Towards More Reliable Pre-Crash Virtual Safety Assessment: The impact of the choice of data types and reference driver models on the assessment of vehicle automation

Licentiate thesis, 2024

Road crashes are a major cause of deaths and serious injuries worldwide. New technologies offer the opportunity to reduce road crashes 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). Methods are in place to assess how safe these systems are. One of these methods employs virtual simulations to predict the impact on safety that the systems would have once released on public roads. However, the process for ensuring that a virtual simulation provides an effective, relevant, and fair assessment of ADASs and ADSs is not always straightforward. This thesis contributes to the development of virtual safety assessment methods by investigating the impact of different data and models on the resulting simulations.

Specifically, the first objective of the thesis is to measure the impact of data selection on the outcomes of virtual safety assessment. Crashes were artificially generated from near-crashes and everyday driving data, using a model of an unresponsive driver. The generated crashes were compared to real-world reconstructed crashes. Automated emergency braking (AEB) systems were then applied to the crashes, to study the impact different data sources have on crash avoidance and mitigation. The results show that those artificially generated crashes are very different from real-world crashes, with lower severity outcomes and criticality.

The second objective of this thesis is to understand if existing reference driver models represent a competent and careful human driver. These models are intended to be benchmarks for ADS safety performance. The models studied in this thesis—from the UN Regulation No. 157—did not perform as the competent and careful drivers they are intended to represent when applied on near-crash cut-ins through counterfactual simulations. Specifically, one model generally showed delayed responses to critical scenarios, compared to humans. The other model instead showed non-human-like behavior, reacting substantially earlier than humans.

The impact of the findings is twofold. First, they can help the development of virtual safety assessment methods by discouraging the use of everyday driving data and near-crash data in counterfactual crash generation. Second, the findings on reference driver models make it clear that models used in regulations must be validated using a range of data types. To continue the work on reference driving models, future work aims at studying how urgency in traffic scenarios impacts drivers’ behaviors. The concept of comfort zone boundaries (CZBs) will be used to study the limits that drivers are able and willing to tolerate in routine driving, and the inclusion of CZBs in the models will be investigated. This research has the potential to contribute to the improvement of reference driver models and virtual safety assessment methods.