A digital twin system for the mobile operation of legged robots was proposed, encompassing the design of architecture, module structure, hardware framework, and software framework. The system enabled reliable and accurate acquisition of environmental states and robot states in mobile scenarios by integrating multiple sensor inputs and data sources. The point and line feature matching theory was used to optimize the autonomous positioning accuracy and the robustness of the legged robot, and the odometer functionality and the real-time mobile mapping were effectively achieved through integration with the environmental modeling data. A general modeling method was introduced to establish a digital twin model that ensured high consistency between the simulated robot motion state and the real robot motion state through error compensation techniques. Experimental results on both datasets and real robots demonstrated that the proposed digital twin system not only operated stably and efficiently across various legged robot platforms but also ensured the real-time state feedback and the odometer positioning accuracy. Compared with ORB-SLAM3, the memory overhead was reduced by about 68.7%, and the CPU usage was reduced by about 17.8%. The hardware experiments showed that the communication delay was basically consistent with the network delay of about 30 ms, which helped to improve the efficiency of task execution.