자율주행 차량을 위한 Latent Diffusion Model 기반의 주행 시나리오 생성

Abstract

This study presents a novel deep learning-based method for scenario generation to validate autonomous driving systems (ADSs). Testing ADSs for every possible scenario in the real world is challenging and costly due to the vast number of potential driving conditions, which can slow down the research and development process. To address this issue, a simulation-based validation method using driving scenarios has been introduced. Current scenarios typically rely on the recorded traffic data or the rule-based traffic agents. However, these conventional method-based scenarios often lack sufficient realism and diversity, which hinders effective validation. This study proposes an approach that leverages a deep learning method for scenario generation to enhance realism and diversity. First, a variational autoencoder (VAE) was employed to encode complex traffic characteristics into a compact latent space. The latent diffusion model was then used to add controlled noise into the latent space and learn this space through iterative denoising processes. Consequently, a new latent space was generated from the learned denoising process using Gaussian noise as the input. The new latent space provided diverse traffic characteristics, which the VAE decode into the driving scenarios that maintain realism while exhibiting diverse traffic. The generated scenarios were validated by comparing them with the scenarios made by VAE, while focusing on their realism and diversity.