SiVIC-ADVeRSce: A Framework for Generating Synthetic Datasets for Visual Perception in Adverse Scenarios and Automated Driving
Wei Xu, Dominique Gruyer, Alexandra Duminil, Sio-Song Ieng · 2024
Autonomous driving systems rely on accurate perception, especially in adverse conditions like weather and light disturbances. However, evaluation in natural adverse conditions poses significant challenges. Synthetic datasets offer a solution by providing a safe, cost-effective, scalable, and reproducible way of assessing autonomous driving systems in adverse scenarios. To address this need, this paper introduces SiVIC-ADVeRSce (SiVIC-Automated Driving system Validation in simulated adverse Road Scenario), a novel framework designed for generating synthetic datasets specifically tailored for training, evaluation, and validation of visual perception functions of Automated Vehicle (AV) in adverse scenarios. Building upon the French national PRISSMA project, this paper presents a conceptual framework, defines critical components for effective dataset generation, and accompanies the corresponding implementation basis. The effectiveness of SiVIC-ADVeRSce is demonstrated through its application in PRISSMA Proof of Concept (PoC), dedicated to Bus Station Automated Services (BuSAS) using an AI-based perception system. Furthermore, this paper proposes potential extensibility in the post-processing phase using AI-based methods like Generative Adversarial Networks (GANs) to create variations of generated virtual datasets.