Published in ESS Conference Proceedings, Hamburg, 2000

THE POTENTIAL OF MICRO-SIMULATION MODELLING IN RELATION TO TRAFFIC SAFETY ASSESSMENT

Jeffery Archer 1 and Iisakki Kosonen 1, 2 Centre for Traffic Simulation Research (CTR), Royal Institute of Technology, P-100 44 Stockholm, Sweden (2) Laboratory of Transportation Engineering, Helsinki Univeristy of Technology (HUT), FIN-02150 Espoo, Finland E-mail: jeffery.archer@infra.kth.se, iisakki.kosonen@hut.fi
(1)

KEYWORDS
Traffic safety, road-user behaviour modelling, safety parameter identification, dynamic interaction simulation.

ABSTRACT
The use of micro-simulation modelling for the assessment of road traffic safety has been largely neglected by the traffic research community due to practical and conceptual problems. These problems emanate largely from the choice of methods and properties that are required to design a representative model of driver behaviour with enough detail to provide useful and relevant safety data, and also the difficulties in identifying safety indicators. For safety assessment purposes a driver model must incorporate an appropriate degree of behavioural variance to allow for lessthan-perfect perception, decision-making and action, thereby giving rise to different levels of accident risk in the interactions of road-users in different traffic situations. The recently started SINDI-project at the Centre for Traffic Simulation Research, Stockholm, aims to break new ground in road traffic micro-simulation for safety assessment. The primary focus is on traffic safety related to road-user interactions at urban intersections, and the possibilities to improve safety through Intelligent Transport Systems. Most critical to the success of the project is the establishment of a representative model of driver behaviour, and the identification of useful and relevant safety indicators that can be generated in the simulation and measured through empirical enquiry for model validation and calibration.

A recent state-of-the-art review of existing road transport related micro-simulation models was undertaken as part of the SMARTEST project funded by the European Unions 4th Framework Program (Algers, et.al., 1997). The main report from this project indicated the existence of some 58 different micro-simulation models, the majority of which had been designed specifically for the purpose of evaluating system efficiency related to the impact and sensitivity of different roadway design strategies. Alternative objectives included the strategic effects of various Intelligent Transport Systems (ITS) and traffic management and travel information systems. One of the main objectives of the SMARTEST project was to identify gaps in existing micro-simulation models (i.e. important traffic related issues for which little or no support was provided, but where an implicit and substantiated need was recognised for their inclusion). Importantly, distinct gaps were identified in relation to: safety evaluation modelling, levels of detail, standards of calibration and validation, standardised performance indicators for the evaluation of output, and the inclusion of pedestrians and cyclists. Why is it then that micro-simulation modelling has seldomly been used for traffic safety assessment? The existing traffic safety problem within the European OECD countries showed a staggering 45000 fatalities, and 1.3 million injuries reported from road traffic accidents during 1997, with the estimated costs of such accidents within the European Union in the region of 160 billion ECU, thereby far exceeding the total annual budget for the EU in this year. Most of the traffic safety problems are attributed to the erroneous behaviour of individual road-users. Research related to prospective traffic safety is therefore highly motivated, and arguably well-suited to methods that have the potential to evaluate the impact of safety countermeasures a priori through the simulation of road-user behaviour at the microscopic level. Another relevant aspect is the fact that it has only recently been possible to achieve the high level of fidelity that is required for this kind of simulation. Increases in computer power and capacity, in combination with new and innovative modelling techniques, have opened the door for a new era of high fidelity micro-simulation modelling. In addition, new technology has also led to improvements in data-collection techniques, providing a wealth of detailed empirical data to

INTRODUCTION
In traffic engineering, micro-simulation has proved to be a particularly useful tool for studying the traffic system where the behaviour of the system as an entirety is largely dependent on the behaviour and interactions of entities, i.e. road-users, at the microscopic level. The use of microsimulation of the traffic system enables new and sometimes controversial measures to be tested without disrupting existing traffic networks, or putting people at risk. Through its ability to indicate the potential of alternative system designs at an early point in project planning, it can also provide a useful and cost-effective platform for establishing a balance between the different, and often opposing system objectives of efficiency, safety, and environmental concerns.

Published in ESS Conference Proceedings, Hamburg, 2000

promote more substantiated model design and development, and ease the burden of validation and calibration. One of the main reasons attributable to the lack of microsimulators that aim at safety assessment concerns the fact that detailed modelling of road-user behaviour in relation to the traffic system requires greater in-depth knowledge of a far more diverse and multi-disciplinary nature. While simple car-following and gap-acceptance driver sub-models are sufficient to provide a normative behaviour for traffic system capacity assessment. The evaluation of safety demands a more complex driver model that incorporates a representative degree of variance, that allows errors to occur as the result of less-than-perfect perception, decisionmaking, and action, thereby causing different levels of risk in the interactions between road-users and the environment. Furthermore, the study of safety at, for example, isolated intersections is hindered by the fact that accidents, the most tangible and accepted measure of traffic safety, occur so seldomly. Therefore the study of safety is dependent on the identification of a number of useful and relevant safety indicators that, as a proxy measure, have a validated relationship to accidents and provide some indication of their outcome in terms of severity.

a)

the establishment of a representative model of road-user behaviour that allows for a sufficient degree of variance in road-user perception, decision-making, and action b) identifying suitable and relevant safety indicators that can be used as a foundation for safety assessment (preferably by panel of safety experts) Specifically for the purposes of the SINDI-project, a specially adapted version of the Windows-based HUTSIM microscopic simulator developed at the Helsinki University of Technology is used (Kosonen, 1999). The HUTSIM simulator is based on object-oriented design and implementation, and provides very high fidelity both in the representation of traffic networks, and dynamic interactions between different road-users and the traffic environment. In the following sections of this paper the establishment of a representative model of road-user behaviour, and the identification of suitable and relevant safety indicators are discussed in conjunction with the work currently underway on the adaptation of HUTSIM for safety assessment.

ASSESSING TRAFFIC SAFETY

Traditionally, the objective safety level of the traffic system is measured by the number of police reported accidents and the severity of their outcomes in terms of personal injury and fatality. As previously mentioned, the collection of accident data relative to a particular urban intersection is impractical and of only limited value for safety research (Risser, 1985). A far better approach, from an empirical enquiry point of view, is to focus on measures of the quantity and quality of road-user behaviour, communication, and interaction in order to obtain an indication of prevailing traffic safety levels in any given intersection (Risser, 1985; Hydn, 1987). Svensson (1998) states that for proxy measures or indicators of safety to be useful they must: a) complement accident data and be more frequent than accidents b) have a statistical and causal relationship to accidents c) have the characteristics of near-accidents in a hierarchical continuum describing the severity levels of road-user interactions, where accidents are placed at the highest level and very safe passages, with a minimum of interaction, are found at the lowest level These preconditions are found in, amongst others, the Swedish Traffic Conflict Technique that has been developed at Lund University (Hydn, 1987). The term conflict is generally defined as an observed situation in which two or more road-users approach each other on a collision course, and where an accident is imminent if neither takes evasive action. Conflicts are described by the estimated time-toaccident(TA) value derived from the difference between the point in time at which one of the road-users takes evasive action, and the estimated time of the collision had it taken place. The TA measure also provides a measure severity in the interactions of road-users if speed is also considered. Recently, Svensson (1998) has extended the use of the TASpeed relationship, to describe the shape of the severity

THE SINDI-PROJECT APPROACH

The SINDI-project (SINDI is an acronym for Safety INDIcators) intends to break new ground in relation to use of micro-simulation as a tool for safety assessment purposes. SINDI focuses on the safety problems of different road-user groups, including pedestrians and cyclists, at different types of urban intersections (e.g. three-way or two-way intersections and roundabouts), with different forms of traffic control (e.g. yield or stop-signs, traffic-lights). Research has shown that roughly one-third of all fatal accidents and more than two-thirds of all reported injury accidents occur in urban areas, most commonly at intersections with a considerably higher number of vulnerable users involved than anywhere else (Archer & Vogel, 2000). A number of interesting studies have been conducted with a view to finding the causes of urban accidents. One of the most comprehensive studies covered 1254 injury accidents over a one-year period in northern England (Carsten, et.al., 1989). Findings suggested that road-users were often unable to, or failed to, anticipate other road-users actions, and often failed to yield. These errors were found to be due, at a lower level, to perceptual and cognitive factors such as failing to look, failing to see, and a lack of judgement. The long-term goal of the SINDI-project is to be able to evaluate the potential of different safety countermeasures such as those that fall under the umbrella-term Intelligent Transport Systems (ITS). Intelligent Speed Adaptation (ISA) is of particular interest due to its high safety potential and suitability to the urban environment (Vrhelyi, 1996). In SINDI the two most important issues that are critical to the success of the project concern:

Published in ESS Conference Proceedings, Hamburg, 2000

hierarchy for a particular intersection by including all interactions between road-users and not just those that are considered as severe or serious (i.e. traffic conflicts). The frequencies of interactions in relation to different levels of severity can be used to represent the safety-related behaviour of road-users at specific intersections (see figure 1). Furthermore, the interaction-frequency - severity-level distributions can be disaggregated to focus on, for example, different road-user types, or specific manoeuvre types.

25 Interaction Frequency 20 15 10 5 0 0 5 10 15 20 25 30 Intersection A Intersection B

the average headway distances or time-gaps for vehicles in car-following mode (Grayson, 1984) b) the traffic flow, where high levels of traffic flow imply the need for more frequent interaction, and therefore increased accident risk potential, It has found that the traffic flow and interaction-severity relationship is not entirely straight-forward (Ekman, 1996) c) the manoeuvre patterns and proportional numbers of different manoeuvre types (Svensson, 1998) d) the traffic engineering design and type of control that is imposed, which consequently has an effect on such safety aspects as driver visibility but also defines the possibilities and limitations for road-user interaction (OCinnide & Murphy, 1994) In the final assessment of the proxy safety indicators it is also useful to obtain as much information as possible in relation to the actual reported accident data during recent years to get an indication of the general characteristics and level of safety at the intersection. The conflict technique provides the possibility to estimate the expected number of accidents, within levels of error variance, and therefore provides a useful means of verifying the empirical conflict studies.

a)

Interaction Severity

Adapting HUTSIM for Safety Assessment

Using HUTSIM for safety assessment implies the collection of a large amount of empirical data collection to determine traffic-flows, average speed and headway distributions, timegap acceptance distributions for yielding and turning vehicles, vehicle type distributions, manoeuvre patterns, pedestrian and cyclist relevant information, and so forth. The data collection for safety assessment purposes must also include conflict studies and other relevant safety indicators of the type described above. Besides the more common methods involving inductive loops, pneumatic tubes and other vehicle registration techniques, SINDI uses additional data collected from experiments using instrumented vehicles that have been driven through the intersections that are studied. The main source of safety data comes from the direct on-site recording of conflicts, and concurrent videofilming of road-user interactions that enables the later study of, for example, less severe interactions, the occurrence of post encroachment times, and the behaviour of different types of road-users. During the simulation of an intersection, brief safety output is generated at run-time in relation the occurrence of severe conflicts, PETs, and accidents in conjunction with the composition and numbers of road-user types, traffic flows, and average free-flow speeds (see figure 2).

Figure 1.

Graph showing the different frequencies of interactions with different levels of severity in two intersections, intersection B shows a larger proportion of severe interactions

The Dutch conflict technique DOCTOR, includes both time-to-collision (TTC) and post encroachment time (PET), where TTC represents the smallest time-gap for two vehicles on a collision course at any point during the conflict, and PET represents the time-gap between the passage of two vehicles who are on a near-collision course to pass over a common point of conflict (van der Horst & Kraay, 1986). Research has shown that TTCs less than 1.5 seconds, and PETs less than 1 second are critical for safety in urban areas. There are also a number of other safety indicators with an established level of validity in relation to accident risk in the traffic system. The most important and well-researched of these is speed. The speed profile of an intersection should include, not only the average free-flow speeds of drivers on their approach, but also average turning speeds, levels of speed adaptation, and speed variance among drivers. The relationship between speed and safety in terms of accidents and their outcomes has been established by a vast amount of international research (e.g. Baruya & Finch, 1994, Vrhelyi, 1996). Other important safety relevant variables are:

Published in ESS Conference Proceedings, Hamburg, 2000

accidents are very rarely possible in simulation owing to the deterministic rule-base that governs how vehicles may or may not interact with one another in a given traffic situation. For safety assessment purposes the relatively strict deterministic rule-base that governs this type of simplified driver behaviour must be loosened in order to allow for a realistic amount of behavioural variance and the possibility for errors to occur. For this purpose, it is essential to know something of the factors that, in a given situation, can increase or reduce the risk for accident involvement. The consensus among the traffic safety community is that roaduser behaviour lies behind the majority of traffic accidents, although the drivers themselves are often not prepared to accept their own shortcomings, and often tend to consider themselves as more proficient than the average driver (Risser, 1985; Rumar, 1985). Driving is a task requiring continual adaptation to the demands of a dynamic and complex environment. Some idea of the demands and limitations of the driver can be derived from the information presented by Hkkinen and Luoma (1991) based on data from Finland and the United States (see table 1). Generally, there are many different models of driver behaviour, each with different advantages and limitations, and very few that have been verified by objective empirical enquiry (e.g. Michon, 1985).
Table 1. Events of an average driver in traffic based on an average speed of 60 kph/h and annual mileage of 200,000 km per time-unit 5 in 1 sec 2 in 1 sec 40 in 1 min 30 in 1 min 1 in 2 min 1 in 2 hrs 1 in 1 mon 1 in 7.5 years 1 in 100 years 1 in 2000 yrs per kilometre 300 per Km 120 per Km 40 per Km 30 per Km 1 per 2 km 1 per 120 km 1 per 2000 km 1 per 150,000 km 1 per 2 million km 1 per 40 million km

Figure 2. The run-time safety output generated from the simulation of an intersection in HUTSIM

More detailed information concerning the exact nature of road-user interactions and other safety relevant information is also written to specially prepared HUTSIM output files for further safety analysis. Depending on the accuracy and validity of the existing behavioural model that determines road-user behaviour, the output data for the road-user interactions (and more specifically interactions categorised as serious traffic conflicts), should match that obtained from the corresponding empirical studies, given acceptable predetermined error variance margins. The aim of SINDI is to generate output data that is of sufficient quality and quantity to be used for safety assessment, preferably by a panel of traffic safety experts. Providing it is possible to generate a suitable safety profile, the impact of various safety countermeasures, such as those commonly used in traffic engineering or different ITS applications, can then be tested.

Event pieces of traffic inform. driver observations driver decisions driver actions driver errors risky situations near accidents accidents injury accidents fatal accidents

MODELLING DRIVER BEHAVIOUR

For safety assessment related micro-simulation, the key to success lies in the ability to model road-user behaviour with a high degree of representivity, accuracy, and detail. In a typical micro-simulation model where efficiency is the main concern, the behaviour of the driver is usually represented by a number of fundamental sub-models that describe behaviour such as car-following, lane-changing, gap-acceptance, and obstacle detection. A degree of variance between (but rarely within) individual drivers is introduced by the inclusion of distributions that can be sampled by randomisation functions. This adds more realism in the movement of vehicles but is generally arbitrary and has limited value. Accidents and near-

Interestingly, Risser (1985) has systematically studied different types of errors in driving behaviour that are related to traffic conflicts. Most errors appear to be the result of a lack or misunderstanding of communication in the interactions of different road-users. Among the behaviours that were found to be related to conflicts among different drivers were: risky passing manoeuvres, badly adapted speed, following too closely, unlawful behaviour at traffic lights, hesitant or risky lane-changing, cutting corners, insisting on or taking others right of way, jerky steering, inadequate lateral distance ,and lack of precaution at intersections. One of the most comprehensive models of driver behaviour is that suggested by Rumar (1985), which has also found support in a number of empirical studies (see figure 3).

Published in ESS Conference Proceedings, Hamburg, 2000

Motivation

Experience Attention Expectation

Physical Environment

Sensory Processes

Short-term Memory

Decision-making is itself prone to variance in the model by allowing for lapses of attention, and the effects of inexperience or expectancy. Motivational factors such as a those related to risk-taking or an aggressive personal style are also included. The resulting actions are then performed after an appropriate period of reaction delay in accordance with the driver profile. Vehicle dynamics are also modelled in the HUTSIM micro-simulator.
Reaction Behaviour

Perceptual Structuring

Limited Channel

Decision Processes

Mechanical Filtering

Perceptual Filtering

Cognitive Filtering

At present, work is continuing with the development of the driver behaviour model at a level which can be described as nanoscopic. While the model itself is dependent on the accuracy of the information contained in behavioural parameter distributions, the possibilities of using fuzzy inference in the existing sub-models (car-following, gapacceptance etc.) are being explored in order to achieve more realistic quantitive and qualitative behaviour. The driver behaviour model incorporated in HUTSIM has been adapted to allow lapses of attention, by adding a random delay to the update frequency of certain vehicles that have received little perceptual input during a longer period of time. The inclusion of sensitivity functions in relation to the effect of time-pressure, visual complexity, anticipation or expectation, and visibility limitations, are currently being implemented along with structural changes in the model that permit similar types of errors to those suggested in the literature (e.g. failure to yield, failure to see, and lack of judgement). The models for other road-user types, such as pedestrians and cyclists, will be less complex in nature. The HUTSIM model has some limitations, most restrictive is the lack of lateral movement. This has however, no effect of the registration of conflict situations as they concern only the point in time and space at which aversive action is taken. This limitation might, have implications for the number of simulated accidents, although deceleration is known to be the most common form of evasive action (Risser, 1985).

Figure 3. Rumars model of driver information processing

The model for driver information processing suggested by Rumar, describes the way decisions are made on the basis of the drivers comprehension of the actual traffic situation, and how decisions are influenced by attention factors, motivation, previous experience, and expectations. Perceptual and cognitive filters are also provided to show how important information may be lost during the different stages of the decision-making process. Rumars model provides a theoretical foundation for the model of road-user behaviour (primarily the driver) that is under development in the adapted version of HUTSIM used in the SINDI-project (see figure 4). The SINDI driver behaviour model is at present slightly more simplified, but does include the effects of the superior functions (motivation, experience, attention, and expectation) in the form of added error variance (denoted e in the diagram), between the three main elements of processing, namely perception, decision-making, and action.

CONCLUSIONS
There is a considerable amount of work remaining in relation to the adaptation, validation, and calibration of the behavioural model and safety output components of the HUTSIM microsimulator. However, at the current stage of development, the SINDI-project shows great promise and has significant implications for future traffic safety research.

Perception
e +

Misjudgements (speed, dist) Environmental Complexity Visibility Restrictions Feedback Loop

Steering

Decision-Making

Speed Control

Incorrect Decisions (due to inexperience, expectancy) Attentional Lapses Motivational Factors

REFERENCES
Algers, S.; Bernauer, E.; Boero, M.; Breheret, L.; Di Taranto, C.; Dougherty, M.; Fox, K.; and Gabard, J.F. 1997. Review of MicroSimulation Models. SMARTEST Project Deliverable D3, Institute for Transport Studies, University of Leeds, Leeds, U.K. Archer, J. and Vogel, K. 2000. The Traffic Safety Problem in Urban Areas., Research Report CTR2000-04, Royal Institute of Technology, Stockholm, Sweden. Baruya, A. and Finch, D.J. 1994. Investigation of Traffic Speeds and Accidents on Urban Roads. Proceedings of the 22nd European Transport Forum, Warwick University. Warwick, U.K. Carsten, O. M. J.; Tight, M. R.; Southwell, M. T.; and Plows, B. 1989. Urban Accidents: Why Do They Happen? Basingstoke: AA Foundation for Road Safety Research, UK.

Action
e +

Reaction Delay Incorrect Control Handling Vehicle Characteristics

Vehicle Dynamics

Figure 4. The theoretical model of driving behaviour that is currently used as a platform for development in the HUTSIM microsimulator

Published in ESS Conference Proceedings, Hamburg, 2000 Ekman, L. 1996. On the Treatment of Flow in Traffic Safety Analysis A Non-parametric Approach applied on Vulnerable Road-Users. Bulletin 136. Dept. of Traffic Planning and Engineering, Lund University, Lund, Sweden. Grayson, G.B. 1984. The Malm study. A Calibration of Traffic Conflict Techniques. Institute for Road Safety Research SWOV, Leidschendam, Netherlands. Hkkinen, S. and Luoma, J. 1991. Liikennepsykologia (Traffic Psychology). Karisto Oy, Hmeenlinna, Finland, p.38. Hydn, C. 1987. The Development of a Method for Traffic Safety Evaluation: The Swedish Traffic Conflicts Technique. Dept. of Traffic Planning and Engineering, Lund University, Lund, Sweden Kosonen, I. 1999. HUTSIM Urban Traffic Simulation and Control Model: Principles and Application. Publication 100, Helsinki University of Technology, Espoo, Finland. Michon, J.A. 1985. A Critical View of Driver Behaviour Models. What do we Know, What Should we do? In Human Behaviour and Traffic Safety, L. Evans and R.C. Schwing, eds. New York: Plenum Press. OCinnide, D.; Dryselius, B.; Gutowski, A.; Gynnerstedt, G.; Risser, R.; McCaul, B.; and Merza, H. 1995. Specification and Evaluation of ATT Based Scenarios. DRIVE II Project, HOPES Deliverable 31. Risser, R. 1985. Behaviour in Traffic Conflict Situations. Accident Analysis and Prevention, Vol.17:2, pp.179-197. Rumar, K. 1985. The Role of Perceptual and Cognitive Filters in Observed Behaviour. In Human Behaviour and Traffic Safety, L. Evans and R.C. Schwing, eds. New York: Plenum Press. Svensson, . 1998. A Method for Analysing the Traffic Process in a Safety Perspective. Bulletin 166, Dept. of Traffic Planning and Engineering, Lund University, Lund, Sweden van der Horst, R. and Kraay, J. 1986. The Dutch Conflict Observation Technique DOCTOR. Proceedings of the workshop Traffic Conflicts and Other Intermediate Measures in Safety Evaluation, Budapest, Hungary. Vrhelyi, A. 1996. Dynamic Speed Adaptation Based on Information Technology: A Theoretical Background. Bulletin 142, Dept. of Traffic Planning and Engineering, Lund University, Lund, Sweden.