PDD: Post-Disaster Dataset for Human Detection and Performance Evaluation

Abstract

Human detection is aimed at automatically labeling specific semantic objects in high-resolution images, which is a key problem in the post-disaster search and rescue (SAR) mission with unmanned aerial vehicles (UAVs). However, for large-scale and efficient search and statistics with UAVs, there is a lack of real-world image datasets of post-disaster ruins, resulting in poor performance in practical applications. In this article, we conducted a comprehensive review of current studies and created a post-disaster dataset (PDD) for human detection first, which contains the real-world post-disaster UAV images from multiple scenes, perspectives, and distances. Second, based on the YOLOv5 algorithm, we verified the validity and superiority of PDD by comparing and analyzing its detection performance with other datasets and on different training proportions. Finally, we compared and evaluated the detection performance of 11 state-of-the-art algorithms on PDD, including faster region-based convolutional neural networks (R-CNN), YOLOv5, YOLOv7, YOLOv8, improved YOLOv5 (im-YOLOv5), neural architecture search (YOLO-NAS), detection transformer (DETR), deformable DETR (DDETR), dynamic anchor box DETR (DAB-DETR), denoising DETR (DN-DETR), and DETR with improved denoising anchor box (DINO), and provided a performance analysis of different deep learning algorithms. The results demonstrate that you only look once (YOLO)-based algorithms have a better real-time statistical performance, while the DETR-based algorithms have more accurate box prediction capabilities. The PDD, codes, and models are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/HaoqianSong/Post-Disaster-Dataset</uri> .

Topics & Concepts