Weather-Robust Traffic Sign Recognition through Hybrid Deep Learning and Blended-Realism Data Generation

Developing Countries, such as Bangladesh, often face major challenges in accurately detecting traffic signs because of adverse weather conditions, such as heavy rain, fog, poor lighting, and nighttime driving, as well as regional differences in how signs look and are maintained. This study presents a weather-robust framework for traffic sign detection specifically designed for conditions commonly encountered on Bangladeshi roads. It relies on a carefully built hybrid dataset of 27,000 images, which combines real-world photographs captured in the field with context-aware synthetic samples generated to mimic six challenging weather scenarios: rainy, foggy, low-light, night, sunny, and blurry conditions. The main contribution of this study is the development of a large-scale hybrid dataset combining real and synthetic images of 11 common Bangladesh Road Transport Authority (BRTA) traffic signs, along with a novel YOLOv10-ResNet50 hybrid model that achieves high detection accuracy under diverse and low-visibility conditions in developing regions. This addresses a critical gap by improving the generalization and performance of the model in challenging real-world environments. The proposed model achieved state-of-the-art performance with 79.3% Mean Average Precision (mAP@0.5), a precision of 0.821, a recall of 0.806, and an F1-score of 0.813, while maintaining robust accuracy in challenging night and low-light conditions compared to the baselines. These results demonstrate the effectiveness of hybrid real-blended data and advanced deep learning for reliable traffic sign detection. More importantly, this study aims to enhance road safety in Bangladesh by supporting driver assistance and autonomous systems, ultimately reducing accidents and saving lives in resource-limited environments.