LMF-FD2M: Fatigue Driving Detection with Fusion of Millimeter-Wave Radar and Physiological Bracelet

Fatigued driving is an important cause of traffic accidents, and the current fatigued driving detection methods based on a single modality have poor adaptability, weak anti-interference and difficulty in balancing the requirements of accuracy and real-time performance. To solve these limitations, this paper proposes a lightweight fatigue driving detection model, LMF-FD2M, based on Multi-sensor Fusion, which uses multi-dimensional physiological signals from millimeter-wave radar and a physiological bracelet. LMF-FD2M adopts an asymmetric two-branch architecture, where the Mamba branch is designed to capture long-term global patterns. In particular, a Fourier Analysis Module with Attention (FAMA) is designed for this branch in order to capture global periodic patterns and effective characteristics of physiological signals in the frequency domain. The CNN-BiLSTM branch acts as a high-precision detector for short-term dynamic changes. Specially, a Focused Linear Attention (FLA) is introduced for this branch in order to focus on local instantaneous variations in temporal domain and suppress noises. Moreover, an adaptive fusion method based on sample characteristics is designed to dynamically modulate the contributions of each modality and integrate the global pattern and local dynamics features. The experimental results on two public datasets (FatigueSet, DROZY) and a self-built dataset (M2FDD) show that our LMF-FD2M model is superior to other state-of-the-art models, and the numbers of parameters and FLOPs are only 7–28% of the baselines, which can meet the needs of in-vehicle deployment.