ForecastIT: CCTV-Driven Urban Traffic Dynamics Forecasting With Road Junction-to-Link Correlations

Abstract

With the rapid proliferation of intelligent driving, Internet of Vehicles (IoV) has become an indispensable infrastructure in urban environments. Accurate and real-time prediction of urban traffic dynamics is critical for collaborative driving, vehicular edge computing, and mobility-as-a-service (MaaS) players (e.g., DiDi, Gaode Map, and Uber). In realistic urban roads, the physical connectivity between intersections and road segments plays a pivotal role in shaping traffic dynamics. By incorporating road structural influences into traffic forecasting, we can obtain more accurate and interpretable prediction results. To explicitly capture such interactions, we abstract intersections as junctions and the connecting road segments between them as links, forming a dual-granularity representation of the urban network. Here, we propose ForecastIT, a Transformer-based spatio–temporal forecasting framework that hierarchically models both junction-level and link-level traffic patterns. ForecastIT integrates adaptive attention mechanisms and cross-scale fusion modules to jointly capture temporal dynamics and structural dependencies. Beyond the IoV, the city’s other major network is the closed-circuit television (CCTV) system, from which detailed vehicle information (e.g., vehicle counts, road occupancy, and traffic violations) can be extracted. In this work, we synergize CCTV streams and IoV networks through the proposed ForecastIT, supporting city-scale real-time traffic prediction without requiring additional construction. Leveraging a month-long CCTV dataset collected from over a thousand urban cameras in Jiaxing, China, our model demonstrates superior accuracy and interpretability compared with state-of-the-art baselines.