A real- time spatial occupancy based framework for predicting green signal time

Urban intersections are responsible for nearly 40% of total vehicular delays in metropolitan networks, making traffic signal optimization a central challenge for sustainable urban mobility. Traditional signal control systems remain constrained by rule-based logic or simple vehicle counting, often failing under heterogeneous and congested traffic conditions common in rapidly developing cities. To address this, the present study introduces a real-time, spatial occupancy–driven framework for adaptive green signal prediction built upon a novel hybrid two-stream detection architecture integrated with a visibility-aware fusion mechanism. The proposed model combines anchor-based and anchor-free detection paradigms through a learnable gating layer, enabling accurate recognition of vehicles of varying scales and orientations even under partial occlusion. A dedicated visibility head further refines confidence estimation, enhancing detection reliability in dense urban scenes. Experiments conducted on real-world intersection video data demonstrate that the framework achieves an mAP@0.5 of 93.7%, a 37% reduction in spatial occupancy estimation error, and a 40% decrease in green signal prediction error compared to baseline systems, while sustaining a real-time throughput of 64.7 FPS and an end-to-end perception-to-prediction latency of 61 ms, confirming its viability for real-time traffic control. Beyond practical deployment, this work contributes a novel methodological bridge between computer vision and intelligent transportation systems, offering the research community a benchmark framework for integrating perception-aware learning into traffic management, signal optimization, and broader smart-city decision infrastructures. Clinical Trial Not Applicable.