DEVELOPING MICRO-SIMULATION FOR THE ASSESSMENT OF SAFETY IN RELATION TO INTELLIGENT SPEED ADAPTATION

Jeffery Archer, Centre for Traffic Simulation Research, Royal Institute of Technology, P-100 44 Stockholm, Sweden

ABSTRACT
The development and use of micro-simulation for traffic safety assessment has been neglected or avoided by the traffic research community due to a number of practical and conceptual problems. These problems emanate largely from the availability of theories and methods that are required to design a representative model of road-user behaviour with enough microscopic detail to generate useful and relevant safety data, and also the difficulties in identifying safety indicators that can be measured through empirical enquiry and used for model validation and calibration. For safety assessment purposes, a model of road-user behaviour must incorporate sufficient behavioural variance to allow for less-than-perfect and erroneous performance. Furthermore, suggested safety indicators must be sensitive to changes in performance and behaviour brought about by use of Intelligent Transport Systems, such as Intelligent Speed Adaptation, which are predicted to have great safety potential in the urban environment. The SINDI-project at the Centre for Traffic Simulation Research in Stockholm focuses on the development and use of micro-simulation software for safety assessment. Emphasis is placed on qualitative and quantitative measures of road-user interaction (e.g. traffic conflicts) in small urban networks, and in particular the possibility to improve urban safety through the use of an ISA-system. For this purpose empirical data has been collected from a specific ISA test site in a small Swedish town using a variety of different measurement techniques. This data will be used in conjunction with micro-simulation in order to estimate projected safety effects at different levels of ISA implementation.

INTRODUCTION
In traffic engineering, micro-simulation has become 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 individual entities, i.e. road-users, at the microscopic level. The use of micro-simulation of the traffic system enables new and controversial measures to be tested without disrupting existing traffic networks, or putting people at risk. Furthermore, it has the ability to indicate the potential of alternative system designs at an early point in project planning, and can also provide a useful and cost-effective platform for establishing a balance between the different, and often opposing primary objectives of transportation engineering related to efficiency, safety, and environmental concerns. The majority of the microscopic simulators in existence today focus on network efficiency evaluation, often in relation to different traffic control strategies and the effects of Intelligent Transportation Systems. There are very few micro-simulators that have focused specifically on safety related issues. This factor was evident in review of existing micro-simulators in the SMARTEST project funded by the European Unions 4th Framework Program [1]. The review identified an implicit and substantiated need for greater levels of detail, improved calibration and validation, standardised performance indicators, and inclusion of vulnerable road users.

The SINDI-project focuses primarily on the safety problems of different road-user groups in the urban environment, particularly at intersections and roundabouts where safety problems are particularly eminent [2], and the potential of Intelligent Speed Adaptation (ISA) speed warning devices due to their high safety potential and applicability to the urban environment [3]. The SINDI-project has an established link with the ongoing ISA-trials in the Swedish town of Borlnge. Data has also been collected from a small urban network in the ISA-trial area and is currently undergoing detailed analysis in order that it may serve as a foundation for a great deal of the modelling and micro-simulation work discussed in this paper.

MODELLING MICROSCOPIC DRIVER BEHAVIOUR

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 small degree of behavioural variance between (but rarely within) individual drivers is often introduced by randomly sampling various statistical distributions that represent different behavioural variables (e.g. acceptable time-gaps in yielding situations). This tends adds more realism in the visual movement of vehicles but is generally arbitrary and has no particular significance to the simulation results. Furthermore, there is often little behavioural consistency between different variable distributions. Accidents or near-accidents are rarely possible owing to the use of deterministic rule-bases that govern how vehicles interact in a given traffic situations. For efficiency and environmental evaluation purposes, the lack of detail in behavioural modelling is not usually a significant issue. For safety assessment purposes however, this type of over-simplified driver behaviour is not sufficient to allow for realistic levels of behavioural variance in order to allow for erroneous and less than perfect performance. A more detailed behavioural approach requires modelling at a deeper (perhaps nanoscopic) level. This also infers the use and representation of exogenous and endogenous performance shaping factors that, in a given situation, can increase or reduce the risk for accident involvement. Driving is a highly complex task that requires continual adaptation to meet the demands of the dynamic and complex traffic environment. Some idea of the demands placed on the driver and the occurrence of errors and accidents has been suggested in a novel and interesting report presented by Hkkinen and Luoma [4] based on data from Finland and the United States (see Table 1). According to their statistics, average driver performance results in some 30 errors per hour, following some 7200 observations, 2400 decisions, and 1800 actions. 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
Event pieces of traffic inform. driver observations driver decisions driver actions driver errors risky situations near accidents accidents injury accidents fatal accidents 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 month 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

Microscopic driver behavior modelling of the type required in the SINDI-project inevitably involves the inclusion of probabililistic error mechanisms. This also suggests a modelling approach based on assumptions from different theories and models related to cognitive information processing and human factors research. The idea being to allow fundamental and established types of errors to occur in accordance with estimated probability distributions, also allowing for a number of well-known internal (e.g. driving experience/lack of experience) and external factors (e.g. situational complexity) to influence these probabilities marginally, either negatively or positively. The probability that any particular generated error will eventually lead to an incident or accident, depends to a very large extent on the dynamics and complexity of the prevailing traffic context, just as it does in real-world situations. It is hoped that the proposed preliminary error probability values can be refined and calibrated through further research at the individual level using driving simulators and/or instrumented vehicles. The model of road-user behaviour suggested in the SINDI project is typical of the information processing approach (see figure 1). The first stage involves the pre-attentive selection of contextual data through sensory processes and emphasizes the limitations of human sensory perception and attention. This is followed by a second stage that involves early attentive cognitive processing, in which the perceptual data is given meaning and stored in working memory. The third stage involves active decision-making, where the perceptually acquired information is assimilated and assessed in accordance with the demands of the environment and the need to maintain perceived risk at an acceptable level. Previous experience, motivation, and other more common long-term performance shaping factors are allowed to influence the decision-making process and thereby the subsequent choice of appropriate action. Similarly, shorter term situational factors, such as the prevailing level of complexity in the system, or the current level of stress or time-pressure experienced by the road-user, are also taken into consideration.

Figure 1. The theoretical model of microscopic road-user behaviour for safety simulation.
3

A key element of this model lies in its ability to allow different types of errors to occur at different levels of perceptual and cognitive processing. The errors types shown in the model are classified according to a taxonomy originally suggested by the human factors researcher James Reason [5]. To enable random error occurrence, perceptual data is processed and stored temporarily in a function similar to that of cognitive working memory. This storage represents perceived object data including position, relative speed, and distance to other static and dynamic objects, and also allows for errors such as: (1) false sensation (2) attentional failures (due to an incorrect or inconsistent sampling of the environment, or a lack of sampling during a shorter time-period due to e.g. distraction, stress, fatigue); (3) errors of judgement such as the incorrect assessment of speed and distance (in accordance with a suitable distortion function) (4) recognition failures (5) memory lapses where information is forgotten (6) inaccurate or blocked memory recall For the representation of higher level cognition functions (i.e. knowledge, skill, and decisionmaking), different types of reasoning errors are introduced through the implementation of cognitive rules in the form of specific pre-formatted behavioural schemas for different traffic situations [6]. Depending of the structure and selection technique, schemas can also be used to represent different long-term performance shaping factors (e.g. level of experience). Schemas can be regarded as an extension of the existing rule-based modelling approach, although they also include a stochastic element related to occurrence of what can be termed rule-based and knowledge based errors. Errors can also occur due to failures at lower levels of perceptual processing, in which case the filling-in of missing information is performed in accordance with the schema, resulting in either recovery or cognitive errors. The use of schemas also allows for more advanced forms of error recovery. Estimating the relative error probabilities for the different error types in each part of the information processing sequence, and the effects of complexity and the long-term performance shaping factors is a very difficult task owing to a serious lack of relevant individual road-user data specific to different driving tasks. For modelling purposes error type probabilities based on previous research are assumed. One such error type distribution has been suggested by Sanders and McCormick [7] specifically for decision errors, recognition errors, performance errors, and critical non-performance errors (see figure 2).

Figure 2. Probabilities for different error types suggested by Sanders and McCormick
4

The SINDI behavioural model initially assumes that all road-users are equally variant in terms of their performance. The consequences of this assumption depend largely on the purposes for which the model is intended and designed, and the expected results. The same type of assumption is also implied for susceptibility to short-term situation specific factors, such as environmental complexity and stress and/or time-pressure. The long-term performance shaping factors such as experience and motivation might also be ignored for the purposes of the suggested model, however modelling the differences among for example, experienced and inexperienced drivers, or aggressive drivers who are less motivated to drive safely and careful drivers, can be (and often is) used to introduce a degree of variance in road-user behaviour and performance along with demographic variables such as age and gender [8,9]. It is hoped that a number of detailed studies involving the use of driving simulators and instrumented vehicles in conjunction with urban road networks and the use of ISA will aid in providing a more detailed and empirically sound foundation on which to model road-user behaviour and performance and the complex effects of endogenous and exogenous influences.

DBI-HUTSIM MICROSCOPIC SIMULATOR

The driver behaviour model suggested above has been incorporated into a specially redesigned version of the high-fidelity time-based HUTSIM microsimulator developed at Helsinki University of Technology [9]. This new version, referred to as DBI (an acronym for Driver Behaviour and Interaction) is also based on an object-oriented design and development approach, but places particular emphasis on the modelling of highly detailed road-user behaviour and performance in interaction with the roadway environment (see figure 3).

Figure 3. Picture of the DBI-HUTSIM simulation environment, detailed information is shown for the vehicle marked with a circle in a secondary status window. The vehicle information includes details of objects in the immediately surrounding environment.
5

DBI not only concentrates on achieving a good representation of driving behaviour and performance, but also on important details related to: vehicle dynamics; various Intelligent Transportation Systems such as ISA; and roadway design and control. Each modelled network area requires rigorous calibration and validation based on a large number of diverse empirical site-specific measurements. The microsimulator also has a number of limitations. Most restrictive is the lack of lateral movement, which has implications for some measures of safety and particularly the severity of simulated accident outcomes. It is therefore critical to model deceleration, which is known to be the most common form of evasive action [10], to a high level of accuracy for each vehicle and driver combination.

TRAFFIC SAFETY MEASUREMENT AND EVALUATION

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. A better approach, from an empirical enquiry point of view is to focus on measures of the quantity and the quality of road-user behaviour, communication, and interaction in order to obtain an indication of prevailing traffic safety levels at any given traffic site [10]. The relationship between behaviour, errors, conflicts, and accidents is not trivial as pointed out by Klebelsburg [11] in the 1960s, but has a direct relevance for the SINDI-project (see figure 4).

Figure 4. The relationships between errors, standard behavior, traffic conflicts and accidents Svensson [12] states that for proxy measures or indicators of safety to be useful they must: complement accident data and be more frequent than accidents; have a statistical and causal relationship to accidents; and 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 [13]. Conflicts are described by the estimated time-to-accident(TA) value derived from the difference between the point in time at which one of the involved 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 road-user interactions if speed is also taken into consideration [12]. Other useful techniques include time-to-collision (TTC) and post encroachment time (PET) [14]. 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. Research has shown that TTCs less than 1.5 seconds, and PETs less than 1 second are critical for safety in urban areas. An extension has been suggested to the classical version of TTC that is particularly useful for assessing safety and for representing safety-critical probabilities for micro-simulation purposes [15]. The extended technique takes into account the full course of vehicles over space and time allowing a more comprehensive picture of prevailing safety levels for a particular traffic site.

There are also a number of other safety indicators with an established level of validity. The most important of these concerns speed and speed variance[16]. 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 outcomes is particularly important for the evaluation of ISA. Other important safety relevant variables are: headway distances or time-gaps for vehicles in car-following mode; traffic flow, where high levels of traffic flow imply the need for more frequent interaction; manoeuvre patterns and different manoeuvre types; and factors related to traffic engineering design and type of control [17].

DBI-HUTSIM SAFETY OUTPUT

During the simulation of an intersection, brief safety output is generated at run-time in relation the occurrence of severe conflicts, TTCs, PETs, and accidents in conjunction with the composition and numbers of road-user types, traffic flows, and average free-flow speeds (see figure 5). More detailed information concerning the exact nature of road-user interactions and other relevant information is also written to special output files for further safety analysis.

Figure 5. Run-time safety output generated from the simulation of a traffic site

DATA COLLECTION
For data collection purposes a small but representative urban test network has been selected in the Borlnge ISA trial area. Initially, more traditional traffic measurement techniques have been used to collect data in relation to speed, traffic flows, car following, and gap-acceptance behaviour. For the measurement of safety, conflict studies have been conducted at various safety critical sites in the network in relation to the interactions of different road-users classes. These conflict studies have been complemented by video-recordings, which also allow TTC values to be calculated. Safety data in the form of TTC and TA values has also been collected using a new technique that involves radar sensors. Information regarding reported accidents has also been obtained to get an indication of the characteristics and safety level of the test site. A further source of date is provided by the ISA-equipped vehicles that are continuously logged during the ISA trial period of approximately 12 months. The logged information includes speed, position, and acceleration/deceleration, and is collected at rates of up to 1 Hz.

FINAL COMMENTS
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 DBI-HUTSIM microsimulator. However, at the current stage of development, the SINDI-project shows great promise in relation to the evaluation of the ISA speed warning system used in Borlnge at varying levels of implementation, and has significant implications for traffic safety research related to the use of microsimulators in the future.

REFERENCES
1. Algers S, Bernauer E, Boero M, Breheret L, Di Taranto C, Dougherty M, Fox K, & Gabard J F (1997). Review of Micro-Simulation Models. SMARTEST Project Deliverable D3, Institute for Transport Studies, University of Leeds, Leeds, U.K. Archer J, & Vogel K (2000). The Traffic Safety Problem in Urban Areas., Research Report CTR2000-04, Royal Institute of Technology, Stockholm, Sweden. 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. Hkkinen S, & Luoma J (1991). Liikennepsykologia (Traffic Psychology). Karisto Oy, Hmeenlinna, Finland, p.38. Reason J (1990). Human error. Cambridge University Press, Cambridge. Sears D O, Peplau L A, & Taylor SE (1991). Social Psychology, Prentice-Hall International, 51-52 & 82-88. Sanders E, & McCormick M (1992). Human Factors in Engineering and Design, McGraw Hill International Editions. Sayed T, Brown G, & Navin F (1994). Simulation of Traffic Conflicts at Unsignalized Intersections with TSC-Sim, Accident Analysis & Prevention, Vol 26:5, pp.593-607. Kosonen I (1999). HUTSIM Urban Traffic Simulation and Control Model: Principles and Application. Publication 100, Helsinki University of Technology, Espoo, Finland. Risser R (1985). Behaviour in Traffic Conflict Situations. Accident Analysis and Prevention, Vol.17:2, pp.179-197. Klebelsberg D (1964). Derzeitiger Sand der Verhaltensanalyse des Kraftfahrens. Zrbeit und Leitsung, Ablt. Arbeitswissenscaft soziale betriebspraxis, 18, 33-37 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. 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. van der Horst R, & Kraay J (1986). The Dutch Conflict Observation Technique DOCTOR. Proceedings of the workshop Traffic Conflicts and Other Intermediate Measures in Safety Evaluation, Budapest, Hungary. Minderhoud M M, & Bovy P H L (2001) Extended time-to-collision measures for road traffic safety assessment, Accident Analysis and Prevention, Vol 33, pp.89-97. Baruya A, & Finch D J (1994). Investigation of Traffic Speeds and Accidents on Urban Roads. Proceedings of the 22nd European Transport Forum, Warwick University, Warwick, UK. Englund A, Gregersen N P, Hydn C, Lvsund P, & berg L (1998). Trafikskerhet. En kunskapsversikt, Stockholm: Studentlitteratur, KFB, pp.153-172.

2. 3.

4. 5. 6. 7. 8. 9. 10. 11. 12.

13.

14.

15. 16.

17.