Abstract
Driving Under the Influence (DUI) has emerged as a significant threat to public safety in recent years. Despite substantial efforts to effectively detect DUI, the inherent risks associated with acquiring DUI-related data pose challenges in meeting the data requirements for training. To address this issue, we propose DUIncoder, which is an unsupervised framework designed to learn exclusively from normal driving data across diverse scenarios to detect DUI behaviors and provide explanatory insights. DUIncoder aims to address the challenge of collecting DUI data by leveraging diverse normal driving data, which can be readily and continuously obtained from daily driving. Experiments on simulator data show that DUIncoder achieves detection performance superior to that of supervised learning methods which require additional DUI data. Moreover, its generalization capabilities and adaptability to incremental data demonstrate its potential for enhanced real-world applicability.