Cross-Subject Vigilance Estimation Based On Adversarial Domain Adaptation Networks

The lack of alertness is one of the most important causes of traffic accidents.Since EEG can reflect the physiological and psychological state of people,EEG-based vigilance level estimation has attracted much attention as a research topic.Around this topic,re-searchers have published many articles and proposed many methods.However,almost all these studies focus on building,training,and applying models for one single driver,which requires obtaining a certain amount of the driver’s EEG signal data and corresponding vigilance labels for training of the model.However,the collection of labelled data is time consuming and labor intensive,which makes it difficult to apply these methods directly to the real-world environment.To solve this problem,domain adaptation methods can be used to eliminate the domain discrepancies in feature distribution between drivers,so that one model can be applied to different people.In this way,the labelled data from different drivers can be used to train the target driver’s vigilance estimation model,while the target driver only needs to provide the unlabelled data.In this thesis,we adopt the adversarial domain adaptation networks to eliminate the domain discrepancies.Such methods have shown good performance in recent research on image recognition.In comparison,we also discussed several other domain adaptation methods,none of which applied the idea of adversarial training.We use the publicly available dataset,SEED-VIG,to evaluate the methods.The SEED-VIG dataset contains EEG,EOG data,and the corresponding real-time PERCLOS indicators collected from 23 subjects when they were driving in a virtual driving system.We extracted the forehead EEG and EOG features as input data and use the PERCLOS indicators as labels to train and test the models.As the experimental result demonstrates,the domain adaptation adversarial networks considerably improves the vigilance estimation precision and has good stability(outperform the baseline method by over 10%).