Work

Exploring driving simulator data classification based on eye metrics

Year: 2022
Type: dissertation
Abstract: Cognitive state denotes for the overall human capacity for cognitive readiness to meet demands in changing everyday situations.Various external and internal conditions, like workload and acute stress affect cognitive processes increasing the likelihood of human errors.Automatic detection of weakened cognitive readiness could be beneficial in many fields where the high human performance is necessity e.g.driving.Changes in the cognitive state are reflected in various biosignals and can thus be detected with the help of machine learning methods.Eyes are the main instrument that allows humans to explore the world around us and they produce a quantity of signals that can be utilized for cognitive states detection.Nevertheless, there is a lack of studies exploring the use of eye metrics in machine learning for cognitive state detection.This thesis explores and evaluates the use of eye metrics in classification of driving simulator data.The thesis demonstrated classification procedure for distinguishing between non-crash, before-crash, after-crash, and crash events in a driving simulator game called CogCarSim.A set of potentially discriminative features from electro-oculography (EOG) and videooculography (VOG) signals were extracted: 7 pupil diameter features, 6 gaze point features, 11 saccade and fixation features from the VOG signal, 10 blink features from the EOG signal.Based on extracted features a set of derivative features were generated.In the suggested approach, several machine learning methods were tested with class balancing techniques and normalization.The best results 69.94% of balanced accuracy was achieved by using the personalized Extreme Gradient Boosting model.The features extracted from pupil and gaze points were found to be the most distinguishing and contributed to the overall performance by 13.78% and 6.31% respectively.The classification performance was increased to 75.84% utilizing both gameplay and eye-based features.While the model trained only on gameplay features performed on par with the proposed model and gave 68.52%.Promising results of the study suggests that the selected approach could serve as a basis for future research such as real-time cognitive state detection and collision prediction.
Language: English
Open Access status: green
OpenAlex ID: W7215526430
Cites: 72
Cited by:
Related to: 9