driving
simulator.
The cognitive architecture
A cognitive architecture compromises theories about the
operation mode of human information processing and aims
at using procedures similar to humans. In other words, it
describes a comprehensive computer model of human
cognition. ACT-R (Anderson, 1993; Anderson 2007) is such
a comprehensive theory of human cognitive capacities. It is
also a modeling environment, used to describe human
cognitive processes. Most of its basic assumptions are
inspired by the progress of cognitive neuroscience. ACT-R
is a framework in which the researcher can create models
(programs) for different tasks. Running this model produces
a simulation of human behavior. The main assumption of
ACT-R is the representation of knowledge as either
declarative or procedural knowledge. Declarative
knowledge, consisting of facts, is represented in form of
chunks, or small logical units which encode simple facts
(e.g. the fact: Berlin is in Germany). Procedural
knowledge, representing knowledge about how we do
things, is represented in form of production rules, condition-
action rules that generate a specific action (e.g. manipulate
declarative knowledge) if the conditions of this rule are
fulfilled.
In other words, ACT-R’s knowledge representation is
split in two kind of memory modules. Modules can be
accessed through their buffers. The state of ACT-R at a
given time is the content of the buffers at that time. Buffers
are connected to the modules and are changed by production
rules. Every buffer and (nearly) every module can be
allocated to a cortex region. This enables an interesting
mapping between buffers and neural processes (Anderson
2007).
Fig. 2: The organization of information the cognitive
architecture ACT-R (Anderson, 1993). The buffers contain
information and are connected to modules associated with
brain regions.
Cognitive model
We introduce now a computational model of human
attention in a car driving task implemented in the ACT-R
architecture. It models human attention and behavior for
driving a car on a straight road, overtaking another vehicle
by lane-change, identifying a traffic sign and crossroads.
Driver Modeling
The goal of this research was to develop an integrated driver
model in the context of embodied cognition, task and
artifact (ETA) framework. Byrne (2001) describes the ETA
framework as understanding of interactive behavior based
on the Cognition-Task-Artifact triad introduced by Gray
(Gray & Altman, 2001). Interactive behavior is a function of
the performed Task, the Artifact (instrument) by which the
task is performed, and the Embodied Cognition, the
cognitive, perceptual and motor capabilities by which a
person acts through the artifact.
Cognitive modeling of human driving behavior should
address all three components. An integrated model
considers the driving related tasks (Task), the interface
between the human and the vehicle (Artifact) and the
276
processes that execute the driving task on the vehicle
(Embodied Cognition). The system must be specified
regarding a detailed description of the artifact being used
and the task to perform. Some successfully implemented
and applied models only emphasize one or two of these
components like the perception-action models of control of
Fajen (Fajen & Warren, 2003), which provides a compact
description of the behavioral dynamics of steering and
obstacle avoidance, control-theoretic models like Donges
(1978), dividing the steering task into a guidance and a
stabilization level or machine-learning models, supporting
automobile drivers steering by sampling an image, assessing
the road curvature, and determining the lateral offset of the
vehicle (Pomerleau & Jochem, 1996).
Driving is a continuously changing task of basic subtasks.
These must be integrated and interleaved. This model uses
three basic components, control, monitoring, and decision
making (see Fig. 3), derived from the hierarchical control
structure of Michon (1985). Michon identified three levels
of skills and control for the driving task: operational
(control), tactical (maneuvering), and strategic al (planning).
He claims that a comprehensive model should take into
account the various levels and also provide an information
flow control that allows to switch from one level to the
other.
The independent subtasks of a simple driving task (see
Fig. 3) were implemented as control, the operational process
controlling the input, monitoring, the tactical process
interacting with the environment, and decision making, also
analogous to the tactical level of Michon (1985), managing
maneuvers like overtaking. These subtasks are processed
serially. Every production of the top level goal drive has
sub-goals, which incorporate the three components.
Fig. 3: Schematic representation of the production rules of
the driver model in a simple crossroad scenario. The title of
a box indicates the current goal and the corresponding
production rules. The arrows show the flow of control and
the asteriks the return to the parent-goal.
Development Environment The theory of ACT-R is
embedded in the ACT-R software in form of Common Lisp
functions. This model is implemented in Clozure
CommonLisp 1.3 and the current version of ACT-R 6.0
under the operating system Ubuntu 9.04. In order to make
the simulation environment interact with the ACT- R
system, it was directly implemented in LISP with simple
graphics and the extension with the LTK Lisp Toolkit. As it
was not possible to make ACT-R directly interact with a
driving simulator, we decided to use a Lisp-implementation
of a driving environment.
Model Specification
As mentioned, the cognitive model of human attention
integrated the three components control, monitoring and
decision making. They are implemented as a loop of
cognitive operations in the ACT-R serial processor.
The UML-Diagram in Fig. 4 shows the behavior of the
cognitive model. To execute the task drive, the model runs
through several states.
Fig. 4: UML-Diagram of the driver model
From the initial state, the model finds the road marks and
sets the near point for stable navigation on the road. The
model then fires a production rule screening for a traffic
sign, changes the state according to the result and sets the
far point. In our model, the near and far point are used as
control components and explained in detail in the next
paragraph. If the model reaches the state find far it can reach
277
the state overtake or will repeat the control loop. If there is
special state like an intersection, the model tests for other
given constraints and according to the result of this test, is
will either go to another special state or repeat the control
loop updating the near and far points.
A crucial advantage of the ACT-R architecture is that the
three components control, monitoring and decision-making
can be implemented directly. This takes into account human
constraints and results in a cognitive adequate model of
human attention.
Control
The control component of attention while performing a
driving task manages the perception of lower level visual
cues and the control over the vehicle (e.g., stopping). The
model uses the simple concept of two salient visual
attributes. This concept is based on earlier findings on
locomotion (Llewellyn, 1971) and steering. Further research
(Donges, 1987; Land & Horwood, 1995) describes steering
as divided in two levels, guidance and stabilization, by
using a „far“ and a „near“ region. Models of steering
developed under this assumption have been proven to be
consistent with empirical evidence.
The perception of this model is based on the perception of
two salient visual points (Salvucci & Gray, 2004), a near
and a far point. These two points are used for guidance,
stabilization and also, to observe other salient attributes. For
the here created artificial road environment, these two points
account to recognize relevant aspects in any situation which
may arouse during driving a car.
The near point determines the position on the road, which
is in the middle of the center line and the border line. To
identify the direction of driving, the far point is used and
usually set on the vanishing point on the horizon or on the
lead vehicle. The far point is also used to identify other
situations and can be set on non-control points like traffic
signs or approaching cars. Fig. 5 illustrates the near and far
points.
Fig. 5: Near and far points for a straight road with a
vanishing point and a road segment with a lead car.
The ACT-R architecture limits the employment of the
control component by using a serial cognitive processor.
The serial processing of the subtasks is typical for the
human bottleneck of information processing. The resulting
model is not an optimal model in a mathematical sense, but
approximates human behavior.
In a driving environment, the majority of lower level
visual control is keeping the vehicle in the middle of the
road lane, for which the near point is used. Although the far
point is used to identify traffic signs, it mainly indicates the
driving direction.
If the far point is not set on the vanishing point on the
horizon, the model uses the combination of near and far
point for determining the current scenario (see also Fig. 8
for an overview of implemented scenarios). If there is a lead
vehicle, the distance between the two points is determined,
and in case it falls below a certain safety distance, the model
can react according to that (e.g. through slowing down or
overtaking). In a crossroad scenario without an approaching
car from the right hand side, the model will set the far point
on the vanishing point of the horizon and continue driving.
After that, this model will not look again for another car at
the crossroad, which is surely an issue for future
implementations. In case there is a vehicle or a stop sign, the
stopping of the car is implemented here by setting the far
point onto the near point. The model will continue a loop
until the other vehicle is not on the crossroad anymore and
out of the safety distance.
Monitoring
After the control component, the monitoring is one of the
most important. Here, the environment is continuously
captured (e.g. the model looks for a traffic signs) and
updated in the declarative memory. In the here implemented
driving environment, the situation awareness mainly focuses
on other vehicles around, change of the scenario (from
straight road to crossroad), or traffic signs. The model shifts
the focus of visual attention towards a certain object which
is then encoded as visual attribute. The shift could be based
on a random-sampling model, checking the different
environment areas with a probability p, which has been
successfully done by Salvucci (2006). Here, the model
monitors particular directions and visual attributes (e.g.
other vehicles, center line) by an attention shift. The
encoded attribute is noted in the declarative memory. As
ACT-R has a build-in memory decay mechanism, it might
be possible to predict driver errors because the chunks
encoding the current environment decay and can be
forgotten if not updated continuously. Another source of
possible driver errors could be the potential failure in
encoding relevant information (e.g. to overlook a traffic sign
or a vehicle).
Decision Making
The information provided by the control and monitoring
component is used to determine if and what decisions must
be made on the tactical level concerning the maneuvering
(e.g. stopping or overtaking). The most common decision
making might be whether to stop or to continue driving.
This decision depends on the traffic sign or on other
vehicles. As described earlier, the execution of stopping
corresponds simply to the use of the near and far points
encoding current position and relevant aspects of the
environment. In order for the model to produce a decision
making process similar to humans, encoding a visual
attribute and shifting visual attention cannot occur at the
278
same time. For this model, the focus of attention is for
example either on the near or far point or encoding a traffic
sign. This restriction through serial processing seems to be a
drawback in the sense of mathematical optimal behavior,
but it describes the bottleneck typical for the human
information processing (Anderson et al, 2004). Through the
implementation of this restriction, it is possible to mimic
human cognitive capacities, simulate the dynamic nature of
human driving behavior, and therefore a cognitive adequate
model of human driving behavior is produced.
The knowledge representation comprehends declarative
knowledge in chunks and procedural knowledge in
production rules. For example, the scenario at a crossroad
was implemented in 73 explicit production rules, which are
highly detailed and is therefore open to future extensions of
the model. The control of attention in the ACT-R
architecture is achieved through three different methods of
shifting attention: First by specific locations or directions,
second by specific characteristics, and third by objects, that
have not been in in the focus of attention yet.
The combination of these methods of attention shift
enables the model to create complex search strategies
through the production rules.
Results and Discussion
We present a simulation environment and a cognitive model
of driver attention during car driving that is able to interact
during run-time. In this work, two driver models were
developed. The first model is able to reliably keep the traffic
lane on a two-lane road and initiate a lane change followed
by overtaking another vehicle. It identifies another vehicle
and decides to overtake it if the safety distance falls below a
certain distance (Fig. 6, scenario 1 and 2). The second
model builds up on the first model and extends its
functionality by identifying crossroad (Fig. 6, scenario 3),
traffic signs and vehicle on the right hand side which have
right of way (scenario 4, 5 and 6).
To obtain an integrated driver model of human driving
behavior, it is essential to develop models in an architecture
which is not task specific and can also model human
behavior also in a different context, like ACT-R. This model
is a first attempt to recognize, still simplified, traffic signs
and crossroads. The development of an integrated driver
model makes a first step towards the vision of accident-free
driving. A majority (over 80%) of the automobile accidents
are caused by the driver themselves. Fig. 1 shows the human
errors while driving. Nearly 16% of the accidents happen
while turning and during exit, followed by disregarding the
right of way (15%) and not-adapted speed (15%).
Theoretically, the cognitive driver model could give a
deeper insight for around 30% of the human errors while
driving. However, it has to be taken into account that the
model is still interacting with a simplified environment and
not yet taking into account driver´s prior experience, which
could be implemented by an increased attention in
potentially high accident risk situations. Our driver model is
one approach to integrate operational (lower-level) and
tactical (higher-level) aspects in the framework of the ACT-
R architecture. The model and the environment do not
present a complete picture of driver behavior yet, but they
form a base to extend the ETA framework in any direction.
Fig. 6: Standard situations while driving a car which can
be handled by the cognitive drive model.
The aspect of limited cognitive resources is one of the
main factors for the adequacy of the model. Based on the
implemented bottleneck, the three components control,
monitoring, and decision making, have to share cognition. If
the model is occupied with attention shift, it cannot
simultaneously update the near point. Also, the model can
only fire on production rule at a time and only one visual
operation can be executed at a time. These processes take a
certain time. For example, in the standard implementation in
ACT-R, one firing of a production rule requires 50ms. This
enables the researcher to compare the produced data with
human data, because the ACT-R architecture produces an
output file. This file contains the time, the active buffer and
the according event. This study did not validate the model
data so far. Future research could compare the output file
data with human data, specially compare the attention shift
of the model to human drivers over eye-tracking and the
reaction times. However, for this validation, it must be
possible from the technical side to either connect the ACT-R
model directly to the simulation environment or to produce
the same output file for the human data as the model does.
Also, only most critical parts of key scenarios can be
validated as no single method is sufficient enough to
understand the complex task of human driving behavior yet.
Conclusion and Outlook
The progress to date in the development of cognitive
architectures has been impressive, yet scientific gaps,
279
technical challenges and practical issues remain. On one
hand, cognitive models help to develop an understanding of
driver behavior and aim to provide a theoretical account for
human attention while driving. On the other hand, they are
powerful and practical tools when implementing human-
centered design and real-world applications. First steps
towards the examination of the source of human mistakes
through distraction from the primary driving task through
secondary tasks like dialing a phone haven been taken
(Salvucci, 2001) showing the feasibility of the architecture
for these task and possible extensions.
The ACT-R architecture enables to elucidate interesting
aspects and provides a theory of human attention while
driving. At the same time, human attention during driving is
a challenging task for the ACT-R cognitive architecture. It
shows the still existing limitations beyond basic laboratory
tasks and pushes the research community to expand the
architecture towards more complex and finally real-world
tasks.
Acknowledgments
We are grateful for the kind advice and assistance of Prof.
Bernhard Nebel, the support of Prof. Wolfram Burgard
(University of Freiburg), and for intensive discussions with
Dario Salvucci (University of Philadelphia). This work has
been partially supported by a grant from the DFG to MR
(Project R8-CSPACE in the SFB/TR8 “Spatial Cognition”).
References
Aasman, J. (1995). Modeling driver behavior in Soar. In:
Leidschendam, The Netherlands: KPN Research
Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S.,
Lebiere, C. & Qin, Y. (2004). An Integrated Theory of the
Mind. Psychological Review, 111, 1036
Anderson, J. R. & Lebiere, C. (1998). The atomic
components of thought. Mahwah, NJ: Lawrence Erlbaum.
Bellet, T., Bailly, B, Mayenobe, P., & Georgeon, O. (2007).
Modelling Driver Behavior in Automotive Environments.
Critical Issues in Driver Interactions with Intelligent
Transportation Systems. Cognitive Modelling and
Computational Simulation of Driver Mental Activities.
Benmimoun, A. (2004). Der Fahrer als Vorbild für
Fahrerassistenzsysteme? Ein fahrermodellbasierter Ansatz
zur Entwicklung von situationsadaptiven FAS. 13.
Aachener Kolloquium Fahrzeug- and Motorentechnik
Boer, E. R. (1996). Tangent point oriented curve
negotiation. IEEE Proceedings of the Intelligent Vehicles
96 Symposium
Boer, E. R. (1999). Car following from the driver`s
perspective. Transportation Research Part F, 2, 201-206
Byrne, M. D. (2001). ACT-R/PM and menu selection:
Applying a cognitive architecture to HCI. International
Journal of Human-Computer Studies, 55, 41-84
Donges, E. (1987). A two-level model of driver steering
behavior. Human Factors, 20, 691-707
Fajen, B. R., & Warren, W. H. (2003). Behavioral dynamics
of steering, obstacle avoidance, and route selection.
Journal of Experimental Psychology: Human Perception
and Performance, 29, 343-362
Gray, W. D., & Altmann, E. M. (2001). Cognitive modeling
and human-computer interaction. International
encyclopedia of ergonomics and human factors, 1, 387-
391, Taylor & Francis, Ltd.
Land, M., & Horwood, J.(1995). Which part of the road
guide steering? Nature, 3, 77, 339-340
Liu, Y. (1996). Queuing network modeling of elementary
mental processes. Psychological Review, 103, 116-136
Liu, Y., Feyen, R., & Tsimhoni, O. (2006). Queuing
Network-Model Human Processor (QN-MHP): A
computational Architecture for Multitasking Performance
in Human-Machine Systems. ACM Transactions on
Computer-Human Interaction 13, 37§70
Llewellyn, L. (1971). Visual guidance of locomotion.
Journal of Experimental Psychology , 91, 245-254
Michon, J. A. (1985). A critical view of driver behavior
models: What do we know, what should we do? Human
behavior and traffic safety, 48552, Plenum Press
Pomerlau, D., & Jochem T. (1996). Rapidly adapting
machine vision for automated vehicle steering. IEEE
Expert, 112, 19-27
Reid, L. D., Solowka, E. N.,& Billing, A. M. (1981). A
systematic study of driver behavior steering control
models. Ergonomics, 24 , 447-462
Rushton, S. K., Harris, J. M., Lloyd, M.R., & Wann J.P.
(1998). Guidance of locomotion on foot uses perceived
target location rather than optic flow. Current Biology, 8,
1191-1194
Salvucci, D. D. (2001). Predicting the Effects of In-Car
Interface Use on Driver Performance: An Integrated
Model Approach. International Journal of Human-
Computer Studies, 55, 85-107
Salvucci, D. D. (2006). Modeling Driver Behavior in a
Cognitive Architecture. Human Factors, 48, 362-380
Salvucci, D. D., Liu, A. , & Boer, E. R. (2001). Control and
monitoring during lane changes. Vision in Vehicles, 9
Salvucci, D. D., & Gray, R. (2004). A Two-Point Visual
Control Model of Steering. Perception, 33, 1233
Statistisches Bundesamt (online 20.12.2011)
www.destatis.de
Wilkie, R. M., & Wann, J. P. (2003). Controlling steering
and judging heading: retinal flow, visual direction and
extra-retinal information. Journal of Experimental
Psychology: Human Perception and Performance, 29,
363-378
280
... Therefore, multiple ACT-R modules are required to interact during modeling. Especially since (conditional) automated driving, Level 3, is prospected in near future, effects of attention and distraction (Haring, 2012), of driver drowsiness (Gunzelmann, et al., 2011), multi-tasking (Kosanke & Russwinkel, 2016) as well as insights on non-driving related tasks (Salvucci, 2009) have been modeled. Ultimately, combination of a human driver model in ACT-R with the possibility of a direct connection to a vehicle, to possibly adjust vehicle behavior according to the needs of the driver, is a promising vision. ...
Conference Paper
Full-text available
In this publication an adaptation of the ACT-R visual module is presented based on the SEEV theory on attention allocation. By including this theory into the methodology of how the visual module works, a top-down control of attention guidance and bottom-up processing capabilities were implemented. The visual field of the model shifts according to current fixations, mimicking human behavior. Finally, we introduce a possibility of linking this new visual module with environmental sensors of a vehicle to generate data for the model without the need of a modeler generating environmental data. As of now the interpretation of the environment could be visualized differently depending on the understanding of the modeler. Now, the modeler benefits by having a time efficient reproducible source for data generation for driver modeling.
... To select an assessment model that can be applied to the truck driver, it must understand the dynamics of truck driving and explain the relationship between the inner workings of the brain and body and its responses to specifi c situations (Crystal and Ellington, 2004) with some clarity as well as categorise the standard driver actions and comprehend the driver perception-action. This is important when examining the foundations for some of the models, as they appear to be based on preconceived ideas of how to drive the vehicle in the fi rst place (Haring et al., 2012;Taatgen, 2005). Meaning that if the model selected has these foundations, it will create more questions than provide answers for and any value to the assessment may be limited to the theoretical or simulated environment. ...
Article
Full-text available
Human factors assessment tools for truck driving, appear to overlook the driver’s relationship to the driving experience. Whilst most of the assessment tools are useful and have a degree of value, there are factors that have not been considered. To completely decompose truck driving and identify all the elements will require a purpose-built assessment tool that aligns itself with normal driving behaviour must demonstrate how a truck driver moves between each series of sub-tasks, as well as making allowances for any issues or consequences that may eventuate from undertaking any action.
Article
Objective The primary objective of the study was to evaluate the effect of a secondary motor task induced by an interactive seating system (IASS) on passive driver fatigue in a monotonous simulated driving task. The effect was compared to that of a state-of-the-art massage seating system (MS), which may reduce monotony through additional tactile stimuli. The secondary objective was to compare the user experience of both systems. Method The independent variables were three conditions: one with the IASS, another with the MS, and a control without intervention. The study included seven dependent variables in total: a rating of subjective fatigue, three parameters measuring lane keeping ability, and three parameters reflecting fatigue-related eye movements. The duration of the simulator ride was 40 min in each condition. The study included thirty-five subjects. Results The assessment of subjective fatigue and lane keeping showed that the use of the IASS resulted in significantly lower passive driver fatigue compared to the massage and control conditions. The alerting effects of the IASS were also reflected by an increased eyelid distance. Frequency and duration of blinks, however, showed no clear patterns of fatigue over time in any of the conditions. Thus, both parameters did not seem be suitable to capture passive driver fatigue in this study. Regarding user experience, the subjects preferred the IASS over the MS as well. Conclusion The IASS showed a strong potential as an effective measure against passive driver fatigue within monotonous driving situations. The MS, on the other hand, induced no measurable effects.
Preprint
Improvements in autonomy offer the potential for positive outcomes in a number of domains, yet guaranteeing their safe deployment is difficult. This work investigates how humans can intelligently supervise agents to achieve some level of safety even when performance guarantees are elusive. The motivating research question is: In safety-critical settings, can we avoid the need to have one human supervise one machine at all times? The paper formalizes this 'scaling supervision' problem, and investigates its application to the safety-critical context of autonomous vehicles (AVs) merging into traffic. It proposes a conservative, reachability-based method to reduce the burden on the AVs' human supervisors, which allows for the establishment of high-confidence upper bounds on the supervision requirements in this setting. Order statistics and traffic simulations with deep reinforcement learning show analytically and numerically that teaming of AVs enables supervision time sublinear in AV adoption. A key takeaway is that, despite present imperfections of AVs, supervision becomes more tractable as AVs are deployed en masse. While this work focuses on AVs, the scalable supervision framework is relevant to a broader array of autonomous control challenges.
Chapter
In an optimally integrated HMS (Human Machine Systems) human and machine understand each other to provide an optimum integration. This is one of the core principles which is applicable for the research frameworks in vehicle navigation domain for effectively conducting research for creating optimal guidance information for the human driver. Creation and integration of human cognitive models for navigation is necessary to follow this principle effectively. BeaCON: Behaviour-and Context-Based Optimal Navigation is an existing research framework in the car navigation domain, for conducting analysis for the research problem “Giving the driver adequate navigation information with minimal interruption”. Currently BeaCON does not use the human cognitive models for navigation for the creation of guidance information and because of that the integration with the human driver is not achieved to an optimum level. In this paper, we present enhancement of BeaCON by integrating behaviour and cognitive models of navigation. Understanding the human thoughts while driving enables BeaCON to have a granular analysis of user cognitive state while creating guidance information, which results further cognitive load reduction for navigation tasks by creating more effective guidance information.
Chapter
This chapter describes model of the human cognitive functions, especially these ones which are important for the creative process. The broad context of this work is a development of the conceptual redesign method with computer support (called CRDP—Computer Redesign Process). This method is based on postmodern principles of the interpretation, on the respect to complexity of the creative process and at the impossibility of its direct control. Psychological approaches (e.g. the mind mapping [1] or creativity timing) are used in this interpretation method and its core is a creating of the interpretation map. The aim of this submission is to describe the emergent design processes for the purpose of their simulation and method’s HCI improving. The model is made in Unify Modelling Language. The fractal approach to the communication between user (designer) and software system (CRDP) was outlined.
Article
Full-text available
Queueing Network-Model Human Processor (QN-MHP) is a computational architecture that integrates two complementary approaches to cognitive modeling: the queueing network approach and the symbolic approach (exemplified by the MHP/GOMS family of models, ACT-R, EPIC, and SOAR). Queueing networks are particularly suited for modeling parallel activities and complex structures. Symbolic models have particular strength in generating a person's actions in specific task situations. By integrating the two approaches, QN-MHP offers an architecture for mathematical modeling and real-time generation of concurrent activities in a truly concurrent manner. QN-MHP expands the three discrete serial stages of MHP, of perceptual, cognitive, and motor processing, into three continuous-transmission subnetworks of servers, each performing distinct psychological functions specified with a GOMS-style language. Multitask performance emerges as the behavior of multiple streams of information flowing through a network, with no need to devise complex, task-specific procedures to either interleave production rules into a serial program (ACT-R), or for an executive process to interactively control task processes (EPIC). Using QN-MHP, a driver performance model was created and interfaced with a driving simulator to perform a vehicle steering, and a map reading task concurrently and in real time. The performance data of the model are similar to human subjects performing the same tasks.
Article
A sequence of driving tasks has been carried out in a driving simulator. The initial tests represented lane tracking along a serpentine roadway and were employed to verify the operation of the simulator and the ability of a computer algorithm to fit linear driver models to experimental data. A second series of tests involved an obstacle avoidance manoeuvre in both a car and a truck. These latter simulator runs were augmented by field trials in an automobile during which driver eye point-of-regard data were recorded. Eye point-of-regard results from both simulator and field trials were compared and employed in formulating a simple driver model for the obstacle avoidance manoeuvre. The results from a preliminary fitting of this model to the experimental data are reported. It was found that a single linear model of the driver's dynamic characteristics can be used to represent adequately all of the driver response data measured in the present study.
Article
In evaluating the performance of the driver-vehicle system and in establishing design criteria for favorable vehicle dynamics, a quantitative description of driver steering behavior such as a mathematical model is likely to be helpful. The steering task can be divided into two levels: (1) the guidance level involving the perception of the instantaneous and future course of the forcing function provided by the forward view of the road, and the response to it in an anticipatory open-loop control mode; (2) the stabilization level whereby any occuring deviations from the forcing function are compensated for in a closed-loop control mode. This concept of the duality of the driver's steering activity led to a newly developed two-level model of driver steering behavior. Its parameters were identified on the basis of data measured in driving simulator experiments. The parameter estimates of both levels of the model show significant dependence on the experimental situation which can be characterized by variables such as vehicle speed and desired path curvature.
Article
The major categories of models of the past two decades are reviewed in order to pinpoint their strengths - and perhaps their weaknesses - in that framework. This review includes such models as McKnight & Adams' task analysis, Kidd & Laughery's early behavioral computer simulations, the linear control models (such as McRuer & Weir's), as well as some more recent concepts such as Naeaetaenen & Summala's, Wilde's and Fuller's risk coping models which already carry some cognitive weight. Having proposed my answers to these questions an attempt is made to formulate an alternative approach, based on production systems as developed by J. R. Anderson.
Article
Andrew Liu, MIT Man Vehicle Laboratory (andy@space.mit.edu) Erwin R. Boer, Erwin R. Boer Consulting (BorromeoRing@aol.com) This work examines the visual scanning behavior of drivers in preparing and executing a lane change. In an experiment, drivers navigated a naturalistic highway environment with moderate traffic in a fixed-base driving simulator. The data collected in the experiment included steering control data, eye movements, and verbal protocols. Drivers began to exhibit significantly different scanning behavior approximately three seconds before initiation of the lane change, showing increasingly more frequent gazes at the rear-view mirror at the expense of gazes to their current lane. As soon as they decided to make a lane change (as indicated in their verbal protocols), drivers shifted their gaze from salient guiding features of the current lane (e.g., tangent point or lead vehicle) to salient guiding features of the destination lane. In addition, drivers exhibited increased gazes at surrounding vehicles (front and back) before and during lane changes for the purposes of situation awareness and decision making. The results support a dual-purpose view of driver gazes for control and monitoring.
Article
In this commentary, it is argued that the car following models discussed in Brackstone, M., and McDonald, M. (Transportation Research – Part F (2000), pp. 181–196) ignore one or more of the following issues that characterize to observed driver behavior. These include: (i) car following is only one of many tasks that drivers perform simultaneously and receives therefore only intermittent attention and control (task scheduling/attention management), (ii) drivers are satisfied with a range of conditions that extend beyond the boundaries imposed by perceptual and control limitation (satisficing instead of optimal performance evaluation), and (iii) in each driving task drivers use a set of highly informative perceptual variables to guide decision making and control (perceptual rather than Newtonian input). To elucidate these issues, a general driver modeling framework is presented in which the car-following task is highlighted (Boer, E. R., & Hoedemaeker, M. (1998). In Proceedings of the XVIIth European Annual Conference on Human Decision making and Manual Control December 14–16. France: Valenciennes; Boer, E. R., Hildreth, E. C., & Goodrich, M. A. (1998). In Proceedings of the XVIIth European Annual Conference on Human Decision making and Manual Control December 14–16. France: Valenciennes).
Article
Understanding the interaction of a user with a designed device such as a GUI requires clear understanding of three components: the cognitive, perceptual and motor capabilities of the user, the task to be accomplished and the artefact used to accomplish the task. Computational modeling systems which enable serious consideration of all these constraints have only recently begun to emerge. One such system is ACT-R/PM, which is described in detail. ACT-R/PM is a production system architecture that has been augmented with a set of perceptual-motor modules designed to enable the detailed modeling of interactive tasks. Nilsen's (1991) random menu selection task serves two goals: to illustrate the promise of this system and to help further our understanding of the processes underlying menu selection and visual search. Nilsen's original study, two earlier models of the task, and recent eye-tracking data are all considered. Drawing from the best properties of the previous models considered and guided by information from the eye-tracking experiment, a series of new models of random menu selection were constructed using ACT-R/PM. The final model provides a zero-parameter fit to the data that does an excellent, though not perfect, job of capturing the data.
Article
While researchers have made great strides in evaluating and comparing user interfaces using computational models and frameworks, their work has focused almost exclusively on interfaces that serve as the only or primary task for the user. This paper presents an approach of evaluating and comparing interfaces that users interact with as secondary tasks while executing a more critical primary task. The approach centers on the integration of two computational behavioral models, one for the primary task and another for the secondary task. The resulting integrated model can then execute both tasks together and generate a priori predictions about the effects of one task on the other. The paper focuses in particular on the domain of driving and the comparison of four dialing interfaces for in-car cellular phones. Using the ACT-R cognitive architecture (Anderson & Lebiere, 1998) as a computational framework, behavioral models for each possible dialing interface were integrated with an existing model of driver behavior (Salvucci, Boer & Liu, in press). The integrated model predicted that two different manual-dialing interfaces would have significant effects on driver steering performance while two different voice-dialing interfaces would have no significant effect on performance. An empirical study conducted with human drivers in a driving simulator showed that while model and human performance differed with respect to overall magnitudes, the model correctly predicted the overall pattern of effects for human drivers. These results suggest that the integration of computational behavioral models provides a useful, practical method for predicting the effects of secondary-task interface use on primary-task performance.
Article
Deze studie beschrijft een cognitief model van het gedrag van een automobilist bij het naderen en afhandelen van kruisingen. Het model bestuurt een gesimuleerde auto in een gesimuleerde verkeerswereld waarin ook andere auto's en fietsers rondrijden. ... Zie: Samenvatting