Urban traffic congestion mapping using bus mobility data

Mobility traces of public transportation vehicles, maintained by local government departments, are a very widely available but hardly used data source for traffic congestion analysis in a city. In this paper, we describe our experiences with the historical bus mobility trace data from the New York City MTA, collected over a period of 3 months in 2014. Predicting the congestion state of a road segment is difficult because of complex spatiotemporal dependencies. Our work focuses on the prediction of congestion state of a road segment, given historical trends and information from neighboring road segments. We leverage deep learning architectures such as LSTMs in order to help capture the longer term dependencies in this modelling task. We demonstrate the feasibility of using such a data source for traffic speed prediction and forecasting by showing that we are able to estimate future speeds in a segment with an error of less than 2 m/s (average RMSE ∼ 1 m/s) despite the presence of large amount of gaps and noise in the data. We believe a system such as this for congestion prediction and mapping using limited data would help both citizens as well as urban planners.