EEG-Based Drowsiness Estimation for Driving Safety Using Deep Q-Learning

Ming, Yurui; Wu, Dongrui; Wang, Yu-Kai; Shi, Yuhui; Lin, Chin-Teng · 2021 · Crossref

DOI: 10.1109/tetci.2020.2997031

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 issue of driving safety by proposing a novel method for estimating driver drowsiness using electroencephalogram (EEG) data and deep reinforcement learning. Fatigue is a leading cause of road fatalities, and while EEG provides direct insight into brain activity, traditional supervised learning methods suffer from inefficient data utilization and reliance on labeled data that is often unavailable in real-time scenarios. The authors aim to overcome these limitations by formulating drowsiness estimation as a reinforcement learning task, specifically using Deep Q-Learning (DQN), to indirectly estimate response time (RT) as a proxy for mental state. The study utilizes EEG data collected from 37 subjects during a 90-minute simulated endurance driving test. The experimental setup involved a maneuverable platform with a converted car and a projected highway scenario. EEG signals were recorded using a 32-channel cap at 500 Hz. To simulate fatigue, random turbulence caused the car to deviate from its lane, requiring drivers to correct their steering. The time taken to react (RT) served as the ground-truth indicator of drowsiness. The authors designed a Deep Q-Network tailored for EEG data, featuring a recurrent convolutional neural network (RCNN) backbone to extract spatial and temporal features. The reinforcement learning framework defined the state as segmented EEG data, the action as discrete adjustments to a predicted RT value, and the reward as the negative absolute difference between the predicted and measured RT. This approach allows the model to learn an optimal policy for estimating RT without requiring continuous labeled data during the testing phase. The results demonstrate that the trained DQN model effectively traces variations in the driver’s mental state, outperforming traditional supervised learning counterparts. The reinforcement learning approach proved superior for real-world applications because it maximizes data utilization by processing continuous session data rather than discarding unlabeled segments, a common limitation in supervised paradigms. The model successfully estimated RT with satisfactory accuracy, confirming the feasibility of using deep reinforcement learning for brain-computer interface (BCI) applications in driving safety. The significance of this work lies in its introduction of deep reinforcement learning to the field of EEG-based drowsiness detection, marking the first application of this paradigm in this specific BCI context. By decoupling the need for explicit labels during the inference stage, the method offers a more practical solution for online monitoring systems. The authors conclude that this approach not only improves upon existing supervised methods but also holds potential for generalization to other BCI scenarios where continuous, unlabeled data streams are prevalent.

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
• drowsiness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data
• Methodological Resource: tool software
• Theoretical Contribution: computational model