Eye Detection For Drivers Using Convolutional Neural Networks With Automatically Generated Ground Truth Data

Abstract

Eye detection is an essential feature for driver monitoring systems acting as a base functionality for other algorithms like attention or drowsiness detection. Multiple methods for eye detection exist. The machine learning based methods involve a manual labeling process in order to generate training and testing datasets. This paper presents an eye detection algorithm based on convolutional neural networks trained using automatically generated ground truth data and proves that we can train very good machine learning models using automatically generated labels. Such approach reduces the effort needed for manual labeling and data preprocessing.