Safety Evaluation Using Counterfactual Simulations: The use of computational driver behavior models in crash avoidance systems and virtual simulations with optimal subsampling

Licentiate thesis, 2023

Traffic safety is a problem worldwide. In-vehicle conflict and crash avoidance systems have been under development and assessment for some time, as integral parts of Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS). Among the methods used to assess conflict and crash avoidance systems developed by the automotive industry, virtual safety assessment methods have been shown to have great potential and efficiency. In fact, scenario generation-based virtual safety assessments play—and are likely to continue to play—a very important role in the assessments of vehicles of all levels of automation. The ultimate aim of this thesis is to improve the safety performance of conflict and crash avoidance systems. This aim is addressed through the use of computational driver models in two different ways. First, by using comfort-zone boundaries in system design, and second, by using a behavior-based crash-causation model together with a novel optimized scenario generation method for virtual safety assessment.

The first objective of this thesis is to investigate how a driver model which includes road users’ comfortable behaviors in crash avoidance algorithms impacts the systems’ safety performance and the residual crash characteristics. Chinese car-to-two-wheeler crashes were targeted; Automated Emergency Braking (AEB) algorithms, which comprised the proposed crash avoidance systems, were compared to a traditional AEB algorithm. The proposed algorithms showed larger safety performance benefits. In addition, the similarities in residual crash characteristics regarding impact speed and location after different AEB implementations can potentially simplify the designs of in-crash protection system in future.

The second objective is to develop and apply a method for efficient subsampling in crash-causation-model-based scenario generation for virtual safety assessment. The method, which is machine-learning-assisted, actively and iteratively updates the sampling probability based on new simulation results. The crash-causation model is based on off-road glances and a distribution of driver maximum decelerations in critical situations. A simple time-to-collision-based AEB algorithm was used to demonstrate the assessment process as well as the benefits of combining crash-causation-model-based scenario generation and optimal subsampling. The sampling methods are designed to target specific safety benefit indicators, such as impact speed reduction and crash avoidance rate. The results of the study show that the proposed sampling method requires almost 50% fewer simulations than traditional importance sampling.

Future work aims to focus on applying the active sampling method to driver-model-based car-to-vulnerable road user (VRU) scenario generation. In addition to assessing conflict and crash avoidance system performance, a novel stopping criterion based on Bayesian future prediction will be further developed and demonstrated for use in experiments (e.g., as part of developing driver models) and virtual simulations (e.g., using driver-behavior-based crash-causation models). This criterion will be able to indicate when studies are unlikely to yield actionable results within the budget available, facilitating the decision to discontinue them while they are being run.