SelectAug: A Data Augmentation Method for Distracted Driving Detection

Yuan Li,Wei Mi,Jingguo Ge,Jingyuan Hu,Hui Li,Daoqing Zhang,Tong Li

DOI: https://doi.org/10.1007/978-3-031-05936-0_32

2022-01-01

Abstract:With distracted driving becoming one of the main causes of traffic accidents, deep learning technology has been widely used in distracted driving detection, which achieves high accuracy when the training and test data are identically distributed. However, this assumption cannot correspond to the real-world situation. In case of a small sample size, we usually utilize open datasets as training dataset. Thus the training data distribution and test data distribution are different, which may induce accuracy plummets. Concentrating on the unforeseen data shifts encountered under different data distributions in distracted driving detection application, it is extremely desired to develop the detection technique with high robustness. In order to alleviate the issue about data shifts encountered under different data distributions, we propose an innovative method, SelectAug, to enhance images by applying the selected important features of the images. The experimental evaluations on the StateFarm dataset show that our method outperforms prior methods, demonstrating its efficacy in detecting distracted driving behaviors scenes. Furthermore, our method also improves generalization performance under different data distributions for distracted driving detection, which allows open datasets to be applied to real-world scenarios.