Scholarly article on topic 'Modeling Driver Response to Imperfect Vehicle Control Automation'

Modeling Driver Response to Imperfect Vehicle Control Automation Academic research paper on "Materials engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Manufacturing
OECD Field of science
Keywords
{"Vehicle control automation" / "Adaptive cruise control" / "Driver behavior" / "Transfer of control"}

Abstract of research paper on Materials engineering, author of scientific article — Bobbie D. Seppelt, John D. Lee

Abstract Drivers are required to monitor multiple critical variables in a dynamic, uncertain environment. Increasingly automated vehicles relieve drivers of the demand of moment-to-moment control but impose attentional demands associated with the need to monitor the status and behavior of the automation. Adaptive cruise control (ACC) is such a system – it automates headway maintenance. Because ACC operates effectively in only a subset of driving situations, drivers must monitor the system and intervene periodically. This paper describes a computational model to assess driver-ACC interaction. In this model, the driver is modeled according to a set of state transition diagrams (applying the modeling formalism developed by Degani & Heymann, 2002, 2007), which represent a driver's mental model (i.e., understanding) of ACC as developed through its interface. This mental model combines with a set of response rules, based on perceptual thresholds, which guide driver response in failure situations. The ACC is modeled as a two-level longitudinal control algorithm (Zheng & McDonald, 2005). The interaction of the driver and ACC predict reaction times for drivers intervening during ACC failures. The model predicts collision outcomes for a set of ACC use cases, which include scenarios within and outside normal ACC operating conditions. As identified from the model, there are a number of situations in which drivers’ inadequate understanding of ACC result in collisions. The findings indicate the need to inform drivers of ACC behavior in a way that allows them to anticipate situations prone to system failure, i.e., as based on an adequate understanding of system boundary conditions and appropriate situations of use. Implications of computational models for assessing driver-automation interaction for higher levels of automation are discussed.

Academic research paper on topic "Modeling Driver Response to Imperfect Vehicle Control Automation"

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedia Manufacturing 3 (2015) 2621 - 2628

6th International Conference on Applied Human Factors and Ergonomics (AHFE 2015) and the

Affiliated Conferences, AHFE 2015

Modeling driver response to imperfect vehicle control automation

Bobbie D. Seppelta *, John D. Leeb

aTouchstone Evaluations, Inc., 18160 Mack Avenue, Grosse Pointe, MI 48230, USA bUniversity of Wisconsin, Department of Industrial and Systems Engineering, 1513 University Avenue, Madison, WI53760, USA

Abstract

Drivers are required to monitor multiple critical variables in a dynamic, uncertain environment. Increasingly automated vehicles relieve drivers of the demand of moment-to-moment control but impose attentional demands associated with the need to monitor the status and behavior of the automation. Adaptive cruise control (ACC) is such a system - it automates headway maintenance. Because ACC operates effectively in only a subset of driving situations, drivers must monitor the system and intervene periodically. This paper describes a computational model to assess driver-ACC interaction. In this model, the driver is modeled according to a set of state transition diagrams (applying the modeling formalism developed by Degani & Heymann, 2002, 2007), which represent a driver's mental model (i.e., understanding) of ACC as developed through its interface. This mental model combines with a set of response rules, based on perceptual thresholds, which guide driver response in failure situations. The ACC is modeled as a two-level longitudinal control algorithm (Zheng & McDonald, 2005). The interaction of the driver and ACC predict reaction times for drivers intervening during ACC failures. The model predicts collision outcomes for a set of ACC use cases, which include scenarios within and outside normal ACC operating conditions. As identified from the model, there are a number of situations in which drivers' inadequate understanding of ACC result in collisions. The findings indicate the need to inform drivers of ACC behavior in a way that allows them to anticipate situations prone to system failure, i.e., as based on an adequate understanding of system boundary conditions and appropriate situations of use. Implications of computational models for assessing driver-automation interaction for higher levels of automation are discussed.

© 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review underresponsibility ofAHFEConference

Keywords: Vehicle control automation; Adaptive cruise control; Driver behavior;Transfer of control

* Corresponding author. Tel.: 1-313-458-8077; fax: 1-313-429-7524. E-mail address: bseppelt@touchstoneevaluations.com

2351-9789 © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of AHFE Conference

doi: 10.1016/j.promfg.2015.07.605

1. Introduction

Drivers must monitor multiple critical variables in a dynamic, uncertain environment. They need to predict how these variables change to know how to adapt their behaviour to effectively interact with automated systems. An important concern is how drivers manage their attention to automated systems and to the driving task.

Technology that automates vehicle control relieves drivers of the demand of moment-to-moment control but imposes attentional demands associated with the need to monitor the status and behaviour of the automation [1]. Adaptive cruise control (ACC) is such a system - it automates headway maintenance. Because ACC operates effectively in only a subset of driving situations, drivers must monitor the system and intervene periodically. Sensors positioned on the front of the vehicle detect preceding vehicles and determine their range and speed. If no vehicle is detected, the ACC operates like a conventional cruise control system, maintaining a set speed. If a preceding vehicle is detected, the ACC modulates vehicle speed to maintain a preset time headway. Typically, ACC operates only for a limited range of speeds and braking authority is limited to less than 0.3 g.

The dynamics of driving and ACC features combine to surprise drivers. Such surprises are often a product of inadequate driver understanding of ACC behaviour and limits. Situations that can lead to surprises include drivers: 1) trusting the ACC more than warranted, 2) engaging the system in situations beyond its original design parameters, and 3) failing to recognize when the system is behaving in ways that are contrary to their expectations [2, 3]. Nilsson [4] introduced drivers to three critical scenarios that required driver intervention. These scenarios were approaching a stationary queue of traffic, a car pulling out in front of the driver's vehicle, and a hard braking lead vehicle. Drivers over-relied on the ACC and delayed their response when confronted with the first scenario, resulting in near collisions. Rudin-Brown and Parker [2] observed similar decrements to hazard response with use of ACC; drivers maintained high levels of trust in the hazard situation and performed better in a secondary task, indicating that they over-trusted and over-relied on the ACC. Hoedemaeker [5] showed that drivers failed to recognize the limitations of the ACC for overtaking manoeuvres on rural roads and in detecting traffic coming from the right at an intersection. Stanton, Young, & McCaulder [3] showed ACC surprises also occur when sensors either do not detect targets in the path of the vehicle (e.g., motorcycle) or respond to non-vehicle targets (e.g., bicyclist or pedestrian). These studies show the need for drivers to understand the limits of ACC and when they must disengage the system. The feedback available from current ACC systems is often inadequate to help drivers develop appropriate mental models that will enable them to understand its limitations, and to predict ACC behaviour [3, 6, 7].

This paper presents a modelling analysis intended to develop a deeper understanding of the driver-ACC algorithm interaction. This analysis is expected to reveal situations in which a driver's inadequate understanding of the ACC's function and operating capabilities will increase collision likelihood. Such situations indicate the need to properly inform drivers of the ACC's behaviour. The analyses included in this paper address the following question: If ACC fails temporarily, what is the consequence to driving safety and how will a driver respond as a function of their understanding?

1.1. Modelling approach

Following the modelling formalism developed by Degani and Heymann [8, 9], we model driver-ACC interaction with state transition diagrams. These diagrams represent drivers' mental model (i.e., understanding) of ACC, as developed through its interface and its response to roadway events. These diagrams combine with a set of response rules, based on perceptual thresholds, to predict driver response in failure situations. The ACC is modelled using a two-level longitudinal control algorithm [10].

1.2. Driver component

Drivers' mental models of automation are a function of a system's purpose and form, its operating procedures, and associated state structure [11, 12]. Mental models are constructed from information provided by the interface, observed system behaviour, and instructions [13]. These mental models guide attention allocation [1], directing operators where and when to look for information regarding the automation and the system being controlled [14, 15]. Because a mental model represents understanding, shaping operator prediction of automation behaviour and

control responses [14, 16], compatibility must exist between the physical system itself, the mental model, and the interface between the two. Degani and Heymann [8, 9] developed a formal approach to describe human-automation interaction. In their work, state transition diagrams model the internal configuration of automation and the state changes that occur in response to events that trigger them. As an operator interacts with automation through an interface by observing and triggering state transitions, the operator forms a model. Essentially, the information provided within an interface defines the driver's mental model.

State transition model of ACC. Fig. 1a depicts a state transition model of ACC. This model includes four states: speed control, distance (or following) control, off, and standby. The transitions to and from these states, which represent discrete state changes, are triggered either by the driver (depicted in Fig. 1 as dashed lines) or automatically (depicted in Fig. 1 by solid lines). Fig. 1b shows the detailed view of the states and transitions internal to the 'ACC active' state. Transition lines in Fig. 1a and b are labelled by the event that triggers the state change. The state transition models in Fig. 1 are consistent with the ISO standard for performance requirements and test procedures for Adaptive Cruise Control Systems [17].

Fig. 1. (a) ACC states and transitions. Dashed lines represent driver-triggered transitions, solid lines - ACC-triggered transitions; (b) detailed states and transitions within 'ACC speed control; and 'ACC distance control'.

Automation interfaces present abstracted information concerning the underlying automation state, structure, and response; they represent the automation to the operator. It is therefore important that an interface captures the automation's behaviour adequately. When the model of automation conveyed by the interface differs from the actual state of the automation, breakdowns in human-automation interaction can occur. The ISO standard for ACC systems is to provide drivers at a minimum an indication of the ACC's activation state (ACC system is active, not active, or off), and set speed. One such an interface is a display of set speed that is only visible when ACC is active. Drivers are also to be informed if the ACC system is not available due to a failure. An example driver model of the ACC system derived from use of the controls and through an interface containing the ISO-specified feedback is shown in Fig. 2a. This model illustrates the disparity that can occur from an over-simplification of the actual complexity of the 'ACC active state' in the information provided in the ACC interface. A visual comparison of the ACC system model (Fig. 1) and the driver model (Fig. 2a) indicates that there are a number of unobserved events. Unobserved events are those that are not displayed in the interface and consequently not represented in the operator model [8, 9]. Unobserved ACC events include: sensor limits, braking algorithm limits, and most of the transitions between follow and resume modes internal to 'ACC active'. Without such information, drivers are likely to develop inaccurate mental models [18].

Composite state transition model. A composite model reveals inconsistencies of the driver model to that of the system model [8, 9]. To create a composite model, the driver model states and ACC model states are combined into state pairs. If these state pairs are inconsistent, i.e., the driver's model indicates that ACC is in a state that does not

Press 'ftsssl' or 'Coast' button

Press Time gap+'or ____%

"Time gap-J

ACC active ACC active

ACC system fault detected

ACC not active

ACC deceleration required > 0.2 g or Current ACC vehicle speed < 25 mph

ACC ACC

off standby

Press Set speed' or 'Resume' button

ACC ACC

off off

»-- Press'Off button ---- Press 'On' button

! ACC not ACC

active standby

Fig. 2. (a) driver model of ACC; (b) composite of the driver model (black boxes) and ACC system model (white boxes). Error states are indicated with a vertical hashed border.

match the actual state of the ACC, there is an inaccuracy in the driver's model, referred to as an error state. Error states in Fig. 2b are indicated with a vertical hashed border. The 'ACC not active - ACC off' and 'ACC off - ACC standby' error states result because the interface does not differentiate between standby and off states. The 'ACC active - ACC standby' error state results from missing feedback on the conditions that automatically deactivate the ACC system.

To communicate the ACC system model and to inform drivers of the ACC's behaviour, an ACC interface should indicate set speed, current speed, set headway, current headway, ACC activation (on, off, active, standby), if the LV is detected, and the events/conditions that trigger transitions, i.e., the boundary of the ACC operating conditions. An inaccurate mental model, with missing transitions between states or missing states, leads drivers to incorrect expectations regarding the behaviour of the ACC.

Control response thresholds. The lower bound for a driver's control response (i.e., disengaging ACC) in response to an ACC failure (operationally defined as a loss of ACC's set following distance) is at the point where an event is first perceptible - a just-noticeable difference (JND). Essentially, such a response occurs provided the perception of a change in a particular variable. For dynamic visual objects, the JND is based on the visual angle subtended by the object at the eye of the following driver (O) and the rate of change of this visual angle (dO/dt). These quantities are defined, according to Hoffman and Mortimer [19], as:

O = W/H (1)

dO/dt = 6 = WVJH2 (2)

where W is the width of the object, Vr is the relative velocity between the driver's vehicle and the object (where "object", as applied in this context, refers to a lead vehicle) and H is the distance between the following driver's eye and the rear of the object. The visual angle and the angular velocity subtended by the observed object are perceptible above the thresholds of approximately 0.017 degrees [20] and 0.17 degrees/second [19], respectively.

In driving, visual angle and expansion rate are effective information sources for detecting deceleration of a lead vehicle only when they exceed these thresholds. The perceptual system is flexibly attuned to specific properties of the task environment in response to a looming event or stimulus, whereby angle and expansion rate strategies are used effectively either alone or in combination as perceptual response thresholds [21]. Importantly, simply perceiving a lead vehicle or perceiving that it is slowing might not provoke a response. Another type of response threshold pertinent to driving is that of a threat threshold. Time-to-collision (TTC) is often applied as a safety indicator for longitudinal control using a determined minimum TTC, e.g., [22, 23].

where R indicates range from the driver's vehicle to the lead vehicle. TTC indicates collision potential, being inversely related to accident risk (higher TTC values indicate a higher accident risk and vice versa). A minimum TTC of 4 seconds is often cited as the boundary discriminating instances in which drivers unintentionally find themselves in a dangerous situation and those in which they remain in control [24].

The value of the response threshold at which a driver will initiate a control response depends on response strategy, which depends on the driver's mental model of the ACC. For instance, a driver with an inaccurate mental model is likely to initiate a control response based on their JND of a driving event that signals an ACC failure (e.g., a looming lead vehicle) rather than on their understanding of how a particular change in driving condition affects ACC performance. The use of response thresholds with values for which drivers initiate a response allows the model to predict driver behaviour.

1.3. ACC component

The control aspects of ACC that determines its response behaviour (to trigger state transitions) are briefly described in this section. The longitudinal control model from Zheng and McDonald [10] was adopted to simulate ACC behavior:

where XLV is the rate of change of acceleration of the LV as a function of time. Based on the results from Zheng and McDonald [10], Kv and Kd are set to 0.75 and 0.1, respectively. The time headway (Th) is set to 1.5s. The maximum acceleration of the ACC system is 0.2g.

1.4. Model outputs

The model generates driver reaction time to failures of the ACC system to maintain the set THW, and a composite model profile, which indicates mismatches between the state changes tracked by the driver and state changes of the ACC. The model also generates velocities and relative positions of the lead vehicle and the driver's vehicle for the duration of a driving event, allowing for calculation of collision frequency and other driving performance metrics.

The model used for these analyses was developed using Simulink 6.0 and Stateflow, 6.0, as part of MATLAB 7.0 (Mathworks Inc., Natick, MA). All parameters relating to the driver component and ACC component were controllable. The driver model and system model of ACC were implemented using Stateflow. The control law for ACC (formula 4) was implemented using Simulink. Initial scenario conditions were input into the ACC component, which output an acceleration profile for the driver vehicle. This profile and scenario parameters were input into the driver model and system model to dynamically transition through their respective states and transitions. Driver reaction time to ACC failure events were calculated from the point of failure to the control response, as initiated dependent on the particular response strategy. The independent variables examined in this study were event type, driver model, and control response. The purpose of the driver-ACC interaction analysis is to reveal how interface design and driving conditions lead to driver-ACC interaction failures.

Each event type is characterized by lead vehicle (LV) acceleration, LV velocity, driver vehicle (DV) velocity, initial time headway (THW) to LV, initial range rate to LV, set velocity, and an ACC sensor operation profile (see Table 1). The event conditions resulted in differing ACC response behaviours at failure: 1-second maximum deceleration followed with transition to standby, acceleration to set velocity within active state, and immediate transition to standby for braking, sensing, and setting failures, respectively. In their unique behavioural implications to ACC, these event conditions test the specificity of a driver's mental model. Further, they represent the situational

2. Method

factors and vehicle dynamics reported in the literature as problematic to drivers in use of ACC (e.g., degraded sensors: [3]; braking limit exceedances: [4]; offset LV outside sensor field: [3]; lateral incursion of a LV not detected by sensor field: [10]; set velocity deviant from LV velocity: [5]). For all event types, the initial conditions describe a following situation with a 1.5 s THW at equal velocity. LV deceleration rate and duration, sensor failure time, and velocity difference to set speed are varied for the braking, sensing, and setting event types, respectively.

Two driver models of ACC were implemented in Stateflow for the driver component of the interaction model: 1) an inaccurate driver model as predicted to result from a standard ACC interface (consistent with Fig. 1); and 2) an accurate driver model as predicted to result from an ACC interface that informed drivers of its activation state, settings, and operating limits.

Five control response thresholds were tested: 1) no response (a control condition); 2) JND of expansion rate; 3) TTC = 4 s; 4) state change from inaccurate model; 5) state change from accurate model. To more accurately model driver response time to ACC failures, a 1.5 s time delay was added to each control response calculation of reaction time to account for cognitive processing and movement time to initiate a brake response [25]. The driver-ACC interaction model generated driver reaction time to ACC failure (mean; standard deviation) and collision outcome.

Table 1. Event inputs to model of driver-ACC interaction.

Driving Event Type Description of Driving, Initial Conditions, and ACC Response Number of Conditions

Braking limit exceedence THWo = 1.5 s; RRo = 0.0 m; dLV = 0.0 : 0.01 : 0.85 g 8,686

Set Vo = 24.59 m/s [55 mph] Tdecel = 0 : 0.1 : 10 s

DV Vo = : 22.35 m/s [50 mph] Tf = 0 s

LV Vo = 22.35 m/s [50 mph] Individual condition run time: 60 s (at 60 Hz)

Degraded sensor / Lateral THWo = 1.5 s; RRo = 0.0 m; dLV = 0.0 : 0.01 : 0.85 g 26,058

range limit exceedence Set Vo = 24.59 m/s [55 mph] Tdecel = 0 : 0.1 : 10 s

DV Vo = : 22.35 m/s [50 mph] Tf = 1, 5, 10 s

LV Vo = 22.35 m/s [50 mph] Individual condition run time: 60 s (at 60 Hz)

Minimum velocity setting THWo = 1.5 s; RRo = 0.0 m; dLV = 0.0 : 0.01 : 0.85 g 8,686

exceedence Set Vo = 13.41 m/s [30 mph] Tdecel = 0 : 0.1 : 10 s

DV Vo = : 13.41 m/s [30 mph] Tf = 0 s

LV Vo = 13.41 m/s [30 mph] Individual condition run time: 60 s (at 60 Hz)

THW = time headway; RR = range rate; DV = driver's vehicle; LV = lead vehicle; Vo = initial velocity; dLV = deceleration rate of LV; Tdecei = time of LV deceleration; TF = Sensor failure time; aACc = acceleration response of ACC. Note: Numbers nested between two numbers with colons on either side are the step sizes

3. Results

The total number of event conditions produced from the initial conditions listed in Table 2 was 43,430. A driver reaction time was calculated for each of these initial conditions for each of the five control responses. Table 2 summarizes these statistics by event type and control response. These results are discussed in response to the analysis question posed in the Introduction: If ACC fails temporarily, what is the consequence to driving safety and how will a driver respond as a function of their understanding?

Table 2. Summary of model findings.

Collisions No Response Expansion Rate TTC=4 Inaccurate Model Accurate Model

Braking 0.54 0.527 0.183 0.463 0

Sensing 1 0.591 0.578 0.181 0 0

Sensing 5 0.785 0.768 0.366 0 0

Sensing 10 0.903 0.889 0.366 0 0

Settings 0.636 0.603 0.162 0.472 0

Mean RT

Braking - 7.024 3.967 4.524 3.063

Sensing 1 - 6.649 3.603 1.928 0.428

Sensing 5 - 7.565 3.744 2.001 0.501

Mean RT

Sensing 10 Settings

Standard Deviation of RT

10.973 5.856

4.59 3.717

2.001 3.633

0.501 2.438

Braking Sensing 1 Sensing 5

0.231 0.247 0.314 0.3 0.24

0.03 0.015 0.361 0.55 0.007

0.049 0

0.014 0.014 0.267

0.125 0

0.014 0.014 0.134

Sensing 10 Settings

For each event type, the 'No Response' column in Table 2 lists the percent of conditions that result in a collision ('Sensing 1', 'Sensing 5', and 'Sensing 10' denote the duration of the sensor failure in seconds). From these findings, it is evident that ACC cannot respond to many LV following situations. The braking capabilities of ACC can accommodate only half of the high-velocity LV braking events. As the sensor failure durations increase, collisions become inevitable for many LV braking events. The prevalence of situations in which ACC's limits are exceeded indicates the need for drivers to remain vigilant to its moment-to-moment state and behaviour. Drivers who rely on their perceptual abilities to detect and respond to failures of ACC are likely to over-estimate its ability to operate within its limits. A TTC=4 response threshold resulted in the fastest reaction times across event types and consequent lowest number of collision situations. Numerous event conditions still exceeded drivers' ability to respond safely when the model relies on this threat boundary as a threshold for when to initiate a control response. For the inaccurate driver model, control responses were initiated at the point of ACC's state change from 'On' to 'Standby' when braking and setting limits were exceeded. In this model, it is assumed that drivers would receive discrete warnings to indicate braking and setting limits; notably though, they would not prevent collisions in all situations. In fact, initiating control responses based on TTC is a better response strategy than relying on warnings that notify of limit exceedences, resulting in faster RTs and consequent fewer collisions. For situations when drivers are notified of ACC's sensing limits (the 'Sensing' rows in the 'Inaccurate Model' column in Table 2), drivers are able to avoid collisions. For the inaccurate driver model, however, drivers were unaware of ACC's sensing limits, and therefore remained in the 'On' state during the sensing failures. Drivers are expected to remain in this state until either a perceptual or threat threshold is crossed alerting them to their incorrect expectation of ACC's state. The 'Accurate Model' column in Table 2 lists the predicted RTs and consequent collision outcomes that result with use of an accurate model. Such a model would require an interface that primes drivers to the conditions that lead to ACC failure. For all tested event conditions, an 'Accurate' model prompted control responses that avoided collision.

The analysis indicated that ACC failures that occur during normal use of the system in common driving situations could result in crashes. Because of the limited time to respond to ACC failures, it is important for drivers to understand the ACC and intervene prior to its failure. A response strategy of braking when TTC exceeds 4.0 seconds, initiated after a failure, often fails to avoid collisions. Drivers who rely on their perceptual capabilities and fail to develop accurate models of the ACC are at risk for collision. Drivers must anticipate situations that will lead to failure based on an adequate understanding of how the ACC interacts with the driving environment.

In-vehicle support displays or training on ACC function and operating capabilities help build adequate driver understanding of ACC [26]. Examples of support displays are warnings that indicate when the ACC's capabilities are exceeded or informing displays that indicate the ACC's behaviour relative to operating limits in real-time [27]. Information on function and intention, by making automation more reliable and predictable, promotes trust [28]. Further, feedback that informs of the automation's behaviour in a variety of situations promotes improved mental model accuracy [7]. Taken together, feedback on the ACC's purpose, process, and performance would support adequate understanding. However, in a visually-demanding, multi-task domain like driving, any provided feedback would need to be both peripheral and easily understood to avoid unintended distraction from the driving task.

Increasingly automated vehicles, particularly those that relieve drivers of both lateral and longitudinal control tasks, may lead drivers to disengage from driving [29]. Initial studies on transfer of control indicate drivers may need at least seven seconds to resume control [30]. Informing drivers of an automated system's behaviour and

4. Discussion

conditions in which they may need to resume control is crucial for drivers to develop accurate mental models to enable timely interventions.

References

[1] Moray, N. (1986). Monitoring Behaviour and Supervisory Control. In K.R. Boff, L. Kaufman, and J.P. Thomas (Eds.), Handbook of

Perception and Human Performance (Vol. 2, Chapter 40). New York: Wiley.

[2] Rudin-Brown, C. M., & Parker, H. A. (2004). Behavioural adaptation to adaptive cruise control (ACC): implications for preventive strategies.

Transportation Research Part F: Traffic Psychology and Behaviour, 7(2), 59-76.

[3] Stanton, N.A., Young, M.S. & McCaulder, B. (1997). Drive-by-wire: The case of driver workload and reclaiming control with adaptive cruise

control. Safety Science, 27(2/3), 149-159.

[4] Nilsson, L. (1995). Safety effects of adaptive cruise control in critical traffic situations. Proceedings of the Second World Congress on

Intelligent Transport Systems: 'Steps Forward', III(Yokohama: VERTIS), 1254-1259.

[5] Hoedemaeker, M. (2000). Driving behaviour with ACC and the acceptance by individual drivers. IEEE Intelligent Transportation Systems

Proceedings, 2000 IEEE, 506-509.

[6] Kazi, T.A., Stanton, N.A., & Harrison, D. (2004). The interaction between drivers' conceptual models of automatic-cruise-control and level of

trust in the system. http://www.psychology.nottingham.ac.uk/IAAPdiv13/ICTTP2004papers2/ITS/Kazi.pdf.

[7] Stanton, N.A., & Young. M.S. (2005). Driver behaviour with adaptive cruise control. Ergonomics, 48(10), 1294-1313.

[8] Degani, A., & Heymann, M. (2002). Formal verification of human-automation interaction. Human Factors, 44(1), 28-43.

[9] Heymann, M., & Degani, A. (2007). Formal analysis and automatic generation of user interfaces: Approach, methodology, and an algorithm.

Human Factors, 49, 311-330.

[10] Zheng, P., & McDonald, P. (2005). Manual vs. adaptive cruise control - Can driver's expectation be matched? Transportation Research Part C, 13, 421-431.

[11] Johnson-Laird, P. (1983). Mental models. Cambridge, MA: Harvard University Press.

[12] Rouse, W.B., & Morris, N.M. (1986). On looking into the black box: Prospects and limits in the search for mental models. Psychological Bulletin, 100, 359-363.

[13] Norman, D. A. (1986). Cognitive engineering. In User-Centered System Design (ed. D. A. Norman & S. W. Draper). New Jersey: Lawrence Erlbaum Associates.

[14] Gentner, D., & Stevens, A.L. (1983). Mental models. Hillsdale, NJ: Lawrence Erlbaum Associates, Inc.

[15] Sarter, N.B., Woods, D.D., & Billings, C.E. (1997). Automation surprises. In G. Salvendy (Ed.), Handbook of Human Factors & Ergonomics, Second edition. Wiley.

[16] Moray, N. (1998). Identifying mental models of complex human-machine systems. International Journal of Industrial Ergonomics, 22(4-5), 293-297.

[17] ANSI. (2010). Intelligent transport systems - Adaptive cruise control - Performance requirements and test procedures. ISO/PWI 15622.

[18] Kazi, T.A., Stanton, N.A., Walker, G.H., & Young, M.S. (2007). Designer driving: drivers' conceptual models and level of trust in adaptive cruise control. International Journal of Vehicle Design, 45(3), p. 339-360.

[19] Hoffmann, E.R., & Mortimer, R.G. (1994). Drivers' estimates of time to collision. Accident Analysis and Prevention, 26, 511-520.

[20] Hills, B.L. (1975). Some studies of movement perception, age and accidents. Transport and road research laboratory: Technical report.

[21] Smith, M.R.H., Flach, J.M., Dittman, S.M., & Stanard, T. (2001). Monocular optical constraints on collision control. Journal of experimental psychology applied: Human perception and performance, 27(2), 395-410.

[22] van Winsum, W., & Heino, A. (1996). Choice of time-headway in car-following and the role of time-to-collision information in braking. Ergonomics, 39(4), p. 579-592.

[23] Hirst, S., & Graham, R. (1997). The format and presentation of collision warnings. In Y. Ian Noy (Ed.), Ergonomics and Safety of Intelligent Driver Interfaces. Lawrence Erlbaum Associates: Mahway, New Jersey, 203-219.

[24] Maretzke, J., & Jacob, U. (1992). Distance warning and control as a means of increasing road safety and ease of operation. In FISITA '92: Safety, the Vehicle and the Road, XXIV FISITA Congress (pp. 105-114). London: Institute of Mechanical Engineering.

[25] Brown, T.L. Lee, J.D., & McGehee, D.V. (2001). Human performance models and rear-end collision avoidance algorithms. Human Factors, 42(3), p. 462-482.

[26] Norman, D.A. (1990). The 'problem' with automation: Inappropriate feedback and interaction, not 'over-automation'. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 327(1241), 585-593.

[27] Seppelt, B.D., Lee, J.D. (2007). Making adaptive cruise control (ACC) limits visible. International Journal of Human-Computer Studies, 65(3), 192-205.

[28] Sheridan, T. (1992). Telerobotics, automation, and human supervisory control. MIT Press: Cambridge, MA.

[29] Merat, N., & Jamson, A. H. (2009). How Do Drivers Behave in a Highly Automated Car? In Proceedings of the 5th International Driving Symposium on Human Factors in Driver Assessment, Training and Vehicle Design.

[30] Gold, C., Dambock, D., Lorenz, L., Bengler, K. (2013). "Take Over!" How long does it take to get the driver back into the loop? Proceedings of the Human Factors and Ergonomics Society 57th Annual Meeting, pp. 1938-1942.