Driver anomaly quantification for intelligent vehicles: a contrastive learning approach with representation clustering

Abstract

Driver anomaly quantification is a fundamental capability to support human-centric driving systems of intelligent vehicles. Existing studies usually treat it as a classification task and obtain discrete levels for abnormalities. Meanwhile, the existing data-driven approaches depend on the quality of dataset and provide limited recognition capability for unknown activities. To overcome these challenges, this paper proposes a contrastive learning approach with the aim of building a model that can quantify driver anomalies with a continuous variable. In addition, a novel clustering supervised contrastive loss is proposed to optimize the distribution of the extracted representation vectors to improve the model performance. Compared with the typical contrastive loss, the proposed loss can better cluster normal representations while separating abnormal ones. The abnormality of driver activity can be quantified by calculating the distance to a set of representations of normal activities rather than being produced as the direct output of the model. The experiment results with datasets under different modes demonstrate that the proposed approach is more accurate and robust than existing ones in terms of recognition and quantification of unknown abnormal activities.