On-Street Parking Prediction Using Real-Time Data

Abstract

Drivers looking for parking account for high percentages of traffic congestion. One promising way to tackle this problem is through parking availability prediction. In this paper four algorithms for on-street parking prediction are described, analyzed and compared using real data. The first one uses only the availability's historical mean and standard deviation with no assumption on its distribution. The second assumes that availability is normally distributed. The third uses real-time information and models the availability variation as a normal random variable. The fourth also uses real-time information but models vehicle arrivals and departures as non-homogeneous Poisson processes. This comparative analysis supports the following: use of real-time parking availability information improves prediction performance, classification metrics are necessary for proper algorithm evaluation, and predicting arrivals and departures independently leads to fewer false positive recommendations at the cost of more false negatives.