Developing a Methodology to Assess Data Completeness of Driving Scenarios for Testing Autonomous Vehicles: A Focus on ODD-Specific Objectives

@inproceedings{HadjSalem2024,
title = {Developing a Methodology to Assess Data Completeness of Driving Scenarios for Testing Autonomous Vehicles: A Focus on ODD-Specific Objectives},
author = {Fouad Hadj Selem and Ghada Ben Nejma and Walid Kheriji and Laurent Durville and Mohamed Cherif Rahal and Stephane Geronimi and Emmanuel Arnoux},
editor = {Andras Kemeny and Jean-Rémy Chardonnet and Florent Colombet and Stéphane Espié},
isbn = {978-2-9573777-5-6},
year  = {2024},
date = {2024-09-18},
urldate = {2024-09-18},
booktitle = {Proceedings of the Driving Simulation Conference 2024 Europe VR},
pages = {145-151},
address = {Strasbourg, France},
organization = {Driving Simulation Association},
abstract = {In the scenario-based safety assessment of automated vehicles, the increasing volume of field data prompts critical questions about dataset adequacy: ”Is our dataset sufficient?” and ”What additional insights can we derive from expanding our data collection efforts?”. Additionally, if our data proves insufficient, what specific elements are lacking, and how can they be quickly identified in a targeted manner? This study presents a novel approach to address these inquiries. Building upon existing methodologies, traffic data is reinterpreted as scenarios, each encapsulating various driving activities. Key parameters characterizing these activities are identified to trans form the data representation into a high dimensional continuous space with an unknown probability distribution. Unlike traditional methods, our approach draws inspiration from deep generative techniques, allowing for the trans formation of this classical representation into a new latent space where the distribution becomes known facilitating the evaluation of dataset completeness. Applied to real-world driving scenarios in Europe, our methodological framework estimates scenario coverage rates consistently aligned with key performance indicators, demonstrating its effectiveness in assessing dataset completeness for scenario-based safety assessment. This approach is tailored to estimate the completeness rate associated solely with a fixed and limited statistical objective. Although it can be defined in multiple ways, very flexibly, it must be set at the beginning of the analysis. This is a limitation, but it can also be seen as an advantage, as it allows for planning with a limited budget and specific goals. Furthermore, this approach can aggregate multiple objectives simultaneously, moving away from the pursuit of a vague goal or the notion of absolute completeness for any given problem or application. It transcends limitations by not constraining number of variables, eliminating independence assumptions, and directly computing coverage rates, providing a rigorous framework for conducting safety assessments and enabling a comprehensive evaluation of ADAS. These findings are crucial for the continuity of scenario-based verification efforts, as this approach continues to emerge as a promising solution in the field of automated vehicle safety assessment.},
keywords = {},
}

Download .bib file

TY  - CONF
TI  - Developing a Methodology to Assess Data Completeness of Driving Scenarios for Testing Autonomous Vehicles: A Focus on ODD-Specific Objectives
AU  - Selem, Fouad Hadj
AU  - Nejma, Ghada Ben
AU  - Kheriji, Walid
AU  - Durville, Laurent
AU  - Rahal, Mohamed Cherif
AU  - Geronimi, Stephane
AU  - Arnoux, Emmanuel
C1  - Strasbourg, France
C3  - Proceedings of the Driving Simulation Conference 2024 Europe VR
DA  - 2024/09/18
PY  - 2024
SP  - 145
EP  - 151
LA  - en-US
PB  - Driving Simulation Association
SN  - 978-2-9573777-5-6
L2  - https://proceedings.driving-simulation.org/proceeding/dsc-2024/developing-a-methodology-to-assess-data-completeness-of-driving-scenarios-for-testing-autonomous-vehicles-a-focus-on-odd-specific-objectives
ER  - 

Download .ris file