Lightweight Deep Learning Framework for Multi-Class Aerial Rescue Object Detection Using Enhanced YOLO
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Rapid victim localization is one of the most critical components of Search and Rescue (SAR) operations following natural disasters such as floods, wildfires, earthquakes, and building collapses. Although Unmanned Aerial Vehicles (UAVs) integrated with deep learning-based object detection have demonstrated promising results, existing systems often struggle with small-scale human detection, environmental occlusions, and real-time onboard processing limitations. Furthermore, many approaches rely solely on single-frame RGB imagery, limiting robustness in complex disaster scenarios. This paper presents a lightweight multi-class aerial rescue detection framework capable of simultaneously identifying humans, fire regions, and emergency vehicles from UAV imagery using bounding-box localization. The proposed system enhances a YOLO-based architecture through small-object feature refinement and sequential frame consistency analysis to improve detection accuracy in high-altitude and low-resolution conditions. The model is optimized for deployment on edge-computing platforms, enabling real-time inference during live rescue missions without dependence on high-end ground stations. Experimental evaluation on diverse aerial datasets demonstrates improved precision, recall, and detection stability compared to conventional single-frame detection methods, particularly in challenging environments with smoke, debris, and partial occlusion. The proposed approach advances toward a practical and autonomous UAV-assisted SAR system that supports rapid situational awareness and decision-making for emergency response teams.
Fields of Study & Subjects
Advanced Neural Network Applications
Fire Detection and Safety Systems
Video Surveillance and Tracking Methods
Publication details
- Journal / source
- Research Digest on Engineering Management and Social Innovations
- Pages
- 50–50
- Publication date
- 31 March 2026
- Language
- EN