Driver Drowsiness Estimation by Parallel Linked Time-Domain CNN with Novel Temporal Measures on Eye States.

This paper presents a vision-based driver drowsiness estimation system from sequences of driver images. We propose a stage-by-stage system instead of an end-to-end system for driver drowsiness estimation. The stage-by-stage system (1) calculates features related to eyes on a frame-by-frame basis, (2) calculates temporal measures on eye states, and (3) estimates drowsiness levels by time-domain convolution with a parallel linked structure. Furthermore, we propose average eye closed time (AECT) and soft percentage of eyelid closure (Soft PERCLOS) as novel temporal measures on eye states to extract information related to driver drowsiness. Extensive experiments have been conducted on a driving movie dataset recorded in a real car. Our system achieves a high accuracy of 95.86% and mean absolute error (MAE) of 0.4007 on the dataset.