In response to the growing demand for pedestrian crosswalk safety and visibility in the era of autonomous driving, this paper proposes and designs a crosswalk environment monitoring system based on virtual-physical twin technology. Leveraging Unreal Engine (UE) and Carla simulation software, the system constructs a high-precision virtual environment, integrating real-time data from on-site sensors and weather APIs to achieve comprehensive and dynamic monitoring of crosswalk conditions. A key innovation of the system is the application of a Vision-Language Model (VLM), which automatically translates monitoring data into understandable natural language text while also enabling threshold-based warnings and safety alerts. This enhances the intelligence and automation of the monitoring system. Compared to traditional monitoring methods that primarily rely on graphical visualization, the proposed approach significantly improves data processing efficiency and safety alert capabilities, offering a more efficient, real-time, and intelligent solution for traffic safety management in the context of autonomous driving.