Safety assessment of automated vehicles using real-world driving scenarios

Automated Vehicles (AVs) have a great potential to change transport fundamentally by making it safer, by reducing travel time, and by increasing mobility and accessibility for all. The level of automation of these vehicles determines the extent to which the driver’s task is accomplished by the AV. With the increasing number of AVs entering the market, the level of automation of these vehicles is increasing. The increasing level of automation will cause a paradigm shift: traditionally, human drivers are responsible for the behavior of the vehicle, even if the vehicle is momentarily controlled by an Automated Driving System (ADS), but with increasing levels of automation, the human driver will no longer be solely responsible. So, the accountability and liability shift from the driver to the vehicle manufacturer, the operator of the vehicle (fleet), and/or the (vehicle) authorities. Due to this paradigm shift, for higher levels of automation, it can no longer be assumed that the human driver intervenes whenever the ADS does not respond appropriately. To guarantee that these ADSs respond appropriately in nearly all situations, new methods for assessing ADSs are required.

Scenario-based assessment is an approach for assessing AVs that is broadly supported by the automotive field. With a scenario-based assessment, the AV under test is subjected to many different test scenarios. These test scenarios resemble situations that may be encountered in real-world traffic, to see whether the AV responds appropriately to these scenarios. One of the main challenges with scenario-based assessment of an AV with a high level of automation is to come up with a set of test scenarios that provides enough confidence that the AV responds appropriately in nearly all situations. One popular approach is to use real-world data that contain scenarios from real-world traffic as a source to automatically generate test scenarios. This dissertation describes new methods for improving this data-driven scenario-based assessment of AVs.

The first contribution of this dissertation is a comprehensive and operable definition of the term scenario in the context of scenario-based assessment of AVs. We define a scenario as a quantitative description of the relevant characteristics and activities and/or goals of the ego vehicle(s), the static environment, the dynamic environment, and all events that are relevant to the ego vehicle(s) within the time interval between the first and the last relevant event. A scenario category is defined as the qualitative counterpart of a scenario and can be regarded as an abstraction of a scenario. To enable a computer to store, communicate, interact with, and interpret scenarios, an Object-Oriented Framework (OOF) is proposed in which scenarios, scenario categories, and their building blocks are defined as classes of objects having attributes, methods, and relationships. The advantage of the OOF is that it promotes clarity, modularity, and reusability of the objects that constitute a scenario.

The second contribution is a novel metric for quantifying the degree of completeness of the collected data that are used for the data-driven scenario-based assessment of AVs. The data are used to estimate unknown probability density functions (pdfs) of the important parameters that are used to describe scenarios. The proposed completeness metric is based on the expected approximation error, which is the discrepancy between the real pdf and the estimated pdf: a lower approximation error indicates a higher degree of completeness.

The third contribution is a novel method for capturing scenarios of a specific scenario category from a data set. For example, the provided method can capture all cut-in scenarios from a data set. One of the benefits of the method is that characteristics of a scenario that are shared among different scenario categories need to be identified only once. As a result, the provided method is easily applied to a wide range of scenario categories, such that a wide variety of scenarios can be obtained from the data.

The fourth contribution is the proposal of two complementary methods for generating test scenarios for AVs. The first method automatically determines the parameters that best describe the scenarios of a specific scenario category. The underlying, unknown pdf of the parameters is estimated and scenarios are generated by sampling parameter values from the estimated pdf. The second method enables the conditional sampling of parameter values, which can be used to, e.g., generate scenarios with predefined starting conditions. The benefits of the presented methods are that the generated scenarios are representative of real-world scenarios, they cover the actual variety found in real-world traffic, and they extend the variety found in the collected data. To measure the extent to which the generated scenarios indeed represent real-world scenarios while covering the actual variety found in real-world traffic, the novel Scenario Representativeness metric is proposed.

The fifth contribution is the proposal of two novel methods for quantifying the risk of an AV. Both methods calculate the risk by combining the outcome of virtual simulations of scenarios generated using the aforementioned methods and the estimated likelihood of these scenarios. The first method quantifies the risk prospectively, i.e., before the actual deployment of the AV on public roads. The quantified risk supports the risk assessment activities of ISO 26262 and ISO 21448, the leading standards in automotive safety. These standards decompose the risk into three aspects: exposure, severity, and controllability. Whereas safety experts’ opinions are traditionally used to provide qualitative, subjective ratings for each of these three aspects, our proposed method computes these aspects in a data-driven, quantitative manner. The second method is the novel data-driven Probabilistic RISk Measure derivAtion (PRISMA) method, which is used to derive Surrogate Safety Measures (SSMs) that estimate the probability of a specific event (e.g., a crash) in real time. As opposed to existing SSMs, which are only applicable in specific types of scenarios, the PRISMA method can be used to derive multiple SSMs for different types of scenarios.

The work presented in this dissertation thus makes a substantial contribution to the full integration of a scenario-based assessment for the type approval of AVs. This, in turn, brings us closer to the large-scale deployment of AVs on public roads.