Threat-Awareness Recognition based on Driver’s Eye Gaze Data

Abstract

This master thesis project focuses on identifying and assessing drivers’ threat awareness in a rear-end collision scenario based on the driver’s gaze direction from a single camera driver monitoring system. The purpose of the assessment is to further be used to enable collision avoidance systems to be adaptive to the driver’s visual attention of threats. A post image processing estimation on the gaze direction with uncertainty, is made by using a developed Kalman filter that have different states depending on eye-movement characteristics, which is enabled by a velocity based fixation-identification algorithm. Different gaze features are developed that represent the driver being attentive of the lead-vehicle. These features are found by mapping the driver’s gaze direction to the lead-vehicle. A naturalistic driving field operational test is used to find scenarios where there is a risk for a rear-end collision. The hypothesis in this project is that for a driver to be aware of a potential rear-end collision, the gaze needs to be directed towards the lead-vehicle. In all the scenarios found in the field test, the driver’s visual attention was directed towards the lead-vehicle before the driver responded by braking. These statistics, supported by findings in the literature study, gave confidence that it is possible to be able to classify the driver as unaware if the visual attention is directed elsewhere than the lead-vehicle. However, fully assessing the driver being aware by only using the gaze-direction measurements should be done with caution given the scope of this thesis, as there can be other cognitive factors distracting the driver from perceiving a threat. Based on the annotations of the videos from scenarios found in the field test, and the calculated gaze-features, a logistic regression model is used to assess the driver’s visual attention of the lead-vehicle. This method is chosen as it is an interpretable and transparent binary outcome classification algorithm. The results show that the model is able classify the visual attention with a true positive rate of 96% and a false negative rate of 75%.