Detecting and Estimating On-street Parking Areas from Aerial Images

Abstract

Parking is an essential part of transportation systems and urban planning, but the availability of data on parking is limited and therefore posing problems, for example, estimating search times for parking spaces in travel demand models. This paper presents an on-street parking area prediction model developed using remote sensing and open geospatial data of the German city of Brunswick. Neural networks are used to segment the aerial images in parking and street areas. To enhance the robustness of this detection, multiple predictions over same regions are fused. We enrich this information with publicly available data and formulate a Bayesian inference model to predict the parking area per street meter. The model is estimated and validated using detected parking areas from the aerial images. We find that the prediction accuracy of the parking area model at mid to high levels of parking area per street meter is good, but at lower levels uncertainty increases. Using a Bayesian inference model allows the uncertainty of the prediction to be passed on to subsequent applications to track error propagation. Since only open source data serve as input for the prediction model, a transfer to structurally similar regions, for which no aerial images are available, is possible. The model can be used in a wide range of applications like travel demand models, parking regulation and urban planning.