A Comparative Analysis of Revealing Temporal Patterns for Driver Drowsiness Detection

The primary cause for road accidents is indicated as the unawareness state of the driver which is the result of being drowsy or distracted. Therefore, it is required for the researchers to design systems that detect drowsiness from camera and alert the driver. There are different deep learning architectures that learn the hidden relations between patterns according to their approaches for processing input data. In this paper, the performances of adopting spatial + temporal and spatio-temporal architectures are investigated through detecting driver drowsiness. For spatial + temporal approach, 2 dimensional ResNet-34 is used along with bidirectional LSTM while spatio-temporal architecture consists of 3 dimensional ResNet-34 inflated from 2 dimensional version followed by fully connected layers. The experiments are performed on NTHU Driver Drowsiness Detection dataset and performance of capturing temporal relations is analyzed for both cases. Experiments show that spatial + temporal approach is superior than spatio-temporal approach in terms of both accuracy and speed.