Towards Urban Vehicle Autonomy: Estimating Urban Congestion from Taxi Pickups and Deliveries

While there are several advantages that can be attributed to the use of autonomous vehicles for personal mobility, their overall impact on surrounding traffic and levels of congestion is as of yet, poorly understood. A crucial stepping stone towards a better understanding of the relation between large fleets of autonomous vehicles and urban congestion is the creation of a predictive model of urban congestion. In this paper, we present a geospatial, time-variant predictive model for urban traffic congestion within Manhattan using a machine learning approach. The model relies on taxi data collected in New York for the month of February 2012 and encompasses features such as number of taxis on the road, time of day, weather, and day of the week and predicts the average speed in a given region. Using a multi-class support vector machine with a Gaussian kernel resulted in an average RMS error under 2 mph with a classification accuracy between 85%-90% for central Manhattan. The relatively high prediction accuracy suggests a strong correlation between the taxi data and the macroscopic traffic pattern and congestion in most of Manhattan. This model constitutes a first step towards analyzing congestion resulting from a fleet of autonomous vehicles.