Master thesis : Determining several levels of cognitive load based on the analysis of ocular metrics

Abstract

Cognitive load is the mental effort required to perform a task. However, this concept does not have a single definition and can also be defined as the amount of information that the brain can handle at a given time. This amount is unfortunately limited. Indeed, if the brain receives to many information simultaneously, it will be incapable of handling them all. For this reason, predicting cognitive load is valuable. Indeed, the knowledge of the cognitive load of an individual could allow actions to be taken in order to reduce it. For instance, the environment of the individual can be simplified or relevant information can be emphasized. The aim of this thesis is to be able to predict the cognitive load based on eye move- ments but also on pupil dilation. It is essential that cognitive load can be detected in a wide range of contexts. Furthermore, it is necessary that the user is not hindered by the device used to predict cognitive load. For these reasons, ocular measures are the most suitable indicator of cognitive load level. Indeed, to collect eye data only a camera is needed, which is not invasive for the user. Different measures have been extracted based on pupil dilation but also on the varia- tion of the field of view. Unfortunately, the extraction of the saccades and fixations from the raw data is not possible. Nevertheless, a research study has been carried out to show that a method of extracting saccades and fixations is possible under certain conditions. For this purpose, a data collection involving a callibration under strict conditions on a large number of participants had to be carried out. The extracted data are subsequently used to train a model which predicts the level of cognitive load. The model used is Ran- domForestClassifier from scikit-learn. In addition, another model was trained on the basis of previously calculated measurements of eyelid movements. Finally, the results obtained on the 5-level labeling are not efficient enough. Moreover, for a large number of potential applications, it is only necessary to be informed when the cognitive load is too high. The model can therefore be evaluated in two areas: acceptable cognitive load and too high cognitive load. This greatly increases performance. Finally, a binary model was also trained for both sets of parameters. The results obtained are unfortunately not up to expectations. However, these results must be tempered. Indeed, the cognitive load is a difficult concept to characterise which makes the labelling complex. This is why potential improvements are explained at the end of this work.