An Intelligent Traffic Management Model Using Large Language Model Agents for Adaptive Traffic Signal Control in Nairobi, Kenya

Abstract

Abstract Traffic congestion has remained a significant challenge in rapidly urbanizing cities such as Nairobi, Kenya, affecting mobility, productivity, environmental quality, and overall quality of urban life. Increased vehicle ownership, inadequate infrastructure expansion, and inefficient traffic management systems have exacerbated delays, air pollution, and economic losses. Conventional traffic signal control systems, largely based on fixed timer logic, have often failed to respond effectively to real-time traffic dynamics, particularly at intersections where delays are most severe. This study developed and evaluated an Intelligent Traffic Management Model for Nairobi. The model integrated Large Language Model (LLM) agents to enable adaptive traffic signal control. Video traffic data were collected from selected intersections and processed using the YOLOv5 object detection algorithm due to its proven accuracy in real-time vehicle detection. These processed vehicle density estimates were then used to calibrate a traffic simulation framework within the Simulation of Urban Mobility (SUMO) environment, which was selected for its flexibility and ability to replicate realistic intersection behaviour. Within the simulation, LLM agents were integrated via an API call to analyze lane-specific traffic conditions and adjust green light durations dynamically based on observed vehicle counts and congestion levels. The research used experimental methodology, conducting scenario-based testing under varying traffic conditions, including peak-hour traffic, off peak flows, road closures, and emergency vehicle passage. Performance was evaluated using key metrics such as average vehicle waiting time, intersection throughput, and responsiveness to changing traffic demand. The results indicated that the LLM-driven adaptive control model significantly reduced average vehicle waiting times and improved intersection throughput compared to fixed-time signal control. The system also demonstrated a higher responsiveness to dynamic traffic variations, contributing to smoother traffic flow and more balanced lane usage. This study concludes that integrating LLM agents into adaptive traffic signal control systems offers a scalable, data-driven approach to improving urban mobility in complex and resource-constrained environments. It recommends that transport authorities in Nairobi and similar cities consider pilot implementations of AI-powered traffic signal optimization as part of broader Intelligent Transportation Systems (ITS) strategies.