Learning Traffic Patterns at Small Airports From Flight Tracks

The majority of reported near-midair collisions that involve a general aviation aircraft occur in the vicinity of nontowered airports. A prior work has investigated the feasibility of creating an automated air traffic control system for these nontowered airports using solutions to a partially observable Markov decision process. Validating such system will require an accurate model of aircraft behavior in the traffic pattern. This paper evaluates the different approaches for deriving traffic pattern models from recorded radar data. The first approach is based on prior trajectory clustering work, where turning points in trajectories are identified and clustered. This method performs well on simulated data, but due to its reliance on noisy heading rates, it has difficulty with real-world data. The second approach uses Bayesian inference techniques to learn the parameters of the traffic pattern model, where a hidden semi-Markov model with a hierarchical Dirichlet process as a prior is investigated. Inference in this model is made computationally tractable using Markov chain Monte Carlo methods. The turning point and Bayesian models are compared with each other using different f-divergence measures, and the latter is found to better represent the observed data.