| Title |
Low-light Data Augmentation-based Object Detection for Nighttime Smart Yard Safety Management |
| Authors |
고영민(Yeongmin Ko) ; 박진선(Jinsun Park) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.11.95 |
| Keywords |
Data augmentation; Object detection; Low-light image enhancement; Smart yard |
| Abstract |
In 24/7 Smart Yard environments, preventing safety accidents in low-light conditions such as nighttime is of paramount importance. However, object detection models used in conventional intelligent safety monitoring systems are predominantly trained on daylight data, leading to a significant degradation in detection performance for key safety objects like helmets, workers, and heavy equipment at night. This issue is exacerbated by the difficulty of acquiring sufficient low-light data from actual industrial sites. To address this problem, we propose a dual approach that combines data augmentation at the training stage with input data optimization at the inference stage. In the training stage, data generated using a diffusion model secures the model's generalization performance, while in the inference stage, real-time low-light enhancement pre-processing maximizes detection performance. In experiments conducted on the YOLOv11 model, the proposed method showed a 13%p improvement in overall object detection performance of mAP@50, from 0.35 to 0.48, compared to the baseline model trained only on the original dataset. |