Autonomous Driving in Adverse Weather: A Multi-Modal Fusion Framework with Uncertainty-Aware Learning for Robust Obstacle Detection

Robust obstacle detection in autonomous driving under adverse weather remains a critical challenge due to sensor degradation, visibility reduction, and increased uncertainty. This study proposes an Uncertainty-Aware Multi-Modal Fusion (UAMF) framework that integrates LiDAR, RGB images, and weather priors through a dynamic cross-modal attention mechanism and Bayesian uncertainty modeling. The model adaptively adjusts the fusion weights between sensor modalities according to real-time weather conditions and jointly optimizes detection loss with a KL divergence regularization to quantify predictive uncertainty. Experimental results on the nuScenes, KITTI-Adverse, and CARLA datasets demonstrate that UAMF achieves superior performance across rain, snow, and fog scenarios, with mAP@0.5 reaching 0.78, 0.72, and 0.65, respectively—representing 12–31% gains over existing baselines. Notably, UAMF reduces false positive rates by up to 40% in low-visibility conditions and exhibits a strong correlation (ρ = 0.85) between estimated uncertainty and localization error. Ablation studies confirm the importance of the weather-aware fusion and uncertainty modules, while visibility-level analysis shows improved robustness under <30 m scenarios. The proposed framework offers reliable uncertainty signals for downstream decision-making and is deployable in real-time on embedded platforms. Future work will explore unsupervised weather parameter estimation, uncertainty-aware trajectory forecasting, and cross-domain generalization.

@article{Li2025,
  title     = {Autonomous Driving in Adverse Weather: A Multi-Modal Fusion Framework with Uncertainty-Aware Learning for Robust Obstacle Detection},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {8},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Zhengqing Li and Baljit Singh Bhathal Singh},
  doi       = {10.14569/IJACSA.2025.0160808},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160808}
}