Preprint: Reliability-Aware Multimodal Fusion for Robust Semantic Segmentation in Autonomous Driving

Abstract

Robust semantic segmentation is essential for autonomous driving, yet real-world perception is strongly affected by adverse weather, sensor degradation, and partial sensor failures. While multimodal fusion is widely used to improve robustness, many approaches treat all modalities equally and do not explicitly account for their varying reliability. In this work, we present ReCoLaF, a reliability-aware conflict-guided late fusion framework for multimodal semantic segmentation under uncertainty. ReCoLaF combines evidential deep learning with Dempster--Shafer theory and introduces a two-stage trust modeling strategy: a class-wise reliability mechanism that estimates the semantic relevance of each modality, followed by a conflict-guided adjustment that reduces the contribution of inconsistent sources at the pixel level. Experiments on the DeLiVER and MUSES datasets show that ReCoLaF improves robustness under degraded sensing conditions and sensor failures, achieving the best average performance across modality-availability scenarios. These results highlight the interest of explicitly modeling both reliability and conflict for trustworthy multimodal perception in autonomous driving.