Scenario Classes in Naturalistic Driving: Autoencoder-based Spatial and Time-Sequential Clustering of Surrounding Object Trajectories

Surrounding vehicles are among the essential features to describe traffic scenarios. Besides maneuver (e.g., turn) and scene (e.g., highway), these features are hard to capture in words or labels. The recognition and evaluation of these scenario features are important for road safety. Consequently, when analyzing naturalistic driving data, the composition of the scenarios is essential in order to be able to evaluate driver behavior, and the effects of the overall system quantitatively. In this work, we propose a method to group surrounding vehicles from the perspective of the ego-vehicle and use it for an improved scenario classification. In a two-step approach, we group each vehicle within a scenario independently. We separate the spatial domain (driving tube) from the time domain (performance style). The spatial domain is clustered using a hierarchical ward algorithm to allow for variation of the cluster depth. With the merged result, we realize an outlier detection and a method to quantify the frequency of trajectories within scenarios. From this, the uniqueness of scenarios, e.g., for resimulation, is quantified. This enables us to identify clusters of similar maneuvers of surrounding vehicles up to, for example, lane change maneuver groups of the same speed and acceleration course.