Abstract Scene Graphs: Formalizing and Monitoring Spatial Properties of Automated Driving Functions
Ishan Saxena, Bernd Westphal, Martin Fränzle
TL;DR
The paper addresses the challenge of validating safety-critical spatial properties of Automated Driving Functions (ADFs) on diverse road scenarios. It introduces Abstract Scene Graphs ($ASG$) to extend Scene Graphs ($SGs$) for formalizing spatial properties, with $ASG = (G_{ ext{A}}, D)$ where $G_{ ext{A}}$ is a directed heterogeneous graph and $D$ a predicate set. A Runtime Monitoring (RM) framework is presented to check ADF behavior against ASG-specified properties, illustrated using real-world NHTSA pre-crash examples. The approach aims to reduce semantic ambiguity and remove reliance on text-based graph queries, enabling clearer, multi-stakeholder specification and verification, with planned extensions to automotive ontology, hardware-efficient RM, and cross-domain applicability.
Abstract
Automated Driving Functions (ADFs) need to comply with spatial properties of varied complexity while driving on public roads. Since such situations are safety-critical in nature, it is necessary to continuously check ADFs for compliance with their spatial properties. Due to their complexity, such spatial properties need to be formalized to enable their automated checking. Scene Graphs (SGs) allow for an explicit structured representation of objects present in a traffic scene and their spatial relationships to each other. In this paper, we build upon the SG construct and propose the Abstract Scene Graph (ASG) formalism to formalize spatial properties of ADFs. We show using real-world examples how spatial properties can be formalized using ASGs. Finally, we present a framework that uses ASGs to perform Runtime Monitoring of ADFs. To this end, we also show algorithmically how a spatial property formalized as an ASG can be satisfied by ADF system behaviour.