Mining contextual information for urban traffic speed estimation with random forest model

Abstract

In most regions of the world, traffic systems are under increasing pressure as the population and number of automobiles grow. To reduce the burden on the urban road network caused by the increasing number of vehicles, it is essential to know the dynamic traffic speeds on the road network at each time of day to help with real-time vehicle planning to prevent congestion. Understanding traffic speeds requires estimation and prediction based on historical speed data, which is made possible with the help of artificial intelligence and large amounts of historical data. Before the model can be trained, it is essential to process the dataset. For the model to truly reflect the dynamics of the road network, historical speeds and trajectories data need to be matched to appropriate road segments; this process is known as map-matching. Once matched, the model needs to consider contextual factors in addition to speed and time data to reflect the performance of the road network more accurately over time. This research project in turn investigates how to improve the accuracy of the map matching algorithm, how to mine historical speed contextual information and train machine learning models for each period based on the data containing contextual information. After testing and experimentation, it can be demonstrated that the map matching algorithm proposed in this research can cope with high sampling rate speed data and perform well in lower sampling rate datasets. For speed prediction, the mined contextual information is also helpful for training machine learning models with high accuracy.