Network-Level Estimation of Traffic Signal Performance Measures Using High-Resolution Event-Based Data from Legacy Detection Systems

Abstract Collecting traffic delay and arrival-on-green (AoG) at the network level to support reliable traffic studies, such as traffic monitoring and traffic control, is currently challenging and time-consuming. This is largely because most detection systems at signalized intersections are legacy systems designed for signal control rather than performance measurement collection. These systems typically use a single detector to cover multiple lanes instead of lane-by-lane configurations, limiting their ability to capture traffic performance measures and often requiring costly system upgrades. To obtain accurate network-level performance measures in a cost-efficient manner and overcome the limitations of restricted detector layouts used in automated traffic signal performance measures (ATSPMs) collection, this study proposes a meta-learning-based approach utilizing model-agnostic meta-learning (MAML) to estimate traffic delay and AoG using high-resolution event-based data. The method leverages event-based data from legacy detection systems while enhancing model transferability across different intersections and detection configurations. To evaluate the model’s performance, 153 signalized intersections in the Greater Tucson metropolitan area in Arizona were selected as study locations, using event-based data from two types of traffic detection sensors. The evaluation results show that the mean absolute percent error (MAPE) for control delay estimation ranges from 12% to 22% for through movements, and from 22% to 27% for left-turn movements. For AoG estimation, the MAPE for through movement with advance detectors is 13%–30%, but locations with presence detectors have a relatively higher error. These results indicate that the proposed method achieves reasonably accurate and reliable estimation of network-level performance measures, regardless of traffic detection configurations and intersection layouts.