An Ontology-based Method to Identify Triggering Conditions for Perception Insufficiency of Autonomous Vehicles

TL;DR

This paper introduces an ontology-based framework to systematically identify and analyze triggering conditions that could cause perception insufficiency in autonomous vehicles, enhancing safety validation processes.

Contribution

It presents a novel ontology-driven method for comprehensive identification of AV perception triggering conditions, addressing a gap in systematic safety analysis.

Findings

Identified 87 potential triggering conditions for an L3 autonomous vehicle.

Tested 20 conditions in real-world scenarios, with 8 causing risky behaviors.

Provides a systematic approach to improve AV safety validation.

Abstract

The autonomous vehicle (AV) is a safety-critical system relying on complex sensors and algorithms. The AV may confront risk conditions if these sensors and algorithms misunderstand the environment and situation, even though all components are fault-free. The ISO 21448 defined the safety of the intended functionality (SOTIF), aiming to enhance the AV's safety by specifying AV's development and validation process. As required in the ISO 21448, the triggering conditions, which may lead to the vehicle's functional insufficiencies, should be analyzed and verified. However, there is not yet a method to realize a comprehensive and systematic identification of triggering conditions so far. This paper proposed an analysis framework of triggering conditions for the perception system based on the propagation chain of events model, which consists of triggering source, influenced perception stage, and triggering effect. According to the analysis framework, ontologies of triggering source and perception stage were constructed, and the relationships between concepts in ontologies are defined. According to these ontologies, triggering conditions can be generated comprehensively and systematically. The proposed method was applied on an L3 autonomous vehicle, and 20 from 87 triggering conditions identified were tested in the field, among which eight triggering conditions triggered risky behaviors of the vehicle.