Research on Image Extraction Model of Road Surface Snow and Ice Area Based on Semantic Segmentation

Xuming Yang¹Haonan Zhang^1*Zhihui Qin^1*Peng Wu^1*Zhong Li¹Ying Han¹Jianping Huang²Bo Hao¹

• ¹Institute of Disaster Prevention, Sanhe, China
• ²National Institute of Natural Hazards, Beijing, China

Accurate identification of road ice and snow is crucial for ensuring winter traffic safety. To address the issue of insufficient segmentation accuracy in existing methods, which is often caused by the loss of details and noise interference in complex road scenes, this paper proposes an improved U-Net model, named Spd-Unet. The model introduces the Spd-Conv module in place of traditional downsampling, which effectively preserves the irregular edges and texture information of ice and snow. Concurrently, it integrates a hybrid attention mechanism combining the Convolutional Block Attention Module (CBAM) and the Spatial and Channel Squeeze & Excitation (SCSE) module to enhance the model's ability to focus on key features. To validate the proposed model, we constructed a dedicated road ice and snow segmentation dataset based on the Canadian Adverse Driving Conditions (CADC) dataset. The experimental results demonstrate that Spd-Unet achieved a mean Intersection over Union (mIoU) of 86.18%, outperforming several mainstream networks such as DeepLabV3+. The model proposed in this study can efficiently and robustly segment road ice and snow regions, providing a reliable technological solution for intelligent transportation safety systems.