AlertNet: Deep convolutional-recurrent neural network model for driving alertness detection

Nissimagoudar, P. C.; Nandi, A. V.; Patil, Aakanksha; H. M., Gireesha · 2021 · Crossref

DOI: 10.11591/ijece.v11i4.pp3529-3538

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical safety issue of drowsy driving by proposing AlertNet, a deep learning model designed to detect driver alertness levels using electroencephalography (EEG) signals. The motivation stems from the reliability of EEG as a direct biological indicator of brain state, specifically for identifying the transition from wakefulness to sleep onset. Traditional methods often rely on manual feature extraction and struggle with the complex, temporal nature of EEG data. AlertNet aims to automate this process by eliminating the need for hand-engineered features and conventional preprocessing steps like artifact removal, thereby reducing computational complexity while improving classification accuracy. The methodology employs a hybrid architecture combining Residual Networks (ResNets) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks within a sequence-to-sequence framework. Raw EEG signals from the Sleep-EDF 2013 and 2018 datasets are segmented into 30-second epochs and normalized. The ResNet component, utilizing 1D convolutions and skip connections, extracts temporal and frequency features while mitigating vanishing gradient problems in deep networks. These features are then fed into an encoder-decoder structure built with Bi-LSTMs, which process data in both forward and backward directions to capture long-term dependencies. An attention mechanism is integrated into the decoder to focus on significant portions of the encoded sequence. To address the inherent class imbalance in the datasets—where wake and deep sleep stages outnumber the critical N1 (sleep onset) stage—the authors employ a combined Mean Squared Error (MSE) and Mean Squared False Error (MSFE) loss function. The model was trained using k-fold cross-validation (k=20 for 2013 data, k=10 for 2018 data) with the RMSprop optimizer. The results demonstrate that AlertNet achieves an overall accuracy of 87.92% and 87.05% on the Sleep-EDF 2013 and 2018 datasets, respectively. The model attained macro-F1 scores of 78.06% and 79.66%, and Cohen’s Kappa scores of 0.78 and 0.79. Notably, the model showed improved performance in classifying the N1 sleep stage compared to existing literature, a crucial capability for early drowsiness detection. The use of the custom loss function effectively balanced error rates between majority and minority classes. Comparisons with state-of-the-art methods indicate that AlertNet’s approach to handling sequential data and class imbalance yields competitive or superior results, particularly in distinguishing subtle transitions in alertness. The significance of this work lies in its ability to automatically classify sleep stages from raw EEG signals with high precision, specifically enhancing the detection of sleep onset. By leveraging deep residual learning and attention-based sequence modeling, the study provides a robust framework for real-time driver monitoring systems. The findings suggest that such models can be adapted for other sequential classification problems involving imbalanced data, offering a viable solution for reducing accidents caused by drowsy driving through automated, biologically grounded alertness detection.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms