SAAM-ReflectNet: Sign-aware attention-based multitasking framework for integrated traffic sign detection and retroreflectivity estimation

Abstract

Traffic sign retroreflectivity is essential for roadway safety, particularly in low-light and adverse weather conditions. Traditional methods, such as handheld retroreflectometers and nighttime inspections, are labor-intensive, costly, and unsuitable for large-scale implementation. To address these limitations, we developed SAAM-ReflectNet, a deep learning framework that unifies traffic sign detection, classification, and retroreflectivity estimation into a single automated pipeline. Our RetroNet backbone, developed as part of this study, extracts robust spatial and semantic features to enhance feature representation. The Sign-Aware Attention Module we designed prioritizes critical traffic sign regions, improving detection and classification accuracy by focusing on the most relevant areas. Additionally, our multimodal fusion layers seamlessly integrate RGB imagery with LiDAR intensity data, enabling reliable retroreflectivity estimation. ReflectNet achieved a mean Average Precision (mAP) of 0.635 at IoU=0.5 and 0.522 across IoU thresholds from 0.5 to 0.95, alongside Root Mean Squared Errors (RMSE) of 0.169 for foreground and 0.147 for background reflectivity. Across 15 evaluation runs, performance improvements were statistically significant compared to all baselines (p < 0.05), underscoring the consistency and reliability of ReflectNet.These findings underscore the reliability, scalability, and transferability of our approach, establishing ReflectNet as a transformative tool for intelligent transportation systems and proactive traffic sign maintenance.