EUVIMEDEUVIMED entdecken
Crossref

Πολυτροπική Ανάλυση Νευροφυσιολογικών Σημάτων με Τεχνικές Βαθιάς Μάθησης

Anastasia Oikonomou

DSpace - NTUA (National Technical University of Athens) · 28.08.2026

DOI: 10.26240/heal.ntua.33334 ↗

Kurz gefasst

Fachartikel in der Zeitschrift DSpace - NTUA (National Technical University of Athens), erschienen am 28.08.2026. Verfasst von Anastasia Oikonomou.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The objective of this diploma thesis is the recognition of the intentions of drivers using neurophysical signals under naturalistic driving conditions.More specifically,it is investigated if the Elecroencephalography (EEG) signal contains predictive information regarding the direction of a turning event (left/ right) within a three second time window before the onset.Firstly, turning events were successfully identified through a procedure based on the steering wheel rotation rate.The recorded EEG data were preprocessed and denoised using techniques such as Independent Component Analysis (ICA), digital filters and the AutoReject algorithm.For the feature extraction and classification, deep learning architectures were employed, such as the EEGNet Convolutional Neural Network and models based on Riemannian Geometry.Model performances were evaluated and compared using cross-validation methods of Leave one Session Out and Leave one Subject Out.For the leave Leave one Session Out out cross validation the EEGnet model yielded the best results (Balanced Accuracy 0.675, F1 Score 0.697, AUC 0.721) and in the Leave one Subject out cross validation the Riemann model achieved the best performance (Balanced Accuracy 0.598, F1 Score 0.580, AUC 0.652).The analysis was extenede to the evaluation of ocular signals and more specifically in the three dimention pupil position vector.Unimodal models were implemented and later compared with multimodal approaches using Early and Decision-Level (Late) Fusion.The performances of these models were compared and for the Leave one Session Out validation the ocular unimodal model achieved the best results (Balanced Accuracy 0.740, F1 Score 0.735, AUC 0.812) while for the Leave one Subject Out approach the decision-level model using a Meta-classifier had the best performance (Balanced Accuracy 0.750, AUC 0.828).Finally, an attempt was made to investigate the predictive capability of the EEG signal prior to the ocular signal, however this approach did not have satisfactory results.

Schlagwörter

driver intention recognitionEEGnaturalistic drivingEEGNetRiemannian geometrypupil positionmultimodal fusiondecision fusionfeature-level fusionmeta-classifier

Zitieren

APA

Oikonomou, A. (2026). Πολυτροπική Ανάλυση Νευροφυσιολογικών Σημάτων με Τεχνικές Βαθιάς Μάθησης. DSpace - NTUA (National Technical University of Athens). https://doi.org/10.26240/heal.ntua.33334

Vancouver

Oikonomou A. Πολυτροπική Ανάλυση Νευροφυσιολογικών Σημάτων με Τεχνικές Βαθιάς Μάθησης. DSpace - NTUA (National Technical University of Athens). 2026. doi:10.26240/heal.ntua.33334

BibTeX
@article{Oikonomou2026,
  author = {Oikonomou, Anastasia},
  title = {Πολυτροπική Ανάλυση Νευροφυσιολογικών Σημάτων με Τεχνικές Βαθιάς Μάθησης},
  journal = {DSpace - NTUA (National Technical University of Athens)},
  year = {2026},
  doi = {10.26240/heal.ntua.33334}
}
Kostenlos starten →