Dynamic Fusion and Domain-Invariant Feature Learning for Object Detection in Complex Weather Environments
Surjeet Singh Parihar, Surbhit Shukla, Shashank Swami · 2026
Adverse weather conditions such as fog, rain, snow, haze, and low light conditions have a devastating effect on the performance of the Real World Outdoor Object detection. These conditions cause the loss in visibility, modality imbalance, and domain shift, making the traditional RGB-based object detection methods lack credibility. The above-mentioned drawbacks of the traditional object detection methods make the traditional non-dynamic modality methods inappropriate for Critical Safety Applications (CSA), where a steady perception is required. To overcome the above-mentioned problem, it is recommended that a unified multimodal framework will be used, which will incorporate adaptive fusion and features learning for weather invariance. It uses two encoders that operate independently for RGB and infrared data, which maintains complementary semantic and thermal data. Dynamic Fusion Module is used to dynamically combine the modality contributions based on the scene conditions so that it may be seen under different conditions of visibility with confidence. A Domain-Invariant Feature Learning mechanism is used to optimize the class distributions with respect to weather domains by categorical consistency regularization, and Confounder Dictionary Module is used to remove structured atmospheric artefacts. The framework divides the object semantics and environmental noise by modeling the weather distortions as causal confounders, which enhances generalization to visualized conditions. Much towards FLIR, BDD100K weather subsets, Foggy Cityscapes and a newly assembled RGB-IR dataset: massive performance improvements, with up to 18% mAP and reduced false positives. The findings attest to the claim that adaptive multimodal fusion and confounder-conscious learning are a strong scalable methodology to obtain reliable object detection in adverse real-world scenarios.