Deep learning-based robust positioning for all-weather autonomous driving

Autonomous vehicles (AVs) require precise localization for safe operation, which is challenging under adverse weather. This paper proposes a deep learning-based self-supervised approach for robust ego-motion estimation, fusing visual and radar data using an attention mechanism. The method predicts reliability masks to mitigate sensor deficiencies, demonstrating robust all-weather performance and cross-domain generalizability. A game-theoretic approach analyzes model interpretability, revealing independent failure modes. This work advances AVs towards safe and reliable all-weather autonomous driving.

Yasin Almalioglu, Mehmet Turan, Niki Trigoni, Andrew Markham