Download files
Abstract
This thesis presents an investigation and a framework for a cognitive load measurement (CLM) system based on eye activity, with the aim of allowing machines to understand human dynamic cognitive load changes, so that they can adapt accordingly during tasks. This is achieved by automatically extracting three types of eye activity captured from infrared (IR) webcams placed near the eye – pupil diameter, blink, eye movement (fixation and saccade); and building cognitive load models by means of supervised machine learning to automatically and continuously estimate the level of cognitive load, perceptual load and task segments.
As the measurement accuracy of pupil diameter and blink from commercial eye trackers is often unknown, and the measurement problem using off-the-shelf IR webcams has not been investigated, an efficient and robust algorithm was developed and evaluated for estimating pupil diameter and blink without parameter tuning. To address the efficacy of the eye-based approach in less controlled environments, experiments are conducted initially to investigate the susceptibility of CLM to interfering emotion stimuli and luminance changes in the task backgrounds, as opposed to tightly-controlled conditions in previous studies. Based on several theories and findings from cognitive psychology, an empirical study is conducted to investigate the impact of high perceptual load and task transition on eye-based CLM.
A novel eye-based CLM system configuration is then proposed and evaluated in a real time context with 2-s sliding windows, achieving encouraging results. Key new insights include: (1) using low cost IR webcams to capture eye activity is a better solution than eye trackers for CLM due to being calibration-free and having accurate acquisition - blink detection accuracy was above 99% and pupil diameter was very close to the manually measured size; (2) eye-based CLM can be performed effectively in less controlled illumination backgrounds and with emotion interference; (3) pupil and blink measures were found unable to index cognitive load when perceptual load is high and during task transitions, indicating that perceptual load and task transition should be estimated separately; and (4) eye measures are able to detect perceptual load, cognitive load and task transition in one system.