Optimization of Traffic Signals Using Deep Learning Neural Networks

Abstract

Reducing traffic delay at signalized intersections is a key objective of intelligent transport systems. Many existing applications do not have the intelligence embedded to learn about the environmental parameters (such weather, incident etc.) that influence traffic flow; therefore, they are passive to the dynamic nature of vehicle traffic. This report proposes a deep learning neural networks method to optimise traffic flow and reduce congestion at key intersections, which will enhance the ability of signalized intersections to respond to changing traffic and environmental conditions. The input features of the proposed methods are composed of historical data of all the movements of an intended intersection, time series and environmental variables such as weather conditions etc. The method can learn about the region and predict traffic volumes at any point in time. The output (i.e. predicted traffic volume) is fed into the delay equation that generates best green times to manage traffic delay. The performance of our method is measured by root mean squared error (RMSE), against other models: Radial Basic Function, Random Walk, Support Vector Machine and BP Neural Network. Experiments conducted on real datasets show that our deep neural network method outperforms other methods and can be deployed to optimize the operations of traffic signals.