🤖 AI Summary
To address the challenge of efficiently identifying and deeply understanding safety-critical corner cases from massive naturalistic driving data in autonomous driving development, this paper proposes a dual-path analysis framework integrating semantic scene graph modeling and virtual reality (VR)-based immersive reconstruction. Methodologically, it introduces a novel closed-loop paradigm that synergizes objective detection—driven by multiple criticality metrics (e.g., TTC, RSS, SFF)—with VR-enabled interactive, multi-perspective subjective validation. Semantic scene graphs explicitly encode relational structures among traffic entities, while high-fidelity VR reconstruction and interpretable visualizations jointly support precise scenario localization and human-factor analysis. Experiments on real-world datasets demonstrate significant improvements in corner case detection rate and localization accuracy, alongside enhanced interpretability of causal factors and deeper human-factor insights. This work establishes a new paradigm for systematic verification of autonomous driving systems.
📝 Abstract
One core challenge in the development of automated vehicles is their capability to deal with a multitude of complex trafficscenarios with many, hard to predict traffic participants. As part of the iterative development process, it is necessary to detect criticalscenarios and generate knowledge from them to improve the highly automated driving (HAD) function. In order to tackle this challenge,numerous datasets have been released in the past years, which act as the basis for the development and testing of such algorithms.Nevertheless, the remaining challenges are to find relevant scenes, such as safety-critical corner cases, in these datasets and tounderstand them completely.Therefore, this paper presents a methodology to process and analyze naturalistic motion datasets in two ways: On the one hand, ourapproach maps scenes of the datasets to a generic semantic scene graph which allows for a high-level and objective analysis. Here,arbitrary criticality measures, e.g. TTC, RSS or SFF, can be set to automatically detect critical scenarios between traffic participants.On the other hand, the scenarios are recreated in a realistic virtual reality (VR) environment, which allows for a subjective close-upanalysis from multiple, interactive perspectives.
