Synthetic Image Generation for Data Augmentation to Train an Unconscious Person Detection Network in a UAV Environment

Abstract

In this paper, we propose a data augmentation method using synthetic data generation for detecting an unconscious person in drone images. First, we extract the most salient and delicate foreground mask from a reference image that simulates an unconscious person situation using U² Net, which is a Salient Object Detection (SOD) model. Second, we apply shadow generation to the foreground mask for the natural appearance of the object. The unconscious person object generated by the foreground mask is synthesized with the background image of the Unmanned Aerial Vehicle (UAV) environment according to the altitude using object resizing. Therefore, we generate the most similar data to the image acquired by the drone. We verified the synthetic data-based image dataset using various object detection models, such as YOLOv4, YOLOv5, and EfficientDet. As a result, the Average Precision (AP) is higher than that of the real-world dataset. Our proposed method could be used to generate synthetic data for detecting an unconscious person and reducing the time cost and human resources needed for various tasks.