Combining knowledge and information - graph-based description of driving scenarios to enable holistic vehicle safety

Abstract

Currently, vehicle safety is based on knowledge from injury values, crash pulses, and driving kinematics which leads to intervention strategies separated into isolated domains of active and passive safety. In this contribution, it is shown how vehicle safety can be approached holistically, allowing for human-centered and scenario-based safety decision-making. For this purpose, information from interior and exterior vehicle sensors can be linked by a mathematical framework, combining the knowledge that is already available in the individual domains. A universal graph representation for driving scenarios is developed to master the complexity of driving scenarios and allow for an optimized and scenario-based intervention strategy to minimize occupant injury values. This novel approach allows for the inclusion of sub-models, expert knowledge, results from previous simulations, and annotated databases. The resulting graph can be expanded dynamically for other objects or occupants to reflect all available information to be considered in case of urgency. As input, interior and exterior vehicle sensor data is used. Further information about the driving situation is subsequently derived from this input and the interaction between those states is described by the graph dynamically. For example, occupant attentiveness is derived from measurable eye gaze and eyelid position. From this quantity, reaction time can be estimated in turn. Combined with exterior information, it is possible to decide on the intervention strategy like e.g. alerting the driver. Physical or data-based functional dependencies can be used to represent such interactions. The uncertainties of the inputs and from the surrogate models are included in the graph to ensure a reliable decision-making process. An example of the decision-making process, by modeling the states and actuators as partially observable Markov decision process (POMDP), shows how to optimize the airbag efficiency by influencing the head position prior to an impact. This approach can be extended by additional parameters like driving environment, occupant occupancy, and seating positions in further iterations to optimize the intervention strategy for occupants. The proposed framework integrates scenario-based driving dynamics and existing knowledge from so far separated safety systems with individual activation logic and trigger points to enable holistic vehicle safety intervention strategies for the first time. It lays the foundation to consider new safety hardware, sensor information, and safety functions through a modular, and holistic approach.