Authors:
Hélène Unrein
1
;
Benjamin Chateau
2
and
Jean-Marc André
1
Affiliations:
1
IMS – Cognitique UMR 5218, ENSC-Bordeaux INP, Université de Bordeaux, Talence, France
;
2
Centre Aquitain des Technologies de l’Informations et Electroniques, Talence, France
Keyword(s):
Cognitive States, Engagement in the Driving Task, Hypovigilance, Cognitive Fatigue, Eye-tracking, CARt.
Abstract:
This study aims to validate a construction methodology of a device able to estimate the cognitive state of an operator in real time. The SUaaVE project (SUpporting acceptance of automated VEhicle) studies the integration of an intelligent assistant in a level 4 autonomous car. The aim of our work is to model the cognitive state of the driver in real time and for all situations. The cognitive state is a natural state that alters or preserves the operator's ability to process information and to act. Based on a literature review we identified the cognitive functions used by the driver and the factors influencing them. Different cognitive components emerged from this synthesis: engagement (Witmer & Singer, 1998), fatigue (Marcora and al. 2009) and vigilance (Picot, 2009). Eye-tracking is a technique used to determine the orientation of the gaze in a visual scene. According to the literature the general dynamics of a visual behavior is characterized by metrics: number of fixations, durati
on of fixation, gaze dispersion... These dynamics are altered unconsciously due to fatigue (Faber, Maurits, & Lorist, 2012) or hypovigilance (De Gennaro et al., 2000, Bodala et al., 2016); and consciously due to engagement in driving (Freydier et al., 2014; Neboit, 1982). We carry out a phase of experimentation in a naturalistic situation (driving simulator) in order to collect data for each cognitive state. Realistic scenarios are constructed to induce cognitive states. The model’s estimation is compared to the real cognitive state of the driver measured by behavioral monitoring (eye-tracking). The model is a CARt (Breiman & Ihaka, 1984) decision tree: Classification And Regression Trees. The CARt aims at building a predictor. The interest is to facilitate the design of the tool as well as its future implementation in real time. We illustrate the construction methodology with an example the results obtained.
(More)