Research on Snow Detection of Urban Public Transport Road Area from the Perspective of DeepVision-WetNet Framework

This paper proposes a lightweight object detection model for snow covered road scenes based on the YOLO11 framework to address issues such as boundary blurring, overexposure of brightness, and texture loss caused by snow on winter roads. Firstly, design an enhanced upsampling and multi-scale convolution module EUMSCB, which achieves efficient collaboration between local details and global semantics through multi-scale depth separable convolution and channel shuffling fusion. Secondly, a lightweight adaptive sampling unit LASU is introduced to achieve dynamic upsampling using learnable offsets and range constraints, significantly improving the structural recovery of snow edges and thin snow areas. Finally, a shared enhanced detection head Detect_SEPS is constructed to improve detection robustness under complex scale distributions through multi-layer shared backbones, decoupled classification and regression branches, and learnable scale weights. Experiments based on a self built road snow dataset show that the proposed method achieves significant improvements in mAP and small object detection performance compared to the original YOLO11 while maintaining real-time performance, providing effective technical support for winter road safety monitoring.