Self-Attention-based Multi-Scale Feature Fusion Network for Road Ponding Segmentation

Road ponding, an omnipresent anomaly in traffic scenarios, poses a substantial hazard to road safety, manifesting as potential vehicular loss of control, minor abrasions, or severe collisions. Its transparency renders road ponding susceptible to environmental influences, presenting a formidable challenge for precise edge detection. Existing methodologies exhibit shortcomings in accurately discerning road ponding. In response, we propose an innovative approach called Fusformer (Self-Attention-based Multi-Scale Feature Fusion Network). Representing the pioneering application of Transformer architecture in road ponding detection, Fusformer integrates an enhanced self-attention mechanism derived from Segformer. This mechanism adeptly captures highly non-local contextual attention, thereby facilitating improved identification of ponding edges. Augmenting this approach, we incorporate a multi-scale feature fusion module to mitigate semantic disparities between high-resolution coarse features and low-resolution fine features. This strategy enhances feature interactions, preventing the loss of intricate ponding details and yielding more precise segmentation results. This work contributes valuable insights to proactive warning research in the realm of road traffic safety.