Traffic Networks via Neural Networks: Description and Evolution

Abstract

We optimize traffic signal timing sequences for a section of a traffic network in order to reduce congestion based on anticipated demand. The system relies on the accuracy of the predicted traffic demand in time and space which is carried out by a neural network. Specifically, we design, train, and evaluate three different neural network models and assert their capability to describe demand from traffic cameras. To train these neural networks we create location specific time series data by approximating vehicle densities from camera images. Each image passes through a cascade of filtering methods and provides a traffic density estimate corresponding to the camera location at that specific time. The system is showcased using real-time camera images from the traffic network of Goteborg. We specifically test this system in reducing congestion for a small section of the traffic network. To facilitate the learning and resulting prediction we collected images from cameras in that network over a couple of months. We then use the neural network to produce forecasts of traffic demand and adjust the traffic signals within that section. To simulate how congestion will evolve once the traffic signals are adjusted we implement an advanced stochastic model.