객체 기반 데이터 증강을 통한 교통 표지 인식 성능 향상

초록

Accurate traffic sign recognition is a core component of autonomous driving and Advanced Driver Assistance Systems (ADAS). However, recognition performance inherently degrades in real-world road environments due to various factors, including dynamic lighting, adverse weather conditions, and image distortions. Furthermore, collecting large-scale datasets that encompass all such environmental variables incurs prohibitive costs and time. While conventional image processing-based data augmentation techniques have been utilized to address this issue, they often fail to adequately reflect the complexities of actual road environments. Therefore, this paper proposes an object-based synthetic data augmentation technique that focuses specifically on the traffic sign object itself, combining augmentations that simulate environmental changes with those that simulate image distortions. By integrating the synthetic data generated through this proposed method with real-world data for model training, key performance metrics showed significant improvement over traditional simple augmentation approaches. The results of this study demonstrate that object-based data augmentation can effectively enhance the generalization performance of deep learning models, even when limited real-world data is available.

키워드