Citation

Rabbi, Riadul Islam (2026) Driver drowsiness detection using time series physiological signals with deep temporal learning architecture. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Driver drowsiness is a leading factor in road crashes, serious injuries, and fatalities; however, many existing systems still rely on single physiological measures, simulator-based validation, and models that are not suitable for real-time deployment. This study addresses these limitations by proposing and evaluating a physiological driver drowsiness detection framework based on Heart Rate Variability (HRV), Electrodermal Activity (EDA), and Skin Temperature recorded using a wrist-worn Empatica EmbracePlus device for highway, rural, and urban driving. The main objective is to develop an accurate, well-calibrated, and computationally efficient three-level drowsiness detection system suitable for near real-time operation in realistic driving conditions. This work implements a complete pipeline, covering multi-environment experimental protocol design, wrist-worn signal acquisition, preprocessing, and time-frequency representation, followed by deep learning-based detection. Drowsiness is annotated into Alert, Early Drowsy, and Severe Drowsy states using driving context and behavioural cues. Six architectures are developed and compared: 1D-CNN, temporal convolutional network (TCN), long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM), gated recurrent network (GRU), and a new multi-level temporal convolutional network with hybrid attention (MTL-ATCN) using single modality HRV, EDA, and temperature inputs. Performance is evaluated using accuracy, precision, recall, F1- score, and AUC-ROC, including Expected Calibration Error (ECE), and Weighted Multi-Class Detection Score (WM-CDS), which combines discrimination, calibration, and computational cost. HRV emerges as the most informative signal, with MTL-ATCN reaching 93.10% accuracy in rural driving and 96.69% accuracy and F1-score, and 99.62% AUC-ROC on the combined dataset, with an inferencelatency of about 45ms. EDA and temperature models achieve peak accuracies of 89.10% and 84.15%, respectively, with ECE around 6-8%, confirming that wristworn physiological sensing with an attention-based temporal convolutional network can provide robust, well-calibrated, and efficient three-level drowsiness detection under real-road conditions. The findings conclude that wrist-worn physiological sensing, particularly HRV-based modelling, combined with the proposed MTLATCN architecture, provides a reliable, well-calibrated, and near real-time driver drowsiness detection under real-world driving conditions.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: Q325.73 .R33 2026
Uncontrolled Keywords:	Deep learning (Machine learning)
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	03 Jul 2026 01:10
Last Modified:	03 Jul 2026 01:10
URII:	http://shdl.mmu.edu.my/id/eprint/16200

Downloads

Edit (login required)