A Lightweight UAV-Based SAR System for Human Detection and Monocular Geolocation

Abstract

Timely localization of missing persons remains a critical challenge in search-and-rescue (SAR) operations, where manual analysis of UAV imagery limits response speed and operational efficiency. This paper presents a lightweight end-to-end UAV-based SAR pipeline integrating real-time human detection, monocular GPS geolocation, and automated alerting into a deployable embedded system. Human detection is performed using a YOLOv8n model trained through a two-stage transfer learning strategy using the VisDrone and HERIDAL datasets. Detected persons are geolocated using a lightweight geometric algorithm that estimates ground coordinates directly from standard UAV telemetry without requiring additional sensors.

The proposed system is evaluated on a custom annotated dataset of 300 UAV frames collected under four controlled flight conditions varying in altitude and camera angle. The final model achieves 0.921 precision, 0.926 recall, and 0.965 mAP@0.5, outperforming single-stage training baselines. The geolocation module achieves a mean localization error of 1.01 m across 60 independent measurements under controlled conditions. For real-time onboard deployment, the model is optimized using W8A16 post-training quantization and benchmarked on a Qualcomm RB3 Gen 2 NPU, achieving a median inference latency of 37.3 ms at 960×960 resolution while preserving detection accuracy.

The complete implementation, trained models, and evaluation dataset are publicly released to support reproducible research and future development in deployable UAV-based SAR systems.