SAFE-Net: Scale-Aware Feature Enhancement for Aerial Person Detection in Flood Disaster Imagery

Arun Kumar S, Komuravelli Prashanth, JANIPIREDDY GANESH MOULI, Gorthi Rama Krishna Sai Subrahmanyam

Track: Proceedings Track

Keywords: Person detection, UAV, Attention

TL;DR: Detecting stranded persons during floods from UAVs with a lightweight attention model

Abstract: Detecting stranded persons from Unmanned Aerial Vehicles (UAVs) is crucial for autonomous search and rescue operations during flood disasters. However, reliable detection in aerial flood imagery remains challenging due to extreme scale variations, strong water reflections and motion blur. In many cases, human targets occupy less than 0.1% of the image area, making them difficult to detect using conventional object detectors. To address these challenges, we propose SAFE-Net (Scale-Aware Feature Enhancement Network), a lightweight detection framework which builds upon the YOLOv8 architecture by replacing standard C2f blocks with a Scale-Aware Feature Enhancement (SAFE) module. The SAFE module improves the representation of tiny human targets through two mechanisms: scale-aware spatial weighting to emphasize extremely small objects, and texture enhancement using depthwise convolutions to recover fine edge information while suppressing noise from water surfaces. We also introduce UAV-based Flood Survivor Detection Dataset (UAV-SURV), a dataset of 6,122 annotated aerial images collected from real flood monitoring videos. Experimental results show that SAFE-Net improves detection accuracy by 10.9% in mAP@0.5:0.95 over YOLOv8n while reducing model parameters by 33.8\%, demonstrating an effective and lightweight solution for UAV-assisted disaster response.

Submission Number: 5