A Static-to-Temporal Framework for Interpretable Camera Lens Soiling Severity Estimation in Autonomous Driving.

Camera lens soiling can severely degrade visual perception in autonomous driving, making reliable soiling severity estimation essential for camera-health monitoring and downstream perception safety. However, existing methods mainly rely on area-based indicators or frame-wise predictions, which insufficiently account for opacity differences, spatial importance, and temporal stability in continuous video streams. To address these limitations, this paper proposes a static-to-temporal soiling framework for camera-soiling severity estimation. First, we propose a Structured Dual-Head Static Model (SDSM) that jointly predicts tile-level four-class soiling distributions and an image-level severity score. The model is coupled with an explicit Structured Severity Score that aggregates local predictions through opacity-aware, spatial, and dominance-related components. Second, to alleviate the scarcity of real temporal soiling data, we construct a Two-Stage Stable Diffusion (TS-SD) pipeline and use the resulting SD-Seq data as mechanism supervision for temporal learning rather than direct single-frame strong supervision. Finally, we introduce a structure-constrained adaptive EMA Module to improve temporal stability while preserving the original single-frame severity scale. Experiments on WoodScape, External Test, and OccNuScenes-Dirt show strong cross-domain severity estimation performance, including a cluster-level Spearman correlation of 0.7876 on External Test. The temporal module further reduces Jitter (MAD) by 51.5%. These results demonstrate an interpretable, cross-domain, and deployment-friendly solution for camera-soiling assessment.