Filter-Pruned 3D Convolutional Neural Network for Drowsiness Detection

Human drowsiness while operating motor vehicles or heavy machinery can have potentially lethal consequences for the operator and others in their immediate vicinity. In this study, we developed a visual-based drowsiness detection system that can analyze videos and make predictions on attention status. A 3D convolutional neural network (CNN) was built for spatio-temporal feature extraction in consecutive frames, and temporal smoothing was used to remove noisy predictions. As a part of an assistance system, a real-time, lightweight and computationally-efficient system is preferable. Thus, we proposed a Scale Module that can be easily integrated into the convolutional layer and estimate the importance of filters. Our results show that scale values calculated from the Scale Module are good indicators for filter pruning, and that filters with small scale values can be removed with negligible loss in the model's performance.