Traffic demand and longer term forecasting from real-time observations

We propose an end to end system comprised of real-time image processing in combination with neural networks to describe upcoming traffic demand in order to forecast short as well as longer term traffic evolution and congestion from real-time traffic camera images. The neural networks use both current and historic traffic information collected by analyzing traffic camera images from a number of roads in an actual traffic network. Specifically we design and train a long short-term memory, a gated recurrent unit and a stacked autoencoder network. We train these networks on data from a single camera location and use each of the three networks to predict traffic density by processing images arriving in real time at all the other camera locations in this traffic network. The results reveal that such a system could be helpful to provide information about traffic demand and formation of congestion for hours into the future. The traffic data used is collected from the traffic network of Goteborg in Sweden.