Can simulation predict traffic safety indicators? Transition from “No” to “Maybe” with the i4Driving human driver model

By Ali Nadi and Hans van Lint

The development of Connected and Autonomous Vehicles (CAV) and Advanced Driver Assistance Systems (ADAS) is rapidly progressing. As these technologies become more prevalent on our roads, guaranteeing their safety becomes of critical importance. The complex and dynamic nature of CAVs and ADAS demands rigorous safety assessments to address potential risks and hazards. These assessments involve comprehensive testing, simulation and validation processes to evaluate their reliability, functionality and ability to interact with other road users and infrastructure seamlessly.

i4Driving’s second innovation is to augment available models with a 4D cognitive layer, to make the most of unveiled patterns and existing behavioural and psychological theories, which is the focus of this blog post.

Current status of Traffic simulation models

Traffic simulation models are invaluable tools for designing and evaluating road transport systems in terms of efficiency and, under certain conditions, environmental impact. However, none of today’s traffic simulation models in practice or academia are valid to assess traffic safety effects. This is mainly due to two rather loosely connected bodies of knowledge, namely traffic flow simulation and traffic safety psychology.

In traffic flow theory, Human Driver Models are essentially mathematical constructs (sets of equations) that mimic human behaviour to control the longitudinal (free driving, car-following) and lateral (lane-changing, crossing) movement of vehicles in a virtual environment. These models mainly assume that drivers accelerate to their optimal velocity as a function of the distance headway. This opened the door to the variety of different longitudinal driving models which, in general, are named stimulus-response models. These models often require incorporating perception and anticipation processes, where the drivers’ model can observe the environment, anticipate the traffic conditions further downstream, and use reaction times dynamics to respond to changes in stimuli (e.g., through a set of actions that can be carried out at the decision point). A similar diversity of modelling approaches can be found for lateral driving behaviour, that controls when drivers change lanes, merge, and diverge. All these models are collision-free by design and fail to cater for the full heterogeneity of driving behaviour in the simulation environment. “Collision-free by design” implies the utilisation of strict control objectives or staying within reasonable parameter limits. These models are highly valuable when assessing, designing, and optimising road infrastructure. In these contexts, the primary focus is modelling the collective average behaviour of drivers, ensuring safe interactions. Consequently, today’s microscopic traffic simulation models are designed to replicate idealised, collision-free behaviour (provided they are well-calibrated). As a result, they may not be suitable for drawing safety-related inferences, for three primary reasons.

Firstly, accidents are infrequent occurrences and, as a result, there is limited quantitative empirical evidence available to study the actual mechanisms/factors leading to them. Given one accident in 106 km and 4 to 10 fatality in 109 km, all interactions are statistically safe.

Figure 1. Distribution of safe and unsafe interaction
Figure 1. Distribution of safe and unsafe interaction

Secondly, there are many plausible mechanisms that lead to accidents: perception misjudgment, errors, and risk taking, physiological, emotional, and mental state, distraction in and outside the vehicle, human interactions, and technological failure, for example (see figure 2 below).

Figure 2: Mechanisms that may lead to accidents
Figure 2. Mechanisms that may lead to accidents

Third, there is huge inter- and intra-driver heterogeneity. Human factors exert a pivotal influence on human driving traits, contributing to a wide spectrum of driving behaviours that can potentially lead to safety-critical situations.

i4Driving multi-scale framework for microscopic traffic simulation

To deal with these challenges, we have developed a generic multi-scale framework for microscopic traffic simulation in the i4Driving project. this framework empowers us to simulate scenarios in which various human factors such as perception errors, risk-taking tendencies, and distractions among simulated drivers lead to interactions resulting in critical traffic situations or accidents. By leveraging this modelling framework, we can unveil and forecast the diversity and complexity inherent in human driving behaviours, thus providing a dependable and lifelike virtual evaluation environment for assessing road safety within the context of mixed traffic.

Figure 3 shows the conceptual framework of the i4Driving human-driver model. In this framework the operational collision-free models are disentangled from all the information processing that take place at strategic, tactical, and/or cognitive levels.

In principle, stimulus and response processes are governed by an information processing mechanism, in which the simulation environment is observed by the driver model and translated into anticipated stimuli, such as distance gaps and speed differences, and respective responses, such as acceleration/deceleration, and steering. These processes can be subject to driving traits and all relevant mental states, such as skills, fatigue, age, gender, and preferences. Such human factors can be included in human-driver models either exogenously or endogenously.

The model contains two mental state variables: situational awareness and driver task demand. Situational awareness encompasses various dimensions of perception, including focus and distraction, as described in research by Endsley (1995) and Wickens (2008). On the other hand, driver task demand is a concept frequently used in studies to explain driving performance under a wide range of conditions, as seen in the works of Precht et al. (2017b) and Teh et al. (2014). In simpler terms, task demand represents the cumulative mental workload imposed on a driver by various cognitive tasks they are engaged in, while situational awareness gauges how well a driver is aware of their surroundings, especially the stimuli necessary for safely and efficiently executing the driving task.

In our framework, both task demand and situational awareness are dynamic, meaning they change over time and space. These dynamic mental state variables have a direct impact on various driving parameters, such as reaction time, the frequency and magnitude of perception errors, and how drivers respond to factors like distance gaps.

The model also includes plausible mechanisms that can consider human’s capabilities in multi-tasking through anticipation reliance. The notion of anticipation reliance is related to how people (think they can) multitask. In this notion, driver do primary driving task with full focus and rely on their anticipation (which can be wrong!) while driving with secondary task. The following illustrative example shows, how the i4Driving human driver model works in a car-following scenario.

Figure 4: Illustrative example of i4Driving human driver model (van lint and Calvert, 2018)
Figure 4. Illustrative example of i4Driving human driver model

In this scenario, under light traffic conditions, the driving task demand (TD) is significantly lower than the driver’s task capacity (TC). However, in dense traffic conditions with a shorter time headway, the task demand approaches the driver’s task capacity (van Lint and Calvert, 2018).

At time stamp 3, the driver receives a phone call, introducing a secondary task. Initially, the phone call’s task demand is high but gradually decreases over time until it reaches a certain threshold. During the phone call, the driver’s task demand surpasses their capacity, resulting in saturated driver workload. This can lead to a diminished situational awareness due to perception errors (van Lint and Calvert, 2018).

In response, the driver relies on anticipation and adapts their behaviour by increasing the time headway to alleviate the workload. Once the call ends, the driver accelerates to their desired speed (Calvert el at., 2020).

Within this framework, we can simulate various aspects of social interaction during driving, such as tailgating, cooperative gap creation, and competitive gap seeking. These interactions can give rise to potentially unsafe situations that may culminate in accidents or near-critical conditions (Schakel et al., 2023). In the i4Driving human driving model, we can model heterogeneous driving behaviours by calibrating the fundamental diagram of task demands. Alternatively, we can achieve this by externally manipulating task capacity.


In the i4Driving project, we bridge two realms of research: traffic flow theory and traffic safety psychology. Our objective is to create a robust library of models that can potentially enable us to forecast traffic safety indicators in environments where both human drivers and autonomous vehicles coexist. This shift from the inability to predict to the possibility of predicting traffic safety represents a significant leap forward in the field. It also lays the foundation for future regulatory and policy frameworks governing autonomous vehicles.


  • Endsley, M. R. (1995). Toward a theory of situation awareness in dynamic systems. Human factors37(1), 32-64.
  • Calvert, S. C., Schakel, W. J., & van Lint, J. W. C. (2020). A generic multi-scale framework for microscopic traffic simulation part II–Anticipation Reliance as compensation mechanism for potential task overload. Transportation Research Part B: Methodological140, 42-63.
  • Precht, L., Keinath, A., & Krems, J. F. (2017). Identifying the main factors contributing to driving errors and traffic violations–Results from naturalistic driving data. Transportation research part F: traffic psychology and behaviour49, 49-92.
  • Schakel, W., Knoop, V., Keyvan-Ekbatani, M., & van Lint, H. (2023). Social Interactions on Multi-Lane Motorways: Towards a Theory of Impacts.
  • Teh, E., Jamson, S., Carsten, O., & Jamson, H. (2014). Temporal fluctuations in driving demand: The effect of traffic complexity on subjective measures of workload and driving performance. Transportation research part F: traffic psychology and behaviour22, 207-217.
  • Van Lint, J. W. C., & Calvert, S. C. (2018). A generic multi-level framework for microscopic traffic simulation—Theory and an example case in modelling driver distraction. Transportation Research Part B: Methodological117, 63-86.
  • Wickens, C. D. (2008). Situation awareness: Review of Mica Endsley’s 1995 articles on situation awareness theory and measurement. Human factors50(3), 397-403.

Author information
  • A. (Ali) Nadi Najafabadi, postdoctoral researcher at Delft University of Technology (TU Delft), the Department of Transport and Planning, Faculty of Civil Engineering and Geosciences.
  • J.W.C. (Hans) van Lint, Professor of Traffic Simulation and Computing at TU Delft.

Leave a Comment

Your email address will not be published. Required fields are marked *