Traffic density estimation using computer vision

Abstract

This paper describes a traffic density estimation system (TDES) that contributes to the field of intelligent transportation systems by providing the necessary queue length information to enable the optimal control of traffic lights. Modern computer vision techniques utilize Convolutional Neural Networks (CNN) to determine traffic queue lengths at intersections, including lane detection to ensure that only relevant vehicles are counted. We applied a trade-off study to different CNN models based on speed, accuracy, and implementation efficiency and found that the YOLOv8 model slightly outperformed the R-CNN and SSD architectures. Intelligent traffic control methods then apply Reinforcement Learning (RL) to this data to determine optimal cycle lengths for the different phases of traffic lights. We used the Simulation of Urban Mobility (SUMO) software to simulate a simplified traffic system consisting of traffic lights that manage different traffic levels as part of the daily and weekly traffic cycle. The simulated system allocated priority to emergency vehicles and buses over normal passenger vehicles. Our simulations indicated that the combination of the TDES and RL traffic control outperforms traditional fixed-time traffic control methods for all traffic densities. The results showed that the TDES achieved a 70% mean average precision and a maximum 50% waiting time error. For a variety of traffic scenarios applied to a single intersection architecture, the improved traffic control method reduced average vehicle waiting time by almost 90%.