Devising a new methodology of traffic signal scheduling for dhaka city leveraging computer vision and multi-objective optimization

Abstract

With the rapid increase of population in the current world, there is an increasing demand for efficient transportation systems. Proper vehicle management in road intersections can ameliorate the efficiency of transportation systems to a great extent. In this regard, one of the most important traffic management steps is traffic signal scheduling at road intersections or junctions. Traffic signal scheduling is also one of the fundamental concepts used in Intelligent Transportation Systems. Many research studies have been conducted on adaptive or traffic-responsive signal scheduling covering vehicle detection, optimization, simulation, etc. However, most of the existing studies have been conducted considering the context of developed countries. These studies have not considered various contexts prevalent in developing countries. Examples of such contexts include non-lane-based traffic, heterogeneous traffic, etc. A special focus covering these distinctive contexts of the developing countries is needed to develop a contextually-appropriate traffic signal scheduling, which is yet to be explored in the literature to the best of our knowledge. To fill the gap in the existing literature, in this study, we propose a new methodology incorporating chronological implementations of computer vision, multi-objective optimization, and simulation for the purpose of developing an efficient traffic signal scheduling mechanism appropriate for Dhaka city. To do so, first, we collect real traffic images from important junctions of Dhaka city and annotate the images by trained human annotators. After a rigorous checking over the annotations, we prepare a complete dataset of 1071 images covering diverse weather and illumination conditions. Then, we perform experimentation on the prepared dataset using state-of-the-art object detection architectures, where YOLOv8n (nano) architecture provides the best inference time (5.3 ms) as well as the second-best mean average precision (68.3 %). Then, we explore three metaheuristics algorithms with two new objectives. Among the three algorithms, we find that real-encoded NSGA-II performs the best. Accordingly, we leverage the best-found model in our rigorous simulation using the simulator, DhakaSim. Our simulation results demonstrate that we can achieve up to 24.1% decrease in average waiting time, 10.4% increase in average vehicle flow rate, and 6.6% increase in average speed.