Model-based head pose estimation for air-traffic controllers

We present a method for estimating the point of fixation of an air traffic controller from a low resolution video sequence. A geometric model of the head is used to estimate head orientation; head pose estimates are combined with a 3D model of the environment to compute the target of gaze. The head model is constructed from a small set of images. Two lighting models are considered: in the first, we only use ambient lighting; in the second, we add a finite distance point source. In both cases, we jointly estimate the albedo of each facet of the head model and the parameters of the lighting model. Because ground-truth data are unavailable, the absolute accuracy of the gaze estimates is unknown. With either method, the results are sufficiently accurate to answer questions of operational interest, such as "is the controller looking out the window.".