Overview
- Date:Starts 31 March 2026, 14:00Ends 31 March 2026, 17:00
- Location:Virtual Development Lab, Chalmers Tvärgata 4C
- Opponent:Assoc. Prof. Simeon Calvert, Delft University of Technology, The Netherlands
- ThesisRead thesis (Opens in new tab)
Road crashes are a major cause of deaths and serious injuries worldwide. New technologies can reduce their number and severity by supporting drivers with advanced driver assistance systems (ADASs), and by taking over the entire driving task—at least under certain conditions—with automated driving systems (ADSs). Virtual simulation is one of the methods for predicting how safe these systems will be once they are released on public roads. However, ensuring that this method provides meaningful and representative results remains challenging. Safety targets are required to ensure that ADAS and ADS assessments are effective, relevant, and fair and that the systems have a positive impact on safety. This thesis focuses on the foundations for formulating and assessing these safety targets, with the aim of supporting the development of ADSs and ADASs, and ultimately improving traffic safety.
The work addresses four main aspects of safety targets. First, the thesis investigates the impact of data selection on the outcomes of virtual safety assessment. The findings indicate that crashes artificially generated from these data can differ substantially from real-world crashes, leading to lower severity outcomes, reduced criticality, and inaccurate benefit estimations. Second, the thesis evaluates the safety performance of the reference models described in UN Regulation No. 157 to determine whether they represent adequate safety targets for ADSs. A comparison of the models’ responses to those of real drivers in safety-critical scenarios reveals that the models do not perform like the competent and careful drivers they are intended to represent. Third, the thesis analyzes lane-changing behavior and its relation to the surrounding driving context. The results describe characteristics of lane-changing behavior that can be used in modeling, including a modified definition of lane-change initiation that incorporates a lateral speed threshold (in addition to the lateral position threshold used to define lane changes in current reference driver models). Finally, the thesis investigates the link between off-road glance behavior and crash risk for car-following scenarios. In line with previous literature, our results suggest that drivers adapt only modestly to time gap-related crash risk, yet they reduce both the frequency and duration of off‑road glances as time to collision gets shorter.
The findings highlight issues in current regulatory behavior model-based safety targets and the challenges that current reference models face, both in their formulation and in the data and methods used to assess safety target validity. Moreover, the findings also suggest that using observed behavior as the sole basis for safety targets for ADASs is problematic, although including components such as urgency and glance behavior may improve their performance and relevance. Overall, the results highlight the importance of ensuring relevant and valid virtual safety assessments through a well-considered choice of data sources and the robust and accurate representation of safety targets through driver models.
The work addresses four main aspects of safety targets. First, the thesis investigates the impact of data selection on the outcomes of virtual safety assessment. The findings indicate that crashes artificially generated from these data can differ substantially from real-world crashes, leading to lower severity outcomes, reduced criticality, and inaccurate benefit estimations. Second, the thesis evaluates the safety performance of the reference models described in UN Regulation No. 157 to determine whether they represent adequate safety targets for ADSs. A comparison of the models’ responses to those of real drivers in safety-critical scenarios reveals that the models do not perform like the competent and careful drivers they are intended to represent. Third, the thesis analyzes lane-changing behavior and its relation to the surrounding driving context. The results describe characteristics of lane-changing behavior that can be used in modeling, including a modified definition of lane-change initiation that incorporates a lateral speed threshold (in addition to the lateral position threshold used to define lane changes in current reference driver models). Finally, the thesis investigates the link between off-road glance behavior and crash risk for car-following scenarios. In line with previous literature, our results suggest that drivers adapt only modestly to time gap-related crash risk, yet they reduce both the frequency and duration of off‑road glances as time to collision gets shorter.
The findings highlight issues in current regulatory behavior model-based safety targets and the challenges that current reference models face, both in their formulation and in the data and methods used to assess safety target validity. Moreover, the findings also suggest that using observed behavior as the sole basis for safety targets for ADASs is problematic, although including components such as urgency and glance behavior may improve their performance and relevance. Overall, the results highlight the importance of ensuring relevant and valid virtual safety assessments through a well-considered choice of data sources and the robust and accurate representation of safety targets through driver models.
Pierluigi Olleja
- Visiting Researcher, Vehicle Safety, Mechanical Engineering