Hierarchical HMMs on Eyes for Driver Drowsiness Detection

Driving with fatigue is one of the main factors of traffic accident, so that timely driver drowsiness detection is very crucial. This paper proposes a novel drowsiness detection method by constructing Hierarchical Hidden Markov Models(HHMMs) on eyes sequences. The method includes two key parts: (1) take Faster-RCNN as backbone to train a robust and efficient eyes extractor by imposing additional constraints. (2) build a two-layer HHMMs to fully exploit temporal context information on different levels. We can calculate the likelihood of drowsy and nondrowsy separately for eye state sequences under two trained HHMMs to determine whether the driver is sleepy or not. In general, our approach achieves 92.98% accuracy and 25 fps speed on the evaluation set of the public NTHU-DDD Dataset. The experimental results show that the proposed method outperforms the existing driver drowsiness detection algorithms.