FMCW Radar-Based Drowsiness Detection With a Convolutional Adaptive Pooling Attention Gated-Recurrent-Unit Network

Abstract

The state of drowsiness significantly affects work efficiency and productivity, increasing the risk of accidents and mishaps. Radar-based detection technology offers significant advantages in drowsiness detection, providing a noninvasive and reliable method based on vital sign tracking and physiological feature extraction. However, the classification of sleepiness levels is often simple and the detection accuracy is limited. This study proposes a frequency-modulated continuous-wave (FMCW) radar-based system with a convolutional adaptive pooling attention gated-recurrent-unit (CAPA-GRU) network to enhance detection accuracy and precisely determine levels of radar-based drowsiness detection. First, an FMCW radar is used to obtain breathing and heartbeat signals, and the radar signals are processed through the wavelet transform method to obtain highly accurate physiological characteristics. Then, the vital sign signals are analyzed both in the time and frequency domains, and the optimal input data is obtained by combining the characteristic data. Also, the CAPA-GRU, comprising a convolutional neural network (CNN), a gated-recurrent-unit (GRU), and a convolutional adaptive average pooling (CAA) module, is proposed for drowsiness classification and monitoring. The experimental results show that the proposed method achieves multistage sleepiness detection based on FMCW radar and achieves excellent results in low classification. The proposed network has excellent performance and certain robustness. Experiments conducted with cross-validation on a self-collected dataset show that the proposed method achieved 90.11% accuracy in binary classification, 80.50% accuracy in ternary classification, and 58.17% accuracy in quinary classification and the study also used a public data set for sleepiness detection, and the detection accuracy reached 97.34%.