Development of an ML model for traffic jam management

Abstract

Urban traffic congestion remains a critical issue due to inefficiencies in static traffic signal control systems. This paper introduces an adaptive traffic control framework combining real-time vehicle detection with the application of YOLOv8 and dynamic signal timing optimization. The YOLOv8 model was trained with a custom data-set of 6,398 annotated traffic images with a vehicle detection performance of 90% mean average precision (mAP). The optimal green signal durations are calculated by the system based on vehicle density and type for each lane with the constraints of 10–60 seconds. A Pygame-based simulator compared the performance of four scenarios: static timing with equal/prioritized lanes and adaptive control with equal/prioritized lanes. The simulation run results of 40 (3 minutes each) revealed a 7–10% reduction in average waiting time and a 12% increase in the number of served vehicles for the adaptive scenarios compared to static systems. The combination of computer vision and machine learning presents a scalable solution for reducing traffic congestion in smart cities.