Entropy-Based Traffic Flow Labeling for CNN-Based Traffic Congestion Prediction From Meta-Parameters

Abstract

Traffic congestion affects quality of life by inducing frustration and wasting time. The congestion is also critical to vehicles with high emergencies such as ambulances or police cars. This leads to additional CO2 emissions. Traffic management requires the accurate modeling of congestion levels. Two main observable parameters identify the congestion state of a city: vehicle speed and density. Congestion has an intuitive definition rather than a quantitative one, and is associated with the disorder and randomness occurring in traffic parameters. Therefore, statistical analysis offers an efficient and natural framework for modeling such disorders. In this study, a differential-entropy-based approach was proposed for labelling purposes. Subsequently, supervised congestion prediction from traffic meta-parameters based on a convolutional neural network was proposed. Traffic parameters includes node localization, date, day of the week, time of day, special road conditions, and holidays. The proposed model is validated on the CityPulse dataset, which is a set of vehicle traffic records, collected in Aarhus city in Denmark over a period of six months, for 449 observation nodes. Simulation results on the CityPulse dataset illustrate that the proposed approach yields accurate prediction rates for different nodes considered. The proposed system can prevent traffic congestion by reorienting the drivers to follow other itineraries.

Keywords

• Shannon entropy
• deep CNN
• smart city
• traffic congestion state prediction