Calibration-Free EEG-based Driver Drowsiness Detection with Online Test-Time Adaptation

TL;DR

This work tackles EEG-based driver drowsiness detection under non-i.i.d. conditions caused by inter-subject variability, removing the need for calibration data. It introduces an online test-time adaptation framework that freezes pretrained BN statistics while updating affine BN parameters, employs a streaming memory bank with energy-based sample removal, and augments with prototype learning for robust predictions. The method operates on EEG segments from a simulated sustained-attention driving dataset and achieves a mean F1-score of 81.73%, beating the best TTA baselines by about 11 percentage points. These results demonstrate strong calibration-free adaptability to evolving EEG distributions in real-time driving contexts.

Abstract

Drowsy driving is a growing cause of traffic accidents, prompting recent exploration of electroencephalography (EEG)-based drowsiness detection systems. However, the inherent variability of EEG signals due to psychological and physical factors necessitates a cumbersome calibration process. In particular, the inter-subject variability of EEG signals leads to a domain shift problem, which makes it challenging to generalize drowsiness detection models to unseen target subjects. To address these issues, we propose a novel driver drowsiness detection framework that leverages online test-time adaptation (TTA) methods to dynamically adjust to target subject distributions. Our proposed method updates the learnable parameters in batch normalization (BN) layers, while preserving pretrained normalization statistics, resulting in a modified configuration that ensures effective adaptation during test time. We incorporate a memory bank that dynamically manages streaming EEG segments, selecting samples based on their reliability determined by negative energy scores and persistence time. In addition, we introduce prototype learning to ensure robust predictions against distribution shifts over time. We validated our method on the sustained-attention driving dataset collected in a simulated environment, where drowsiness was estimated from delayed reaction times during monotonous lane-keeping tasks. Our experiments show that our method outperforms all baselines, achieving an average F1-score of 81.73\%, an improvement of 11.73\% over the best TTA baseline. This demonstrates that our proposed method significantly enhances the adaptability of EEG-based drowsiness detection systems in non-i.i.d. scenarios.

Figures (5)

• Figure 1: Comparison of the transfer learning algorithms. DA utilizes both labeled source and unlabeled target domain data during training. Subsequently, DA makes a prediction with the pretrained model. DG uses only labeled source domain data during training and predicts with a pretrained model on unseen domains. TTA learns labeled source domain data during training, then dynamically adapts the pretrained model with input target domain data. This encourages effective predictions during inference with the adjusted model.
• Figure 2: Overview of the proposed test-time adaptation framework. Incoming EEG segments from the target subject are filtered by entropy and stored in a memory bank based on energy score and persistence. The memory bank is used to update only the learnable parameters of the BN layers via entropy minimization with energy-based regularization. Class prototypes are continuously updated from the memory to support robust prediction under distribution shifts.
• Figure 3: Experimental paradigms of the sustained-attention driving dataset were conducted in the following sequence for each trial: response offset, deviation onset, and response onset. Drowsy drivers exhibited longer RTs compared to alert drivers.
• Figure 4: UMAP feature visualization from target subjects. The row and column indicate the target subjects and adaptation methods, respectively. The first column shows the features from the ERM (no adaptation) model. The second and third columns show the features when adapting the model using existing BN configurations. The fourth and fifth columns show the features when adapting the model using TTA methods. The last column shows the features of our proposed method. We show that the proposed method significantly enhances performance and disentanglement.
• Figure 5: Performance comparison when applying our BN configuration to the existing TTA baselines. The gray horizontal line denotes the average F1-score of the source pretrained EEGNet8,2 without adaptation methods. Existing baselines with their original BN configuration are presented on the left, while the baselines with our BN configuration are shown on the right. The overall performance of the existing TTA baselines significantly improves after manipulating the BN layers.