Abstract
Our actions and intentions characterize the movement patterns of our eyes. Our visual exploration is driven by a mixture of cognitive processes and a conflict between the inspection of detail and the maintenance of an up-to-date overview. As a consequence, the determination of the influence of separate behavioral factors is challenging. The work at hand examines how eye movement sequences can be compared to each other. This process is at the core of almost every eye-tracking study as it answers questions such as: "Does gaze behavior of a patient group differ from his/her control group?," "How does experts’ visual exploration differ from novices’?," "How does the composition of a painting influence the observer’s gaze?" Therefore, several eye movement processing steps are revised: The issue of data quality is discussed with focus on methods and benchmarks to assure good quality during pupil detection, gaze mapping and eye movement identification in dynamic scenarios. Furthermore, eye-tracking data is integrated with physiological parameters such as ECG, galvanic skin conductivity and pupil dilation. The fusion of these complementary physiological sensors helps to disambiguate gaze and attention allocation. This thesis proposes a novel method for the comparison of visual scan patterns, which is based on the frequency of short snippets of the whole eye movement sequence. Combined with current techniques in machine learning, the method is adaptable to a multitude of applications. Visualization and aggregation procedures for frequently traversed gaze trails are demonstrated on the basis of the finding that these short patterns are highly characteristic to many applications. The proposed comparison technique is evaluated against state-of-the-art approaches on a new collection of data from a broad spectrum of eye-tracking experiments, ranging from static viewing tasks to highly dynamic outdoor scenarios. It effectively predicts the observer’s task in a conjunction search task and in the more complex Yarbus experiment significantly above chance level. Furthermore, it is possible to assess driving fitness and the driver’s secondary task, as well as to classify the expertise of neurosurgeons. This new approach is able to identify the influence of a single experimental factor upon eye movement sequences. In contrast to all competing methods, it generalizes well over a broad spectrum of experimental designs.
Author and committee
dc:creator, dc:contributor.*- Author
-
- Kübler, Thomas Christian Alexander
Identifiers
dc:identifier.*- Identifier
- hdl:10900/74458