BiLSTM-Based VAE-GAN for Predicting Future Road States in Autonomous Driving

초록

The ability to accurately predict future road conditions is essential for the advancement of autonomous driving systems. This study introduces a BiLSTM-based VAE-GAN framework that leverages both temporal and spatial information to generate high-quality future road images. The proposed architecture combines the reconstruction capabilities of Variational Auto-Encoders (VAEs) with the adversarial training of Generative Adversarial Networks (GANs), while incorporating Bidirectional Long Short-Term Memory (BiLSTM) to effectively capture temporal dependencies in sequential driving data. To train the model, diverse datasets were collected from the CARLA simulation environment, encompassing various road conditions and vehicle states. The training process minimizes reconstruction loss, KL divergence, and adversarial loss, enabling the generation of visually consistent and semantically accurate future road images. Quantitative evaluations using PSNR and MSE metrics demonstrate the model's ability to outperform conventional VAE-based approaches, achieving high structural similarity and low reconstruction errors. The results highlight the potential of the proposed framework to enhance decision-making and lane-keeping performance in autonomous vehicles. By predicting future road states with high fidelity, the BiLSTM-based VAE-GAN framework lays the groundwork for integrating generative models into real-world autonomous driving applications, contributing to safer and more reliable driving systems.

키워드