Lightweight Multispectral Detection and DEM-Constrained Ray Consistency Localization for UAV-Assisted Search and Rescue

Abstract

Reliable target detection and geographic localization are critical for unmanned aerial vehicle (UAV)-assisted search and rescue (SAR) yet remain challenging in complex outdoor environments. Small targets in UAV Red–Green–Blue–Infrared (RGB–IR) imagery suffer from background clutter, occlusion, low illumination, and infrared thermal diffusion, while localization is vulnerable to unstable viewpoints and terrain-induced ray uncertainty. This study presents an integrated UAV-SAR framework coupling lightweight multispectral detection with Digital Elevation Model (DEM)-constrained geographic localization. For detection, the Asymmetric Fusion and Context-aware Detection (AFC-Det) network leverages asymmetric dual-stream encoding, cross-modal mutual prompting, and high-resolution anchored aggregation to enhance small-target representation from RGB–IR pairs. For localization, the Global Context-Regularized Huber Ray Consistency Optimization (GCR-HRCO) improves geolocation via global ray aggregation, multi-ray geometric consistency, Huber robust optimization, and DEM-based terrain constraints. Experimental results demonstrate AFC-Det achieves 45.4% average precision (AP) and 44.7% AP for small objects (APs) on the VTSaR dataset, with 1.7 million parameters, 8.0 GFLOPs, and 107.2 FPS, generalizing well to M3FD (54.6% AP). On SAR-DAG_raycast, GCR-HRCO reduces mean horizontal error from 6.85 m to 3.31 m and RMSE from 8.16 m to 4.33 m. Collectively, these results demonstrate the effectiveness of the proposed detection and localization components.