User-adaptive Eyelid Aperture Estimation for Blink Detection in Driver Monitoring Systems
Authors: Luis Salgado
Date: 19.05.2020
Abstract
This paper presents a new method for eyelid aperture estimation, suitable to be used in Driver Monitoring Systems (DMS) to measure blink patterns such as microsleeps and any other metric that assess the fatigue level of the driver. The method has been designed to work real-time and in continuous operation, by introducing a novel online Exponential Weighted Moving Average (EWMA)-based Bayesian estimation process, which ensures dynamic adaptability to drivers with different physiognomy features, and also to changes due to physiological states (e.g. drowsiness). Our method has been implemented in the framework of a DMS, to take advantage of existing facial landmark detection and tracking mechanisms, and to provide real-time functionality for driving platforms (such as the NVIDIA Drive PX 2). The method is evaluated against a large labelled dataset, and compared to baseline and previous existing methods, showing an excellent balance between adaptability, performance, and robustness.
BIB_text
title = {User-adaptive Eyelid Aperture Estimation for Blink Detection in Driver Monitoring Systems},
pages = {342-352},
keywds = {
Eyelid Aperture, Blink Detection, Driver Monitoring, Computer Vision, ADAS
}
abstract = {
This paper presents a new method for eyelid aperture estimation, suitable to be used in Driver Monitoring Systems (DMS) to measure blink patterns such as microsleeps and any other metric that assess the fatigue level of the driver. The method has been designed to work real-time and in continuous operation, by introducing a novel online Exponential Weighted Moving Average (EWMA)-based Bayesian estimation process, which ensures dynamic adaptability to drivers with different physiognomy features, and also to changes due to physiological states (e.g. drowsiness). Our method has been implemented in the framework of a DMS, to take advantage of existing facial landmark detection and tracking mechanisms, and to provide real-time functionality for driving platforms (such as the NVIDIA Drive PX 2). The method is evaluated against a large labelled dataset, and compared to baseline and previous existing methods, showing an excellent balance between adaptability, performance, and robustness.
}
isbn = {978-989-758-419-0},
date = {2020-05-19},
}
