Understanding and modelling car drivers overtaking cyclists: Toward the inclusion of driver models in virtual safety assessment of advanced driving assistance systems
Licentiate thesis, 2019
The total number of road crashes in Europe is decreasing, but the number of crashes involving cyclists is not decreasing at the same rate. To help car drivers avoid or mitigate crashes while overtaking a cyclist, advanced driver assistance systems (ADAS) have been developed. To evaluate and further improve these ADAS to support drivers as they overtake cyclists, we need to understand and model driver behaviours.
This thesis has two objectives: 1) to extract and analyse cyclist-overtaking manoeuvres from naturalistic driving data and 2) compare driver behaviour models for overtaking manoeuvres that can be used in counterfactual simulations for evaluating ADAS safety benefits.
The drivers’ comfort zone boundaries (CZBs) when overtaking a cyclist were identified and analysed using naturalistic driving data. Three driver models that predict when a car driver starts steering away in order to overtake a cyclist were implemented: a threshold model, an evidence accumulation model, and a model inspired by a proportional-integral-derivative controller. These models were tested and verified using two different datasets, one from a test-track experiment and one from naturalistic driving data. Model parameters were obtained using computationally efficient linear programming.
The results show that, when an oncoming vehicle was present, the drivers were significantly closer to the cyclist before steering away. This finding indicates that the presence of an oncoming vehicle is a crucial factor for the safety of the cyclist and needs to be taken into account for the development of ADAS that maintain safe distance to the cyclist. Furthermore, the quantification of the CZBs has implications for the development of ADAS which can estimate the time-to-collision to an oncoming vehicle or a cyclist to be overtaken, providing timely and acceptable warnings—or interventions—when drivers exceed their usual CZBs. A comparison of the models shows that all three are highly variable in detecting steering away time for different drivers. Furthermore, differences were discovered in detected steering away time between models fitted to test-track experiments and naturalistic driving data. Future work may focus on using larger, more diverse datasets and investigating more advanced models before including them in counterfactual simulations.