A Multi-scale Fusion and Dynamic Upsampling Model for Road Surface Snow Detection

A Multi-scale Fusion and Dynamic Upsampling Model for Road Surface Snow Detection

연구 분야: Verification

학회: International Conference on Cognitive Computation and Systems

The issue of road surface snow accumulation during winter is one of the leading causes of traffic accidents. Therefore, real-time road surface snow detection is a crucial research topic in the field of traffic management. Currently, manual sensor-based and traditional machine learning detection methods face challenges such as high labor consuming and resource costs, low accuracy, substantial computational load and poor robustness. To address these issues, this study proposes a real-time road surface snow detection model, YOLOv8-FCD, which incorporates multi-scale fusion and dynamic upsampling, aiming for efficient winter road snow detection. The model integrates the C2f_Faster module into the backbone network, achieving a lightweight design while ensuring detection accuracy. Additionally, the SPPF module incorporates a C3STR module based on Swin Transformer to enhance backdrop differentiation and snow feature extraction. The neck network then uses the DySample upsampling operator to increase the accuracy and resilience of the model while successfully lowering the number of parameters. Finally, utilizing a real-world road surface snow dataset gathered by in-vehicle cameras operated in Canada throughout the winter seasons, testing results show that the proposed YOLOv8-FCD achieves a significant gain in accuracy, mAP@50%, and mAP@50–95% when compared to other benchmark models.