Temporal and spatial feature based approaches in drowsiness detection using deep learning technique

Abstract

Drowsiness is a term which seems to be very simple but for a moment, it becomes a critical issue for many drivers and workers while they are performing their duty. Many people’s lives may collapse into trouble because of drowsiness. Therefore, such a real-time system is needed which can be easy to develop and configure for early as well as accurate drowsiness detection. As per requisite, we have adopted a large realistic dataset which includes 30 h video of 60 different participants in three classes, i.e. alert, low vigilant and drowsy. In our proposed work, we have selected the videos with extreme classes, i.e. alert and drowsy only. Further, we have designed two different models based on temporal and spatial feature by employing computer vision as well as deep-learning approach. In one model, temporal features are obtained by computer vision techniques followed by long short-term memory (LSTM) and the second model adopts spatial features extraction through convolution neural network (CNN) followed by LSTM. Although the temporal model is more complex and has less accuracy than spatial model, in spite of this, the study shows that the temporal model is far better in terms of training time than spatial model by establishing the comparison using confusion metrics and Area under Curve (AUC)–Receiver-Operating Characteristic Curve (ROC) score.