Мультиспектральні методи релокалізації БПЛА: аналіз сучасних підходів та перспективи розвитку

Abstract

This article provides a comprehensive systematic review of multispectral relocalization methodologies for Unmanned Aerial Vehicles (UAVs) operating in Degraded Visual Environments (DVE). Traditional Visual Simultaneous Localization and Mapping (VSLAM) algorithms, which rely exclusively on the visible electromagnetic spectrum (RGB sensors), experience severe performance degradation – often losing tracking accuracy by 40-60%—under conditions of dynamic illumination, nighttime operations, and aerosol interference such as fog or smoke. To mitigate these critical vulnerabilities, the research explores the paradigm of cross-modal Sensor Fusion, specifically the integration of standard optical data with long-wave infrared (LWIR) thermal signatures (8-14 µm). Because thermal emissions depend on the thermodynamic properties of materials rather than external lighting, they offer superior penetration through obscurants and maintain highly stable descriptors across day-night transitions. Based on a rigorous analysis of over 50 peer-reviewed publications from 2015 to 2024, the study categorizes current fusion architectures into three primary paradigms: Early Fusion (pixel-level concatenation utilizing Convolutional Neural Networks), Late Fusion (decision-level integration ensuring modular redundancy), and Hybrid Fusion (adaptive spatial alignment using Cyclic Fuse-and-Refine Blocks). The comparative quantitative analysis reveals that Hybrid Fusion achieves the most optimal balance, reducing positioning errors to 0.15-0.25 meters in urban environments and demonstrating a 35-40% performance advantage over monospectral systems under severe visibility constraints.