Towards a characterization of safe driving behavior for automated vehicles based on models of "typical" human driving behavior

Abstract

Automated driving is expected to play a central role in future mobility systems by enabling, among other benefits, mobility-as-a-service schemes and better road utilization. To this end, automated vehicles must not only be functionally safe. They should also be perceived as driving safely by other traffic participants and have a positive impact on traffic safety. However, to the best of our knowledge, there is no consensus yet on what "driving safely'' means. This article proposes a new characterization of safe driving behavior for automated vehicles based on models of "typical" human driving behavior. Such behavior (specially from attentive, experienced drivers) is known to lead to interactions of mid to low severity (i.e., low collision risk). Automated vehicles displaying similar behavior would interact with other traffic participants in a recognizable, predictable, and safe way. As a first step towards this characterization, machine-learning-based models (autoencoders) were developed from longitudinal, naturalistic driving data (from NGSIM). Autoencoders are relatively inexpensive computationally and can monitor whether a vehicle behaves "typically'" or not based on anomaly detection principles. Our initial results show that the proposed approach can readily separate typical (safe) from anomalous (unsafe) driving behavior in the considered data set.