subscribe: Posts | Comments | Email

Instrumentation: the driving simulator

Instrumentation: the driving simulator

This is Chapter 3 of the Thesis from Adaptive control to adaptive driver behaviour, by van Winsum, 1996.

Other chapters of this thesis can be found here:


3.1. Background

The preparation of the experiments discussed in this thesis required a substantial amount of software design and implementation for the TRC driving simulator. A full description of the functionality and implementation of the simulator is beyond the scope of this chapter. The reader is referred to other documents for more detail and background (for example Van Wolffelaar & Van Winsum, 1992; Van Wolffelaar & Van Winsum, 1994).

The driving simulator of the Traffic Research Center (TRC) was developed as an instrument for behavioural research of driving. The GIDS project in which the TRC was involved at that time required a simulation testbed for mathematical driving modeling. Because of the dynamic complexities of driver tasks in multi-vehicle traffic, a dynamic traffic simulation was required as a test tool (Van Winsum and Van Wolffelaar, 1993). The objective of GIDS, an acronym for Generic Intelligent Driver Support, was ‘to determine the requirements and design standards for a class of intelligent co-driver (GIDS) systems that are maximally consistent with the performance requirements and performance capabilities of the human driver.’ (Michon and Smiley (1993). A prototype system was developed as a demonstrator for the essential features of the GIDS concept. One of the functions of the GIDS system was to compare required driving behaviour with actual behaviour. Required behaviour was modeled for a number of driving tasks and implemented in a computer system (Van Winsum, 1991; McLoughlin et al., 1993). It was decided at that time to design and implement a dynamic traffic simulation model and connect this with the driving simulator under development. From that moment the driving simulator evolved as a dynamic driving simulator with an interacting traffic world that could be connected to the GIDS system to serve as a test bed. After this, the simulator was also used as a testbed for other external driver support systems.

However, most importantly, the simulator is an instrument for the study of driver behaviour. Because it is used by researchers with very different questions and requirements, flexibility in software design has been considered to be important. This was accomplished by using the object-oriented computer language C++, and a multi-purpose UNIX machine instead of a single purpose dedicated simulation machine. To further increase flexibility for the researchers and to facilitate the design and testing of the experiments reported in this thesis, a fourth generation simulation language, SSL (Scenario Specification Language), was developed for the specification of experiments, together with a specification language (NSL, Network Specification Language) and software tools for roadnetwork creation. Data-sampling and data-processing facilities were added to facilitate experimentation.


3.2 The structure of the simulator

The simulator is composed of a number of software and hardware components that are connected via interfaces. The ‘conventional’ driving simulator consists of a physical car mockup, a car model implemented in software and a graphics system, together with a static road network environment. A dynamic traffic environment is added to this. The structure of these components as well as the object relations are shown in figure 1. In this figure several types of relations can be seen. An “is-a” relation specifies that a certain object type is a subtype of an other more abstract object type. For example, a BMW-car is some kind of car. This means that it inherits the functionality of the more abstract object type car. A “has-a” relation specifies that a certain objects has another object as a member. For example, a car has a traverser. The heavy printed arrows specify the direction of the flow of information. For example, there is an information flow from the object roadnet to the object sensor. This means that a sensor requests information from a certain instantiation of the object roadnet.



Figure 1. Logical structure of components of the driving simulator and relations between objects.


In addition to this, a number of facilities related to data-sampling and processing and experimental control are added to the simulator.


Car cabin. The steering wheel, clutch, gear, accelera­tor, brake and indica­tors of the car (a BMW 518) are connected to a Silicon Graphics Skywriter 340VGXT compu­ter (IRIS). Electromotors and other electronic appliances are built in the car to excert forces on the pedals and steering wheel and to send data on the steering wheel, accelerator pedal, brake pedal and indicators to the IRIS computer for further processing by the car model.


Car model. The IRIS computer processes these signals in a separate process referred to as the car model. The car model is described in more detail in Spaargaren (1994). It computes the longitudinal and lateral speed and acceleration that are the result of physical characteristics of the car and the input from the car cabin. From this the new coordinate position in the artificial world the car is driving in is computed. The output of the car model is used by the car traverser and by the graphics system. The traverser constitutes the link with the dynamic traffic process while the graphics system presents the output of the full system in a real-time visual format to the driver.


Graphics system. On a projection screen, placed in front, to the left and to the right of the driver, an image of the outside world from the perspective of the driver with a horizontal angle of 150 degrees is projected by three graphi­cal videopro­jectors that are controlled by the graphics software. Images are presented with a rate of 15 to 20 frames per second, resulting in a sug­gestion of smooth move­ment. The visual objects are buil­dings, roads, traffic signs, traffic lights and other vehicles.

In addition to this, the sound of the engine, wind and tires is presented by means of a digital soundsampler recei­ving input from the simulator computer.


Logical network (Roadnet). The logical network is the static environment in which the simulator car and traffic operate. The static environment consists of a database with a network of roads, traffic signs, traffic lights and buildings. This database is used for the visualization of the environment by the graphics system and by the artificially intelligent traffic to evaluate the present situation. The database can be generated in two ways:

– by NSL (Network Specification Language). This is a user specification language, created for the TRC simulator (Van Winsum, SSL/NSL specification release 1.2, 1994), by which a network of roads can be specified as an ASCII text. This text is processed by an NSL interpreter program that generates a road network database that is used by the simulator (Van Winsum, 1994, NSL scanner/ parser/interpreter computer program).

– by means of an interactive graphical program written in C++/OSF Motif (Van Winsum, 1993, program WORLDED). The user can specify a network of roads by means of click and point operations. The output of the NSL interpreter can also be used as input for this program to visualize and change the network.


The network consists of a structure of three base tables: a table with intersections, a table with paths and a table with segments. An intersection is a point in the network coordinate system with 1..n, {n >= 1}, outgoing paths. Coordinates are in meters. The following relations hold:


– n = 1: the intersection forms a terminal point in the network. If cars approach this intersection they cannot proceed beyond the intersection and provisions are made to ensure that the car turns around in the opposite direction as soon as the intersection is reached. The intersection has no physical layout and has the appearance of an ending road. The implication is that it is not possible for cars to move off the logical world.

– n = 2: the intersection is a virtual intersection in the sense that it has no specific layout and is not treated as an intersection by the traffic. The only purpose of creating such an intersection is for the convenience of the network constructor.

– n > 2: the intersection has more than two branches.

An intersection is of a certain type (f.i. a roundabout), it can be controlled by traffic lights with a certain control strategy, and it contains a list of references to outgoing paths. This list is ordered such that the path connections to the intersections are counterclockwise. In addition to this the intersection contains information about the layout, which is used by the graphics system and by the traffic.


A path is a logical connection between two intersections and always has one direction. It must start at one intersection A and end at one intersection B, where A may be equal to B. If A=B then the path is logically a circular path. All paths  have precisely one path in the opposite direction, referred to as a counterpath. It has a list of references to segments with 1..n elements, {n>=1} . This list is ordered such that the segments are in successive order. A path also contains information on right-of-way at the intersection at the end of the path, whether entry into this path is allowed, a reference to a traffic light at the end of the path if there is one, and information on buildings on the right side of the segments on the path.

To every path an ordered list with references to cars is attached. This list is ordered such that it reflects the order of the cars on the path and it may be empty. Cars can be added or removed at any time during the simulation process. In this way the simulator car and the computer controlled cars are connected to the static environment. Because every car is an object in the software-engineering conception that it has its own functions and data-structures, every car performs its own administration of detailed position (coordinates, distances from the last intersection and from the edge of the road etc.) in relation to the logical network.

The concept of path corresponds to the terminology of graph theory. Using that terminology, intersections are nodes.


The combination of nodes and paths may be described as a directed graph with the following properties:

– Suppose the network is represented as the graph G=(P,Z), with P being the set of intersections or nodes and Z being the set of ordered relations between the intersections, then P = {0..n} with n > 0. The fact that all intersections are member of a set ensures that all members occur once. The number of the intersections are in successive order. A set of intersections is, for example, {0,1,2,3}, meaning that there are 4 intersections. The set {0,1,3,4} is incorrect because the number 2 is missing.

– Z contains the relations between two nodes A and B, for example {{1,2}, {1,3}, {1,1}}. If {A, B} is a member of Z then {B,A} is also a member. This shows that all paths have a counterpath. A road can be traveled in two directions and this is the reason that every path has a counterpath. If only one-way traffic is allowed there are still two paths because physically it is possible to enter a one-way street into the wrong direction although legally it is not allowed.

– The fact that Z is described as a set suggests that the member {A,B} may occur only once. This restriction has been abandoned for practical purposes. There may be more than one instantiation of the relation {A,B}. In that sense Z is not a set but a collection. This restriction was loosened because sometimes there is more than one road between two intersections.

– A further restriction to the graph specification is that all nodes must occur in at least one relation, that is, a node that is fully unconnected is not allowed.


A segment is represented as a line through the middle of a roadpiece. It can be either straight or curved and is undirected. Segments are members of ordered lists connected to a path and the ordered list must contain al least one segment. A segment must be a member of one and only one ordered list. Segments represent the physical layout of the road, while a path represents the logical presence of a road. The direction depends on the path the segment is in. If the segment is straight the two end points are given in coordinates. If it is curved the segment contains the necessary information on the curvature, such as the radius, the centerpoint of the arc etc. A segment has a certain lane-width. At present only two-lane segments are allowed.

Traffic signs, buildings and traffic lights are connected to the network and have a certain position, angle, and type. Within the simulator program this roadnet representation is implemented as the separate object class in the roadnet module (Van Winsum, 1992, computer program class c_roadnet, roadnet.c). This object performs its own administration and can be queried from outside via an interface.

The following is an example of a definition of a simple network with NSL.

Define Inter[0] {

X := 100; Y:= 100;


Define Segment[0] {

Type   := Straight;

StartX := Inter[0].X;

StartY := Inter[0].Y

Length := 100;

Angle  := 90;



In this definition a straight road of 100 meters with an absolute angle of 90 degrees is created, starting at coordinate position [100, 100]. Paths are added automatically by the system. Since this definition of a network would result in a path without an end node, the system creates an end node (intersection number 1). Since the lane-width is not specified, the segment is assigned the default lane-width of 3 meters by the NSL system. In this case the NSL interpreter creates 2 intersections, 2 paths and 1 segment, no traffic signs, traffic lights or buildings. NSL contains a number of geometric transformation methods and rules to assists the user and to make it easier to build the network.


Traffic. Traffic consists of a list of cars that may be controlled by a human driver (the simulator car) or by an artificially intelligent ‘driver’. Every car has a number of properties, such as a length, a width, a wheel-base and so on and a number of objects that are needed for driving in the logical world. These objects are a traverser, a sensor and a decision (control) mechanism. In the case of a human-controlled car the decision mechanisms is of course the human driver who, together with the car model, determines the movement of the car. In the case of a computer controlled car the decision mechanism is composed of a set of decision rules. Traffic is implemented in the simulator program as a separate object container class (Van Winsum, 1992, computer program class c_traffic, traffic.c). It contains all kinds of methods for adding or removing cars from a traffic list. The class traffic contains references to cars which may be very different in type. The mechanisms of late binding and virtual classes and inheritance, which are properties of the object-oriented methodology used, ensure that in the future all kinds of other moving objects such as pedestrians and bicyclists may be added to traffic. Every car has its own instantiation of a traverser, sensor and control object. These objects also may be of different types. For example, a human controlled car (the simulator car) would need a somewhat different traverser than a computer controlled car or maybe a pedestrian.

In the case of a human driver, the output of the car model, i.e. the speed and the angle of lateral displacement, are input for the traverser. For computer-controlled cars, the output of the artificially intelligent decision mechanism is the input for the traverser. The traverser calculates the lateral position (with respect to the right side of the road), the longitudinal displacement with respect to the road, it connects the car to the network of roads, checks which path is selected if the car is on an intersection and performs a number of other checks to maintain the position of the car accurate with respect to other traffic. It uses deadreckoning techniques in this process. The traverser is the interface between traffic and the road network and it also connects the simulator car with the interactive traffic world. The traverser is implemented as a separate object class in the simulator program, such that every car has a reference to its own instantiation of a traverser object (Van Winsum, 1992, computer program class, c_traverser, travers.c)

The sensor is an object that really consists of a set of sensors. Both the human controlled car and the computer controlled cars have a sensor object but they use it differently. In general, the sensor is used to ‘look’ into the network. In this way every car, including the simulator car, can evaluate the present type of road and curvature, evaluate the distance and speed of traffic in front etc. This means that the sensor is an interface between the network and the car in that it requests information from the network. The human controlled car uses this information for data storage purposes and to give input to driver support systems. The computer controlled cars use this information for the decisions they are required to make concerning their speed and course. Sensor is implemented as a separate object class in the simulator program (Van Winsum, 1992, computer program class c_sensor, sensor.c). Every car has a reference to its own instantiation of a sensor object.

The control mechanism for the human driver is the human information processing system that uses visual information received via the graphics system, to exert the controls in the car cabin. These car control signals are processed by the car model. The output of the car model is used to update the graphics and as input for the traverser that connects the simulator car to the network. The control mechanism of the computer controlled cars consists of a set of decision rules. Every computer controlled car has rules for different driver tasks on the tactical level. These tasks are related to curve negotiation, car-following, overtaking, negotiating intersections, speed choice on straight roads and processing road sign information. The car evaluates which tasks are presently performed and computes a required speed and lateral position. Since multiple tasks can be performed at the same time, a decision mechanism selects the appropriate speed and lateral position together with the required acceleration and wheel-angle to reach this state, after all tasks have been evaluated for the present car . This results in a natural and human-like behaviour that contributes in an important way to the fidelity of the simulator. For computer controlled robot cars the artificial intelligence is implemented in a separate object class in the simulator program (Van Winsum, 1992, computer program class c_control, control.c).

3.3 Data collection and processing

A large quantity of performance data can be collected with any sampling frequency. Examples are time-to-collision, time-to-intersection, time-to-line crossing, lateral position, speed, acceleration, brake force and so on. The user creates an ASCII text with keywords that specify the sample frequency and the data to sample with that frequency. The data are then sampled during a simulation run and stored into a binary file. The real-time handling of data-storage during a simulator run is controlled by a separate object class c_data that is implemented in the simulator program (Van Winsum, 1992, computer program class c_data, newdata.c).

After a simulator run the data can be visualized and preprocessed with a graphical program written in C++ and X-windows/OSF motif (Van Winsum, 1994, program DATAPROC).


For the experiments described in this thesis the real-time sampling of time-based information was required. The variables used are TTC (time-to-collision), TLC (time-to-line crossing) and THW (time-headway during car-following). These measures are defined and implemented as follows:


– THW is defined as D/u

for u > 0, else THW = infinite (undefined)

with D = bumper to bumper distance in meters along the path between

the simulator car and the lead vehicle, and

u = speed of simulator car in m/s


– TTC is defined as D/(u – ulead)

for (u – ulead) > 0, else TTC = infinite (undefined)

with ulead = speed of lead vehicle in m/s


– TLC is calculated differently depending on whether the car is on a straight road or in a curve.

In general, TLC = DLC/u,

for u > 0, else TLC = infinite (undefined)

with DLC = distance to line crossing along the vehicle path and

u = speed of simulator car in m/s.


DLC is solved goniometrically using the cosine rule. Normally, the car is not driving in a straight line but it alternates between curves to left and to right. The radius of the vehicle path is calculated using the coordinates of the centerpoint of the curve the car is driving. This centerpoint is calculated as the point where the perpendicular lines through the frontwheel and the rearwheel, with slipangles added to the wheelangles, intersect. Rv, the vehicle radius, is then computed as the distance between the center of gravity of the car and the centerpoint of the vehicle curve. Rv1 then is the distance between the front (left or right) wheel and the centerpoint of the vehicle curve. TLC then measures the time until either the left or right front wheel crosses the left or right lane boundary, given the current vehicle path and speed.

First the case for straight roads is described if the vehicle makes a left turning curve, see figure 2. DLC is computed as a*Rv1. Since Rv1 is known, only a has to be computed, using the cosine rule.



Figure 2. Determination of the length of the arc DLC for driving on straight roadsections.

– The length of the linepiece A is computed as Rv1-(dleft/cos(ha)), with dleft being the distance between the left frontwheel and the lane boundary (in a line perpendicular on the road) and ha the angle between the line perpendicular on the road and the line from the front wheel to the centerpoint of the vehicle curve.

– The length of the linepiece C is computed as  (2*A*cos(ß)+Ö((2*A*cos(ß))2-4*(A2-Rv12)))/2

Then a = arccos((A2 + Rv12 – C2)/(2*A*C))

and DLC = a*Rv1


Figure 3. Determination of the length of the arc DLC for driving on curved roadsections.


Figure 3 shows the situation for determining the TLC while the car is negotiating a road curve. Again, DLC is determined as a*Rv1. In this case a is computed differently.

– The length of linepiece A represents the distance between the centerpoint of the roadcurve and the centerpoint of the vehicle curve.

– Angle ß is computed as the angle difference between the line from the centerpoint of the vehicle curve to the centerpoint of the roadcurve and the line from the centerpoint of the vehicle curve to the left front wheel (if the vehicle turns towards the inner lane boundary).

– Angle a1 is computed as arccos((A2 + Rv12 – Rr2)/(2*A*Rv1))

– a= ß – a1 and DLC = a*Rv1


In addition to this, vehicle control information was required for the experiments. The position of the accelerator pedal, expressed as a percentage pressed, the position of the brake pedal and the force excerted by the foot on the braking pedal were used in the studies on car-following, while steering wheel angle was used in the study on steering performance and curve negotiation. The results of these control actions, such as speed, acceleration, heading angle and lateral position, were sampled and processed as well.

For every experiment automatic data processing programs were written to extract and process the required data. These data were then transformed into a format suitable for processing by SPSS.


3.4 Scenario Specification Language (SSL)

SSL is a user specification language that was defined and implemented as a tool for specification and design of experiments. It contains most of the functionality of the simulator. A description of SSL then essentially gives a description of the functionality of the TRC simulator. For a full specification of the language the reader is referred to the SSL/NSL specification document (Van Winsum, 1994).

An ASCII file with SSL commands is analyzed by a scanner and parser module during initialization of the simulator program and syntactical errors are reported to the user. (Van Winsum, 1994, SSL scanner/parser/interpreter modules). If no syntactical errors are found, the SSL commands are converted to an internal data-structure that is interpreted in real-time by the SSL-interpreter during execution of the simulation process. Since the simulation process is a dynamic process in which the state is determined by SSL specifications, the human driver, the behaviour of traffic and by the process operator who interacts with the computer via the user interface, the course of events is not deterministic. However, SSL commands can be used to force identical situations for all subjects in an experiment. Since SSL commands are often conditional, the state of the traffic world can be queried and events can be triggered if some condition is true or false.

Scenarios are defined in a SSL text file. A scenario is a predefined list of situations with a start and an end condition: the scenario starts when the start condition is fulfilled and terminates when the end condition is fulfilled. A scenario may involve 0..n cars, referred to as participants, in addition to the simulator car. A participant is a car that performs conditional actions. A scenario may be used for controlling traffic and traffic lights, for indicating when data must be stored, for communication with the driver with spoken or written messages, for sending messages to other devices, and so on. SSL is not exclusively a language for specification of traffic situations during an experiment. It also may be used for rapid prototyping of driver support systems, for creating test situations and for debugging. It is important to note that SSL is often used to override default settings and default behaviour. For example, if a computer-controlled car is created with SSL, the car follows its own rules unless specified differently with SSL.

The following is an example of an SSL description.


Define Scen[1] {

Var { time; }

Start {

When ( Part[MainTarget].LeadCar != Absent and

Part[MainTarget].DisToLeadCar < 50 );

Scen[].NrTimes := 1;   time := runtime();


End {

When ( runtime() – time > 20 );


Define Part[1] {

Start {

Part[].CarNr  := Part[MainTarget].LeadCar;

Part[].MaxVelocity := 50/3.6;


End {

Part[].MaxVelocity := 100/3.6;





This scenario specifies that if there is a lead vehicle and the distance to it is less than 50 meters then this lead vehicle starts to drive with a maximum speed of 50 km/h during 20 seconds. After 20 seconds (at the end of the scenario) this vehicle pulls up to a speed of 100 km/h.

SSL files contain the full script for an experiment and are thus complete specifications of an experiment. This ensures repeatability and detailed documentation of experiments. Since researchers are able to make their own SSL script files they can design and test experiments with a minimal dependency on technical staff and computer programmers.


3.5 The use of the simulator in the experiments

The driving simulator offers a number of advantages compared to studying driver behaviour on the road.

1) The sensors of the simulator car and of other cars used in the car-following experiments contain important information that is much harder to obtain with current technology in a real world test situation. This information is vital as input for the control of experiments and data-sampling. For the experiment on curve driving the measurement of TLC information during curve negotiation would be very hard to obtain in real world driving. A simulator is the only practical way to obtain complex measures such as the TLC in curves. Although time-to-collision information may be obtained during on-road experiments it is measured more practically and efficiently in the simulator.

2) All kinds of situations can be generated and tested that would be very hard or impossible to generate in the real world. In the curve negotiation experiment the drivers are required to negotiate a number of different road curves with a specific lane width and radii. Roads with the precise characteristics required by this experiment are very hard to find in the real world. During the car-following experiments the lead vehicle was sometimes required to drive with a certain fixed time-headway in front of the simulator car. This would be difficult to establish in on-road experiments.

3) The responses of drivers to maneuvers too dangerous to be tested in real world driving can easily be tested in the simulator. This is especially important in the car-following and braking experiments discussed in the chapters 5 to 9.

4) Situations can be brought under experimental control. This is important for the comparability of the results since all subjects have encountered precisely the same situations. In on-road experiments traffic density and weather conditions are hard to control. In this respect a simulator has important advantages compared to real world experiments.


In the experiments performed in the context of this thesis, the time-based safety margins TLC and TTC play an important role. The perception of TTC has been studied in a large number of experiments (see chapter 6). These studies have given strong support for the idea that TTC information is extracted from the optic flow field. The expansion of the image on the retina gives sufficient information for the extraction of TTC information without requiring the driver to assess speed or distance information. Since the graphical properties of optical perspective, visual angle and optical expansion rate are the same in the TRC simulator as in real world driving, there is reason to assume that the driving simulator is suitable for the type of research discussed in the chapters 5 to 9. An important prerequisite for a smooth optic expansion is a high graphical frame rate. In order to obtain a high frame rate, the visual scenes in all experiments are limited to the essential components while substantial effort has been invested in the design of fast algorithms for traffic handling and experimental control.


Chapter 10: thesis traffic psychology


This is chapter 10 of the thesis from 1996 by van Winsum. It concerns a number of behavioural studies into driver adaptation that have been performed in a research driving simulator.

Other chapters of this thesis can be found here:


General discussion and conclusions


10.1 Testing the adaptation model for the case of individual differences: discussion of results from the experiments


The adaptation model predicts that factors  that affect operational performance will normally result in an adaptation of behaviour on the tactical level, such that constant safety margins are maintained. Individual differences in operational performance are then predicted to be reflected in individual differences in behaviour on the tactical level. The results of the experiments support the adaptation model applied to the relation between individual differences in behaviour on the tactical level and the operational level for both the car-following task and the curve negotiation task. The general results from these experiments and the relevance for the adaptation model are discussed in this chapter.

If individual differences in skills and operational performance result in adaptation of behaviour on the tactical level then this behaviour must be consistent and characterized by individual differences. This implies that, in addition to the transient effects on tactical behaviour discussed in previous paragraphs, some level of consistency and constancy must exist in, for example, speed choice and choice of headway during car-following. If the adaptation model also applies to individual differences then at least some part of the between-subjects variance in behaviour on the tactical level must be explained in terms of the between-subjects variance in operational performance. The tasks of curve negotiation and car-following were selected for closer examination. Speed choice during curve negotiation is considered as an example of the effect of lateral control performance on behaviour on the tactical level. Choice of time-headway in car-following is described as an example of the effect of longitudinal control performance on tactical behaviour.

Experiment 1 deviates from the other five in that it not only examines the effect of individual differences in operational performance but also the effect of a situational factor, i.e. curve radius. The results have been published in Human Factors. The experiment was performed in a driving simulator that was programmed by the people from Carnetsoft. Explaining speed choice as a function of curve radius has been a long lasting problem that has been investigated in a large number of studies. The problem is usually described in terms of a relation between lateral acceleration and choice of speed. The underlying process has never become clear. However, the results of experiment 1 clearly suggest that the inverse relation between lateral acceleration and speed, often referred to in the literature, is the result of a process of adaptation of speed choice together with a strategy of maintaining constant safety margins. Speed choice in curves proves to be a consistent measure of tactical behaviour. Also measures of operational performance prove to be stable and consistent within the driver. This indicates that the important prerequisite for the validity of the adaptation model that both operational performance and behaviour on the tactical level are consistent and characterized by clear individual differences is fulfilled for the case of lateral control performance and speed choice in curves. Steering is discussed as the factor that affects choice of speed in curves. A model of steering is presented that suggests that steering errors are affected by individual differences in steering competence and by required steering-wheel angle. A larger required steering-wheel angle then results in larger steering errors. The situational factor road radius, together with speed, affects required steering-wheel angle. A smaller radius increases required steering-wheel angle and thus steering error, which is compensated or adapted for by choosing a lower speed. The same reasoning applies to individual differences in steering performance. This is measured independently during straight road driving. Drivers with poorer steering competence are characterized by larger steering errors which is compensated for by choosing a lower speed in curves according to the adaptation model. Summarizing, in experiment 1 the adaptation model is tested in two different manners for the case of speed choice in curves:

– curve radius affects operational performance which affects speed choice, and

– steering competence affects operational performance which affects speed choice.

These hypotheses are supported by the results of experiment 1. The results indicate that a smaller curve radius and poorer steering competence increase steering errors and result in such speed reductions that TLC is kept on a constant minimum value. These results then strongly support the adaptation model discussed in paragraph 2.5 and the value of the concept of a time-based safety margin that is controlled during driving.


Experiments 2 to 6 consider the task of car-following. During car-following the driver never knows whether the lead vehicle will brake, and if it does, how hard it will brake and for how long. It is then assumed that the driver has learned the quality of his or her braking performance from previous experiences and that this results in the choice of a preferred  time-headway (THW). THW is the time available to the driver to reach the same level of deceleration as the lead vehicle in case it brakes, without becoming involved in a collision. Braking performance is assumed to affect the time required to reach the same level of deceleration as the lead vehicle. Adaptation of THW may then be regarded as a compensation strategy for drivers with poorer braking performance.

The detailed examination of the car-following task introduces some specific problems. First of all, it is not immediately clear which aspects of operational performance play a role in choice of time-headway. This is examined in the experiments 2 to 6.

Secondly, the literature on choice of THW during car-following is not very extensive. The literature on braking is limited as well and mainly restricted to emergency braking (braking as fast as possible), see chapter 5. This implies that the theoretical perspective on braking and car-following had to be developed during the course of experimentation and that the number of experiments required to test the theoretical model is larger for the case of car-following than for speed choice in curves.

Thirdly, an important limitation in the study of car-following is that the details of operational braking performance can only be compared between different drivers if they start braking at the same distance- or time-headway. This means that, in studying braking performance, drivers will have to be forced into time-headway conditions they would not choose themselves, which may result in differential effort allocations as a function of the discrepancy between preferred THW and actual THW. This was illustrated by the results of an experiment by Heino et al. (1992). They found that particularly drivers who normally follow at a larger THW increase their mental effort, as measured by heart rate variability, when they are forced to follow at a smaller THW. This means that the methodological prerequisite of measuring braking performance in forced-paced situations may, to some degree, obscure individual differences in braking performance because of between-subjects differences in effort allocation. Nevertheless in the present studies, this method is preferred to the alternative where braking performance is measured while drivers follow at their preferred THW. Drivers who follow at a smaller THW would be forced to brake faster compared to drivers who follow at a larger THW, and this would damage the comparability of braking performance between drivers.

The results of experiments 3 and 4 demonstrate that choice of THW is consistent and constant over different speeds. In experiment 3 preferred THW is measured at speeds of 40, 50, 60 and 70 km/h. Speed has no significant effect on preferred THW and the within-subjects reliability of the THW’s is high. This is confirmed by the results of experiment 4. The high consistency in choice of THW has been confirmed in an on-road experiment by Heino et al. (1992). They reported a correlation of 0.85 between time-headways measured on two different stretches of road. Other studies on the consistency of THW are discussed in chapter 6. The results indicate that choice of THW is independent of speed and consistent within the individual driver and that clear and reliable individual differences exist in choice of THW. This is an important prerequisite for the application of the adaptation model to individual differences.


Experiment 2 examines the relation between preferred THW and the ability to brake as fast as possible,  the speed of stimulus encoding and response preparation. The additive factor logic (see Sternberg, 1969) is applied to examine the locus of effect of operational performance differences. The search for differences in the ability to brake as fast as possible stems from the tradition in the literature on braking where the quality of braking is generally examined in terms of the ability to brake as fast as possible. Experiment 2 may therefore be regarded as a search for individual differences in the limits of performance. The braking parameter that is studied in detail is reaction time (RT), defined as the interval between the moment the lead vehicle starts to brake and the moment the subject starts to release the foot from the accelerator. Again, this approach stems from the dominant view in the literature on braking, i.e. that differences in braking performance originate from perceptual factors measured by RT. Differences in the speed of stimulus encoding regarding the braking action of the lead vehicle would suggest that drivers with a smaller preferred THW (short followers) perceive the braking of the lead vehicle earlier. Differences in response preparation would suggest that the state of motor readiness is reached earlier by short followers compared to long followers. The results indicate that choice of THW is not related to individual differences in RT for a decelerating lead vehicle, to differences in stimulus encoding or to differences in response preparation. From this it is concluded that differences between short and long followers cannot be explained in terms of “limits of perceptual and motor skills”. However, differences in preferred THW appear to be related to braking performance in quite another way. Differences in response execution speed as a function of preferred THW are restricted to braking situations characterized by uncertainties concerning the braking of the lead vehicle, the required deceleration and the duration of braking, as is always the case in real world car-following situations. The results suggest that individual differences in the transformation of visual feedback to the motor response may be related to choice of THW. The results have been published in Ergonomics.


Experiment 3 considers these aspects in more detail and examines the use of time-to-collision (TTC) during braking and the way the braking response is executed. The process of braking is connected explicitly to the literature on time-to-collision (TTC). TTC is defined as the time required for two vehicles to collide if they continue at their present speed and on the same path (see for example Van der Horst, 1990). In the literature it is often suggested that the perception of TTC from the optic flow field is an important skill related to the initiation of braking. But curiously, only a few experimental studies have connected the concept of TTC to the braking response. The general conclusion from the literature is that TTC is underestimated and that there are large individual differences in the ability to accurately estimate TTC. In experiment 3 the hypothesis is tested that preferred THW is related to the sensitivity to TTC information. According to this reasoning, drivers who are more sensitive to TTC information are better able to judge the moment to start braking while drivers who are less sensitive to TTC information run the risk of starting to brake too late. This may result in a compensatory larger preferred time-headway for these drivers. The results indicate that both the initiation and the control of braking are strongly determined by TTC on the moment the lead vehicle starts to brake. Short followers differ from long followers in the control of braking: short followers brake harder and more efficiently, and, most importantly, the intensity of braking is more sensitive to TTC differences, compared to long followers. Yet, a confounding factor may have affected the results. Because the absolute levels of TTC differ between short and long followers in this experiment, short followers may have been forced to brake more efficiently.


Experiment 4 explicitly controls this confounding factor. Braking performance is measured with identical initial time-headway for all subjects. The subjects are unaware of the fact that the lead vehicle will brake and of the required deceleration and the duration of braking. A model of braking is discussed in which the process of braking is divided into three separate phases: the RT phase, the open-loop ballistic phase and the closed-loop phase. The RT phase is defined as the interval between the moment the lead vehicle starts to brake and the moment the foot starts to be retracted from the accelerator pedal. The open-loop phase is operationally defined as the period that starts when the subject retracts the foot from the accelerator after the lead vehicle has started to decelerate and ends when the brake pedal is touched. During the closed-loop phase visual feedback is used to control the process of braking. It is defined operationally as the period between the moment the brake pedal is touched by the foot and the moment the maximum brake position is reached. It is hypothesized that the speed of the open-loop ballistic response is determined by TTC on the moment the driver detects the deceleration of the lead vehicle, while the duration of the closed-loop phase is determined by the number of decelerations in the brake pedal signal (movement corrections). The results show that reaction time is not related to preferred time-headway. This confirms the results of the experiment 2. The open-loop phase of the motor response appears to be very sensitive to TTC, and especially to TTC on the moment the foot is retracted from the accelerator pedal. This supports the hypothesis. Also, the results indicate that short followers are characterized by a faster open-loop response that is not caused by a smaller TTC. This suggests that short followers program their movement speed to a higher level compared to long followers. The duration of the closed-loop phase of the motor response is, in accordance with the hypothesis, strongly related to the number of movement corrections. Short followers exhibit a faster closed-loop response with fewer movement corrections. The results also indicate a strong effect of total movement time on preferred THW, strengthening the conclusion that short and long followers differ in both the open- and closed-loop phases of movement. This suggests that short follo­wers are more sensitive to the task requirements in braking situations, confirming the results of experiment 3.


Experiments 5 and 6 test the hypothesis that short followers differ from long followers in the sensitivity of the braking response execution to TTC information. Both experiments apply the model of braking discussed in chapter 7. In experiment 5, the RT phase, the open-loop and the closed-loop phases of the braking process are manipulated independently. If short followers differ from long followers in either of these phases then the factor “preferred THW” should interact with any factor that manipulates these phases. The RT phase is manipulated with the factor initial THW on the moment the lead vehicle starts to brake. The duration of the open-loop phase is manipulated by the factor initial deceleration. The level of deceleration (3 vs. 6 m/s²) of the lead vehicle is expected to affect the TTC on the moment the subject detects the braking of the lead vehicle and thereby the duration of the open-loop phase. The closed-loop phase is manipulated by the factor secondary deceleration: as soon as the foot touches the brake pedal (this is the moment the closed-loop phase starts) the deceleration of the lead vehicle changes. This requires the use of visual feedback in order to change the programmed motor response. Although the results show that the respective phases of the braking response are affected by the manipulations, the predicted interactions of preferred THW with the factors that manipulate the open- and the closed-loop phases are not statistically significant. The pattern of results suggests that task-specific factors resulted in undesirable startle reactions and vigilance effects.

Because of this the final experiment 6 applies multiple measurements per manipulated factor, a higher frame-rate and shorter task duration, in order to prevent startle reactions and vigilance effects. The main hypothesis is that short followers differ from long followers in the sensitivity of the motor response to TTC. TTC is manipulated with two levels of initial deceleration of the lead vehicle (3 vs. 6 m/s²), in random order. The results indicate, in support of the main hypothesis, that the open-loop response of short followers is more sensitive to differences in TTC compared to long followers. The assumed causal chain is that individual differences in some basic perceptual-motor skill affect the quality of the braking response. The driver is assumed to adapt the choice of THW during car-following accordingly. In this way drivers protect themselves against poorer operational performance. However, it may be argued that short followers have had more practice in braking resulting in improved operational performance because of learning effects. To rule out this explanation it is examined whether short followers differ from long followers in perceptual-motor performance in tasks unrelated to braking. In order to test whether choice of THW is related to more general perceptual-motor skills that require the transformation of visual information to a motor response, performance on a lateral tracking task and a longitudinal tracking task is tested. The results clearly indicate that short followers perform better on both the lateral tracking tasks and the longitudinal tracking task. In addition to this, performance on both types of tracking tasks is significantly correlated. This strongly suggests that:

1) Short followers differ from long followers in perceptual-motor skills related to the transformation of visual information to a motor response,

2) these differences in skill are not acquired as a function of differences in following behaviour,

3) these differences in skill affect the quality of braking performance in the sense that short followers tune the braking response better to the requirements of the situation, giving them a higher sense of control,

4) resulting in the choice of a larger time-headway for drivers with poorer operational performance and a smaller time-headway for drivers with better operational performance.


10.2 General conclusions and next steps


The impact of vehicle factors and situational factors related to road, weather and temporary state on driver behaviour and the underlying mechanisms of behavioural effects have been addressed in this study. Mechanisms related to individual differences in driver behaviour have been tested from the perspective of the adaptation model. It is clear that the system components vehicle and environment have an important effect on driver behaviour, mediating the effects on accident involvement and traffic safety in general. Adaptation mechanisms are best studied by measuring driver behaviour as a function of vehicle factors, individual differences in skills, situational factors and temporary states instead of accidents, because these factors affect behaviour directly while they affect accident involvement indirectly. One of the reasons for the lack of progress in driver modeling, referred to in chapter 1, is the abundance of determinants and factors that operate simultaneously. This has resulted in several theories that apply only to a limited problem domain. The adaptation model integrates the operational and the tactical level of driver behaviour into one framework. As discussed in chapter 2, driver models and studies in traffic psychology usually examine only one of these levels. It is suggested that these levels should always be studied in their mutual relationship. For example, if the effect of a roadmeasure on speed is examined it should also examine the effects on operational performance at the same time. Of course practical problems may prevent this and this is one of the reasons why simulators may be useful. However, the results suggest that measurement of behaviour on one level may be meaningless when behaviour on another level is excluded from examination. Several other driving tasks such as speed choice on straight roads, gap acceptance at intersections, stopping for traffic lights, overtaking and so on need to be examined within this framework.

According to the adaptation model, drivers with poorer operational performance protect themselves by adapting behaviour on the tactical level, resulting in a lower speed or larger time-headway. The other side of this reasoning is that drivers with better perceptual-motor skills and good operational performance drive at higher speeds or follow at smaller time-headways. However, it is not by any means intended to suggest that drivers with higher speeds are not dangerous because they have a highly developed skill level. Undoubtedly, some drivers who follow other vehicles at a close distance or who drive faster than average are not characterized by better operational performance. The suggested relation is a probabilistic one, and not mechanistic. However, the line of reasoning makes clear that the concept of risk becomes more meaningful if skills and level of performance are added to the equation. This is to say that a certain speed may not be as risky for one person as for the other if they differ in certain required perceptual-motor skills, from the same perspective as the fact that flying an F16 fighter plane is considerably more risky for the author of this thesis than for an experienced pilot.

Also, it is often assumed that higher speeds and shorter following distances are associated with a high accident risk although a number of studies do not confirm this simple relation. The effect of variabi­lity within the traffic system on accidents is one of the reasons why Summala (1985) promoted the introduction of speed limits. This has greatly reduced the accident risk in a number of countries. Speed limits reduce the variability of speed in the system and this reduces accident risk. Brehmer (1990) predicted that accident probability is lowest for cars driving with the average speed, but increa­ses for drivers who deviate more from the average speed, either to lower or higher speeds. He referred to a study of Solomon (1964) on the relation between speed and accident rate on US highways, that supported this hypothesis. Munden (1967), referred to in Rooyers et al. (1992), repor­ted a U-shaped relation between speed and accident rate as well. Brehmer also predicted that accident rates are higher in environments where the variance of the speed distribu­tion is highest. A study of Greenberg (1964) was referred to which demon­strated a positive correlation between accident rate and speed distribution for a sample of US roads. Numerous authors have mentioned that it is an establis­hed fact that accident risk is related to driver speed, and that speeding therefore can be regarded as a form of driver error, related to poor speed perception skills or poor hazard perception. However, whether a higher speed is riskier compared to a lower speed with identical speed distributions is an unresol­ved matter. A similar point can be raised with regards to headways during car-following. Shorter time-headways are usually associated with a higher risk of rear-end collisions. In a large-scale study on the relation between time-headway and accident risk in several countries a relation was found between rear-end accident rates per 100 million vehicle kilome­ters and time-headway (Benjamin, 1980). This relation was however strongly affected by the flow of traffic or traffic density. Traffic volume affected both time-headway and the number of rear-end collisions so that a causal relationship between close following behaviour and rear-end acci­dents could not be established. It was already demonstrated in the fifties by the studies of Herman (referred to in Forbes, 1972) that, even in car-following situations with conservative headways, normal speeds and short response times of drivers, flow distur­bances by the platoon leader (the first car in the chain) may cause a chain reaction that makes it impossible for drivers downstream to avoid a collision. This indicates that the relation between speed choice, choice of time-headway and accident risk is not as straightforward as often suggested.

The general principle of behavioural adaptation demonstrates the inherent flexibility of human behaviour. This flexibility resembles the issue of ‘human behaviour feedback’, discussed in chapter 1, which has puzzled many traffic safety researchers and triggered fierce discussions about the effects of safety measures. The adaptation model may offer the concepts and methodology to clarify the issue of this ‘human behaviour feedback’ in more coherent terms. Driver adaptation of tactical behaviour to effects of safety measures on operational performance may be an important determinant for the success or failure of intended safety changes in the road-vehicle-driver system.

Although the process of adaptation appears to be ‘normal behaviour’, it also seems clear that certain factors prevent adaptation resulting in increased accident involvement. Examples of this are the consumption of alcohol and the case of the young male driver. Citing from paragraph 2.3.4: “The interaction of BAC level and age on accident involvement suggests that both factors share a common locus of effect, in the sense that the factor that causes the higher accident rate of young drivers is aggravated by alcohol. In the discussion of the effects of alcohol it was suggested that the lack of compensation for impaired performance may be the cause for the large role of alcohol in accident causation. Evidence was presented that drivers are unaware of performance decrements under alcohol which is possibly the cause for the absence of compensatory speed changes and effort. From the same perspective it may be suggested that young and inexperienced drivers have not yet learned to recognize the effects of situational factors on their perfor­mance and thus fail to compensate for these effects resulting in speeds that are too high for the circumstances”. Clearly this is an issue that needs to be investigated further. There are some indications that alcohol inhibits the perception of feedback from the driving task. The assumed lack of adaptation in young (male) drivers also needs to be explored further. The theory presented in this study offers a framework to examine these issues.

An important next step is the further validation and testing of the adaptation model. In the present study only a limited part of the model was tested. For example, the principle of effort allocation under forced paced conditions and the effects of this on operational performance need to be tested in further studies. The six experiments described in this study are only a first step in the direction of testing the limits and scope of the model of adaptation.


research driving simulators


There are numerous universities in the world that have a research group in Human Factors or traffic behaviour studies. A lot of these would benefit a lot if they would have a research driving simulator. This is a tool that enables researchers to perform measurements on human car driving behaviour in a controlled environment. The minimum hardware and software of a research driving simulator is:

  • fast computer with a GPU that has outputs for 3 monitors: one for the left, middle and right out-of-the-window views
  • a GPU with sufficient on-board memory and bandwidth so complex visual scenes can be rendered over a large horizontal field of view
  • pedals, buttons, shifter and steering wheel with a rotation of 900 degrees and sufficient force feedback
  • real time simulation software that renders traffic scenes
  • software for scenario generation and experiment generation
  • software for building and modifying virtual environments
  • software for data inspection and analysis.

The advantages of a driving simulator over experiments in the real world with instrumented vehicles are:

  • instrumented vehicles are extremely expensive and inflexible
  • in a simulator you have complete control over the environment, lighting and traffic situations
  • the number of variables that can be measured in a simulator are many more, compared to what can be measured in the real world
  • identical circumstances for all subjects in a simulator
  • safe environment to do your research.

Because you can control your environment and circumstances in a driving simulator, unexplained variance is lower in a simulator. Since the statistical power of tests is determined by the ratio of explained and unexplained variance (or statistical noise) , you need fewer subjects and can do more reliable measurements in a simulator. Also, experiments on the road that are potential dangerous for the subjects can be done safely in a simulator. That’s why studies on the effects of alcohol and drugs on driving behaviour or the effects of medication, prolonged driving, fatigue, or tests with people with neurological conditions can be done much safer in a simulator.