A Multi-view Driver Drowsiness Detection Method Using Transfer Learning and Population-based Sampling Strategy

Driver drowsiness is an important factor in traffic safety. Thus, many researchers endeavor to develop a reliable driver drowsiness detection system. However, the large variation of the relative driver face-camera pose and the lack of large-scale public drowsiness detection dataset have brought great challenges to train a generic drowsiness detection model. This paper proposes a new network to accurately detect driver drowsiness from various viewing angles based on transfer learning and population-based sampling strategy (TLPSN). Firstly, the multitask cascaded convolutional networks (MTCNN) is chosen to detect the driver face. Then, with a transfer learning network based on the O-Net model in MTCNN (ONET-T), the driver face image cropped by MTCNN is taken as the input to inference the driver drowsiness state. The effectiveness of O-Net model is verified compared with other traditional pre-trained models. A population-based sampling strategy is adopted to generate the training set from a new dataset collected through a driver-in-the-loop platform. The results show that the proposed method has strong robustness to the variation of pose while ensuring high accuracy. Besides, transfer learning can significantly promote the generalization ability of the model.