One way to investigate how humans interact with their environment is by studying how they visually search and gather relevant information in order to make decisions. Visual search can be analyzed through visual scan paths, the time ordered sequence of eye fixations and saccadic movements. To create visual scan paths, researchers often use eye fixation detection algorithms, many of which rely on threshold parameters set by the researcher, to automatically identify eye fixations from data collected by eye tracking devices.

However, the choice of threshold parameters used by eye movement detection algorithms is crucial, as many different factors, such as the participant population and the task to be completed, might affect what threshold values can accurately identify eye fixations. Inaccurate thresholds might result in visual scan paths that do not resemble the visual scan path carried out by an individual (i.e., the ideal visual scan path). For example, an inaccurate threshold might fail to identify eye fixations that took place, or combine multiple consecutive eye fixations together into a single eye fixation and place it somewhere in the environment that the individual never actually observed. As such, using inaccurate thresholds might affect our ability to understand and interpret an individual’s visual search (e.g., what information was observed, as well the order it was observed in) and decision-making process.

In this dissertation, novel procedures and algorithms are introduced to facilitate the identification and selection of accurate thresholds. First, an automated procedure was developed to automatically select accurate thresholds based on the impact of threshold values on eye movement metrics (e.g., number of eye fixation), expanding upon prior research efforts by automating a process that was previously largely manual. Second, two approaches are proposed to approximate the trend of similarities between ideal visual scan paths and visual scan paths created at different thresholds, used in prior studies to determine accurate thresholds, without the need to know or use ideal visual scan paths. Using ideal visual scan paths is not always feasible, as one needs to know the expected eye movements of individuals a head of time or needs to engage in the arduous and time consuming process of manually defining the ideal visual scan paths from the data collected. Third, and lastly, a classification framework was developed to identify similar visual scan paths that might showcase variations of a common visual scanning strategy using multiple similarity metrics.