Multi-Level Drowsiness Detection Based on Deep Feature Fusion of Eye and Head Pose

Abstract

Drowsiness detection is a significant problem, most existing non-intrusive methods estimate drowsiness only by single images, without leveraging the temporal information available in the frame sequence. The lack of temporal information leads to the inability of drowsiness detection to indicate consecutive behaviors. To this end, we present a drowsiness detection method, which takes into account both eye and head pose deep feature representation by conducting feature fusion. Then, the fused feature is fed into the LSTM (Long Short-Term Memory) network to enhance the accuracy of the drowsiness detection model through temporal information. The experimental results on the NHTU-DDD dataset and the self-constructed dataset show that the proposed method outperforms six existing advanced approaches.