Pilot's Intent Inference and Aircraft Trajectory Prediction with

Applications to Air Traffic Control

Javier Lovera Yepes[1], Inseok Hwang[2], and Mario Rotea[3]

School of Aeronautics and Astronautics

Purdue University, West Lafayette, IN 47907

Email: jlovera1, ihwang, rotea @ purdue.edu

Air Traffic Control (ATC) is responsible for safely managing and controlling aircraft that operate within the continental Unites States airspace. Due to increased operations, the ATC system is rapidly reaching its maximum capacity. Therefore, new decision support tools for both controllers and pilots have received increasing attention in recent years to help alleviate their workload and increase their efficiency. Some of these decision support tools are founded on the capability of accurately tracking and predicting an aircraft's position in time. Specifically, these capabilities are necessary in applications such as aircraft conflict detection and resolution, or the development of the free flight concept, where pilots will be able to select their course, speed, and altitude in real-time.

We propose an algorithm that could be used as a decision support tool for both current and future concepts of ATC. The algorithm is capable of accurately estimating the states of a maneuvering aircraft (tracking) and inferring the pilot's intent, which are used jointly to compute accurate trajectory predictions in the horizontal plane. For tracking we implement a hybrid estimation algorithm which assumes that the aircraft operates in one of a finite set of dynamical models. These models correspond to the aircraft's modes of operation (e.g. Constant Velocity, Coordinated Turn, etc.). In this way we improve the state estimation performance of a single Kalman filter when the aircraft dynamics deviate from the valid region of the model used in the Kalman filter, which is the case of unexpected maneuvers or changes in modes of operation (e.g. from constant velocity straight flight to coordinated turn flight). Previously, this type of algorithms have also been used for trajectory predictions, however the predictions can significantly diverge from the actual aircraft trajectory due to uncertainties about the environment, the flight plan, and the regulations that influence future pilot's actions or intents. To prevent this and extend the prediction look-ahead time, we use the estimated states together with the pilot's inferred intent. For intent inference we have developed a probabilistic intent inference algorithm that identifies the pilot's intent based on the estimated states, policies and regulations of ATC, the aircraft's flight plan, and environment information such as weather cells, or neighbor aircraft's information. The probabilistic intent inference algorithm has a set of intent models representative of possible pilot's actions in the horizontal, vertical, and speed dimension, which are compared to the actual aircraft states (i.e. position and velocity) to identify the current pilot's intent.  We are motivated by the possibility of creating an algorithm that exhibits the strengths of each technique (i.e. hybrid estimation and intent inference) while mitigating their weaknesses; e.g. short term prediction accuracy of intent inference predictions and long term prediction accuracy of hybrid estimation algorithms that ignore environment. The result is a real-time trajectory prediction algorithm which outperforms previously developed algorithms in intent inference and trajectory prediction accuracy.