Road ponding presents a substantial threat to vehicular safety, particularly in foggy conditions where reliable detection continues to be a major challenge for advanced driver assistance systems (ADASs). To address this issue, we propose an aggregation-broadcast-coupling dynamic wavelet network (ABCDWaveNet), a novel deep learning framework specifically designed to achieve robust ponding detection in fog-affected environments. The central architecture of ABCDWaveNet improves detection performance by utilizing dynamic convolution for adaptive feature extraction under reduced visibility, together with a wavelet-based module that improves feature representation across both spatial and frequency domains, thereby effectively alleviating fog-related interference. In addition, ABCDWaveNet incorporates multi-scale structural and contextual information and employs an adaptive attention coupling gate to dynamically integrate global and local features, leading to improved detection accuracy. For realistic evaluations under compounded adverse weather conditions, we introduce the Foggy Low-Light Puddle dataset. Comprehensive experiments confirmed that ABCDWaveNet attained state-of-the-art results, with notable intersection over union gains of 3.51%, 1.75%, and 1.03% on the Foggy-Puddle, Puddle-1000, and Foggy Low-Light Puddle datasets, respectively. Furthermore, with an inference speed (FPS) of 25.48 on the NVIDIA Jetson AGX Orin, the proposed framework demonstrates strong suitability for development in ADAS applications. These results highlight the effectiveness of ABCDWaveNet, presenting valuable advancements for proactive road safety under challenging weather conditions.
Zhang et al. (Sun,) studied this question.