Cross-Camera Distracted Driver Classification through Feature Disentanglement and Contrastive Learning

2024-11-20

arXiv.org

AI Summary

To address poor generalization in driver distraction classification across cameras caused by viewpoint variation, this paper proposes DBMNet—the first framework that jointly models viewpoint-invariant feature disentanglement and action-aware contrastive learning. DBMNet employs a differentiable disentanglement module to explicitly separate viewpoint-invariant representations from action semantics, and introduces a self-supervised contrastive loss coupled with multi-domain feature alignment. Built upon a lightweight backbone, the model achieves enhanced robustness across cameras, time periods (day/night), and datasets. On the 100-Driver day-night subset, DBMNet improves Top-1 accuracy by 9% over the state of the art. It also achieves superior performance on cross-domain evaluations across three major benchmarks—AUCDD-V1, EZZ2021, and SFD—demonstrating its strong domain generalization capability.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Multi-instance/Multi-view Learning
• Application Domains: Mobility, Driving & Flight

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

The classification of distracted drivers is pivotal for ensuring safe driving. Previous studies demonstrated the effectiveness of neural networks in automatically predicting driver distraction, fatigue, and potential hazards. However, recent research has uncovered a significant loss of accuracy in these models when applied to samples acquired under conditions that differ from the training data. In this paper, we introduce a robust model designed to withstand changes in camera position within the vehicle. Our Driver Behavior Monitoring Network (DBMNet) relies on a lightweight backbone and integrates a disentanglement module to discard camera view information from features, coupled with contrastive learning to enhance the encoding of various driver actions. Experiments conducted on the daytime and nighttime subsets of the 100-Driver dataset validate the effectiveness of our approach with an increment on average of 9% in Top-1 accuracy in comparison with the state of the art. In addition, cross-dataset and cross-camera experiments conducted on three benchmark datasets, namely AUCDD-V1, EZZ2021 and SFD, demonstrate the superior generalization capability of the proposed method.

Problem

• Improving accuracy in cross-camera distracted driver classification
• Enhancing model generalization across varying camera positions
• Reducing computational costs while maintaining high performance

Innovation

• Lightweight backbone for efficient processing
• Feature disentanglement to remove camera view bias
• Contrastive learning enhances driver action encoding