Designing a deep-learning network for traffic density and volume prediction

Abstract: It is relatively easy to know the day to day traffic flow on a highway without taking into account on which lane the cars are driving on. It is more difficult to understand how traffic over time evolves while looking at the lanes, or more specifically, by looking at traffic volume on the specific lanes on a highway, can we have a good guess for where the vehicles should be on a different lane of the same highway in the future? And could it also be possible to predict future congestion in one juncture by looking at incoming traffic from the past at another or multiple junctions somewhere else? This project serves as a precursor for possible future projects by looking at how traffic volume on a given road with 5 lanes evolves over time. The idea is to use past data of traffic volume and then make a model to predict how traffic volume looks in the future specifically for these lanes at certain hours. The historical traffic data was fed into something called an artificial neural network. The process of how an artificial neural network is operating is inspired from how neurons in the brain are operating. A signal is sent from one neuron to another, if this signal is strong enough it will be passed on to the next neuron. However, if the signal is too weak, it will not be passed to the next neuron. These neurons can also receive multiple signals from different neurons. Signals sent between neurons will strengthen the connection between these two. The more signals being sent, the stronger the connections. An artificial neural network works in a similar manner. Between each artificial neuron there is a connection, also known as weight. These weights must be adjusted in such a way so that the network will give good enough results. The traffic data is split up in three parts, one part is the traffic volume at different times, the other part is the corresponding future traffic volume on the same lane and the final part is the test data. Feeding the network the data and adjusting the weights with respect to this information, is called supervised learning. Supervised learning is when the input to the network and its corresponding output are both known. What is unknown, however, are the weights. When training has been completed, the third part of the data is used to test the accuracy of the network.