A Machine Learning-Enhanced Digital Twin Framework for Real-Time Traffic Signal Management at Intersections

This paper presents a trajectory-calibrated, data-driven digital twin (DT) of a real signalized intersection, built to improve microscopic accuracy between real traffic and its virtual model. Motivated by past work that validates DTs mostly with aggregate measures (flows, speeds, delays), we create a modular video-to-simulation pipeline using publicly available drone footage from the South Cleveland Avenue-Daniels Parkway intersection in Fort Myers, Florida. The system converts aerial video into time-stamped vehicle paths, maps them into the Simulation of Urban Mobility (SUMO) coordinate space, builds laneconsistent routes and departures, and runs them in a microscopic simulation. Calibration uses path-level metrics (RMSE, timing drift (TD), lane agreement) along with common measures such as queues and delays. Results show that the approach closely reproduces real traffic with strong spatial and temporal accuracy and offers a clear SUMO-TraCI platform that supports future machine-learning-based prediction and control.