Entropy-guided robust feature domain adaptation for electroencephalogram-based cross-dataset drowsiness recognition

Abstract

To accurately identify driver drowsiness using Electroencephalogram (EEG), a significant amount of time is typically required to gather data from either the same subject or multiple subjects under the same experimental condition. The prolonged period required for data collection impedes the practical application of EEG. In this paper, we consider the challenging task of cross-dataset driver drowsiness recognition, where the EEG data collected by different devices are expected to be conveniently recognized by a classifier trained on an existing dataset. In order to solve the problem of distribution drift, we propose an Entropy-Guided Robust Feature (EGRF) adaptation framework. This framework utilizes the Entropy Minimization (EM) technique for domain alignment and an ensemble-like structure, consisting of multiple branch classifiers and a dropout layer, to enhance the robustness. The proposed method has been tested across two public datasets and achieves 2-class recognition accuracies of 77.4% and 85.9%. The innovative feature-robust block, combined with the EM loss, leads to substantial performance improvement, proving the effectiveness of our framework for cross-dataset driver drowsiness recognition. Our research demonstrates a promising direction for development of a calibration-free driver drowsiness monitoring system in the future.