AI-Based Object Detection in Shared Mobility: A Synthetic Data Approach for Rare Event Recognition

Detecting non-empty bicycle baskets in shared mobility services is a low-frequency yet operationally important task for maintenance efficiency. However, the rarity of such events makes it difficult to secure sufficient training data, while conventional manual inspection imposes a substantial operational burden. To address data scarcity and class imbalance, this study proposes an AI-based basket state monitoring framework with a crop-based synthetic data generation pipeline. The proposed method first detects basket regions from scene images using YOLOv8n and then edits cropped basket ROIs, rather than full images, to generate synthetic non-empty samples. This approach reduces structural distortion and improves the reliability of training data under limited real-data conditions. The same basket-centered ROI workflow is also applied to basket status recognition, where MobileNetV3-small is used as the classification model. Experimental results showed that all settings using synthetic data outperformed the real-only baseline. Specifically, the real-only setting achieved a Recall of 0.235 and an F1-score of 0.339, whereas the proposed framework improved Recall to 0.797 and F1-score to 0.788. These results suggest that the proposed system improves monitoring reliability and may help reduce operational burden in real-world settings.