An Extensive Analysis of Different Approaches to Driver Gaze Classification

Driver Monitoring Systems (DMS) enable Intelligent Vehicles to capture the in-cabin scene and help determine the driver’s level of attention and ability to take over. The task of driver gaze classification is the most important proxy for determining driver attention for DMS. In recent years, different approaches for driver gaze classification have been proposed. However, results and comparisons are barely valid. Different metrics are presented and datasets are kept private and are often collected under constraints that do not reflect realistic driving behavior. This work aims to provide an in-depth discussion and comparison of existing methods for driver gaze classification based on a dataset that is elaborately collected and constitutes realistic driving from real customers under no supervision. In particular, we evaluate the approaches with means of a nested leave-one-driver-out cross-validation on 20 different drivers. Moreover, we analyze the impact of the number of drivers in the training dataset on the generalization ability for unseen drivers and introduce a new error-based metric that allows us to assess how well a model is trained. Observations are that for end-to-end approaches, misclassifications between regions far apart occur more often and for all drivers, whereas the feature-engineered approach appears better qualified to build gaze estimators with limited data.