SADA: Unsupervised Domain Adaptation for Reliable Scene Awareness

Although deep neural networks are extensively employed in autonomous driving, they must be adjusted to handle unforeseen environmental conditions that were not included in their initial training. Unsupervised domain adaptation (UDA) techniques have proven effective in transferring learned knowledge from one domain to another when the labels in the target domain are unavailable. However, their potential application in scene awareness remains underexplored. This paper introduces SADA (Scene Awareness with Unsupervised Domain Adaptation), a novel approach designed to enhance the reliability of drivable area and lane detection tasks in autonomous driving systems. We leverage UDA techniques in unified multi-task networks to address domain shift challenges, ensuring robust feature extraction under diverse and complex conditions. The proposed technique is evaluated on different real-world datasets and the results demonstrate that the proposed methods outperform the baseline scheme in terms of detection accuracy and consistency.