Lerch, David (Fraunhofer IOSB), Bhardwaj, Suraj (University of Siegen), Martin, Manuel (Fraunhofer IOSB), Diederichs, Frederik (Fraunhofer IOSB), Stiefelhagen, Rainer (Karlsruhe Institute of Technology)

Self-Supervised Driver Distraction Detection for Imbalanced Datasets

Scheduled for presentation during the Regular Session "S36a-Behavior Modeling and Decision-Making in Traffic Systems" (FR-LM-T36), Friday, November 21, 2025, 11:50−12:10, Surfers Paradise 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

The role of driver distraction in the occurrence of road traffic accidents is of pivotal importance on a global scale. Furthermore, the effective detection of distraction is com- plicated by the imbalanced nature of typical driving datasets, which often lack sufficient observations of certain behaviors. This study examines the task of driver distraction detection using the State Farm Distracted Driver Detection and the Drive&Act datasets. The primary objective is to address the issue of data imbalance in a label-free manner in an unlabeled training dataset. To address the issue of data imbalance, we propose a novel data-loading technique, Clustered Feature Weighting (CFW). This label-free approach integrates trans- fer learning, unsupervised clustering, and weighted random sampling. The proposed method improves class balance during model training, thereby enhancing overall performance and robustness. The methodology encompasses model selection and embedding extraction, variance analysis, clustering, and weight generation. Experiments are conducted using both supervised and self-supervised learning (SSL) methods. The results demon- strate the effectiveness of SSL methods in enhancing the adapt- ability and robustness of driver distraction detection systems, particularly in handling non-uniform class distributions. CFW significantly improves class balance within training batches, thereby enhancing the accuracy and generalization of driver distraction detection models. Our code will be published on GitHub.