Computational Methods for Eye-Tracking Analysis: Applications to Autism

Though eye-tracking technology has developed considerably over the last hundred years, eye-tracking analysis is still in its infancy. This thesis describes computational techniques and methods we have developed for augmenting this analysis. These methods correct deficits in current approaches, extend traditional techniques to gain greater clarity, and provide frameworks for viewing gaze patterns from new perspectives. We use our methods to study autism, a disorder characterized by social and communicative deficits, in order to gain insight into how these individuals attend to the world around them, and to determine what factors may be motivating their attention. We begin by showing how current fixation algorithms for eye-tracking analysis provide an incomplete picture of gaze behavior. We present a simple linear interpolation model (SLIM) that can provide a more complete, but still compact, picture. We apply this model to the scanning patterns of toddlers with autism and show results which coincide with known deficits in face processing. Furthermore, by adapting standard fixation algorithms to perform temporally greedy box-counting, we provide evidence that the incompleteness of standard algorithms may be due to the fractal qualities of the underlying scanning distributions. Examining distributional aspects of scanning provides only an overview of differences. For this reason we examine standard, fine-grained, region-of-interest (ROI) eye-tracking analysis where regions are drawn around areas and measures, such as how long a subject looks at areas, are calculated. Typically, dynamics of scanning are ignored. To correct this, an entropy measure, as an index of exploration, is proposed and applied to children with autism. We show a pervasive pattern of inattention in autism differentiating 4 year old, but not 2 year old, children with autism from typical children, and discuss how atypical experience and intrinsic biases may affect development. As an alternative to ROI analysis, which can be a subjective and laborious top-down approach, a bottom-up evaluation, based on computational modeling of low-level features, is offered. We use these models to examine preferences for low-level features in autism, and show that children with autism attend more to areas of contrast and less to areas of motion. We also use these same models for gauging the gaze distance between individuals. We use these techniques to highlight the heterogeneity of autism, showing how gaze patterns of individuals with autism are as different from each other as they are from typical controls, and discuss the factors which might lead to this heterogeneity. Finally, we conclude with a discussion of the advantages of the methodologies that we have presented, and discuss the results of our work as they pertain to both the computational and methodological advances we have accomplished and the insights that we have obtained regarding autism.