Learning and Predicting Pilot Behavior in Uncontrolled Airspace

摘要

A new probabilistic method is presented for trajectory prediction in the absence of intent information. Predicted trajectories are drawn from a novel navigation model that represents aircraft motion as a Markov chain of maneuver mode change points, learned by clustering prior observations of aircraft mode change points by position and type of maneuver. The navigation model maps current aircraft state in continuous space to a discrete sequence of mode change points, and generalizes this to predict the position of the aircraft at some time in the future. This approach addresses the weakness of previous Markov model-based approaches that require a fine discretization of the state space by representing only features of the state space where a branch in aircraft trajectory is likely to occur. Learning these features can be achieved with a much smaller training data set than required for other a large, finely discretized Markov model. Since the number of nodes in the navigation model is small relative to the size of the state space, this approach also allows very computationally efficient prediction to be performed in real time. The resulting trajectory predictions are shown to be far more accurate over along time period than can be achieved with current practical algorithms such as TCAS and TSAA, particularly in the challenging uncontrolled airport environment.