Visual Cognitive Driving Distraction Detection using EEG

Overview & Synopsis

Driving distraction shift the attention away from safe driving towards a competing task. Dual task paradigm defines distraction in terms of excessive workload and limited attentional resources. To fill the gap in driver’s safety research, the thesis aims to uncover the mechanics of detecting visual cognitive distraction by using physiological signals. There are 2 sets of experiments conducted with 18 participants. The first experiment with 10 participants aims at evaluating the effects of distraction and synchronising EEG signals. For statistical analysis, Shapiro Wilk test and two-way ANOVA have been used. The results state that the two visual cognitive distraction tasks are performed sequentially and reaction time is affected by the switch. For the second experiment with 8 participants, an integrated system has been created in C++, which involves a 4-wall CAVE system, a driving simulator using SCANeR simulation software, a simple EEG headset with 20 channels (Enobio 20), SMI Eye tracker, BeGaze software. As qualitative methods, SSQ and NASA TLX have been used to evaluate user feedback. As quantitative methods, maths equations, eye movements, EEG signals, and driving performance have been used. For data analysis, Matlab and SPSS and for statistical analysis, Shapiro Wilk test and one-way ANOVA have been used. The results indicate that experiment demands mental capacity, rather than physical capacity. Delta band frequencies in frontal lobe of EEG signals do not have a significant impact on discriminating between driving and distraction, but delta and theta band at central and parietal do. The results also revealed that theta and alpha band frequencies are indeed an important features in both experiments. In a task where subject can control when they are engaging towards the secondary task, beta band showed its statistical importance being the only one able to discriminate up to different level of cognitive tasks. The most striking similarity between both experiments were, some of the statistically significant features are overlapped between brain regions. A comparative analysis of machine learning techniques (SVMs vs Random Forests) in selecting relevant features of EEG signals, referring to the coherence values, has also been conducted. Results indicate that although in general random forests perform better than SVM classifier, F1-score for SVM demonstrates the highest scores (0.90).