Modeling of high-level driving decisions from naturalistic driving data for multi-agent traffic simulation

Traditional hierarchical driving stacks for simulated traffic agents have inherent shortcomings in generating trajectories that are both feasible and naturalistic on high and low abstraction levels.

In his upcoming presentation at the "Machine Learning for Virtual Vehicle Development (CAE & VR)" event on the 25th of February 2021, Christian from AAI will discuss the concept, implementation, and the results of an integrated multi-agent simulation platform to support the development and validation of advanced driver assistance systems (ADAS) and automated driving (AD) functions.

Our simulation facilitates the validation of vehicle-to-vehicle interactions in various traffic scenarios leading to greater user adoption of such systems either as an external observer or as part of the traffic environment (e.g. as driver-in-the-loop). We propose a new driving stack model for simulated traffic agents that is able to mimic high-level decisions from naturalistic traffic data and can employ any trajectory planning method that is able to propose a variety of trajectories. The agent model focuses on human-like behavior, decision-making, and driving performance based on a supervised machine learning approach. Througout our presentation, we will show that even though our approach is based on modelling nanoscopic traffic-agent characteristics, realistic traffic behavior can be achieved both at the microscopic and macroscopic level. The results will be demonstrated based on an exemplary dataset from German highway traffic. Finally, important challenges around the evaluation of traffic simulation performance evaluation will be addressed, concluded with a brief outlook on our future developments.

Follow the link and join the "Machine Learning for Virtual Vehicle Development (CAE & VR)" event on 25 February 2021 to learn more about the event and related contributions around the topic: "How Automotive Simulation Engineers can benefit from Machine Learning".