An accident data-driven scenario generation methodology for testing automated driving systems in a mixed traffic environment

  • Simon Perveen

Student thesis: Doctoral ThesisDoctor of Philosophy


To reduce road traffic accidents caused by human drivers, more and more vehicles
equipped with automated driving systems (ADSs) (SAE Level 3 ≥) are released. Since
ADSs replace human drivers, the expectation on their safety performance is very high. It
has been predicted that millions and billions of miles are required to be driven to estimate
the safety performance of ADSs. The industry aims to implement scenario-based test
strategies in virtual environments to reduce this effort. However, the parameter space for
scenario specification is multi-dimensional. Hence, a guided search is required to identify
solely the critical scenarios to avoid testing irrelevant scenario combinations.
This thesis proposes a non-simulation-based scenario generation method using the
STATS 19 accident data to identify critical scenarios that occurred under adverse
environmental conditions. The methodology is based on two studies and focuses on
mixed traffic scenarios in intersections. Adverse environmental conditions are known to
reduce the safety performance of ADSs. However, due to their low frequency, such
patterns are hidden in accident data. Hence, the first study uses the k-medoids clustering
algorithm to extract hidden patterns for scenario specification from the STATS 19 data.
Important information, such as vehicle dynamics, is missing in the STATS 19 data.
Vehicle dynamics, however, are essential to test the STATS 19 scenarios in a virtual
environment. Defining random vehicle dynamics would end in a large number of
irrelevant tests. Hence, the second study aims to optimise the search for relevant vehicle
dynamics parameters with high injury and collision risks. The second version of the Nondominated Sorting Genetic Algorithm (NSGA-II), a metaheuristics-based optimisation
algorithm, has been used to carry out this expensive search activity.
For scenario specification, the k-medoids algorithm successfully identified hidden
patterns in the STATS 19 data containing adverse environmental conditions (e.g., snow,
fog et al.). The NSGA-II algorithm has successfully identified injury critical vehicle
dynamics parameters with great coverage in the scenario design space. Moreover, by
automating the identification process of injury critical scenarios in a modular and timeefficient way, this work has the potential to: a) accelerate verification and validation
activities and b) improve the safety of ADSs in mixed-vehicle fleet intersections.
Date of Award2023
Original languageEnglish
Awarding Institution
  • Coventry University
SupervisorChristophe Bastien (Supervisor), Jesper Christensen (Supervisor), Karthikeyan Ekambaram (Supervisor), Stratis Kanarachos (Supervisor) & Dobrila Petrovic (Supervisor)

Cite this