A Method for Accurate Driver Status Monitoring using Domain Adaptation after Deployment

Abstract

With rapid advances in AI technology, autonomous driving is close to becoming a reality. Nevertheless, most car accidents are still caused by the driver's forward-looking negligence, and driver's intervention is still required. Therefore, monitoring the driver status has become an essential task for preventing car accidents. Many studies have attempted to solve the concern by applying a pre-trained neural network. However, the performance of a pre-trained neural network is deteriorated due to the distributional shift between training data and field data. In this paper, we propose a method to retain the performance of the pre-trained neural network by mitigating the detrimental effect of the distributional shift. In addition, we will show that the proposed method can be implemented on an embedded platform where memory size and computing power are limited.