Transformer-Based Network for Robust 3D Industrial Environment Understanding in Autonomous UAV Systems

Autonomous navigation of unmanned aerial vehicles (UAVs) in unstructured industrial environments remains challenging due to irregular geometry, dynamic obstacles and sensor uncertainty. Classical Simultaneous Localization and Mapping (SLAM) systems, though geometrically consistent, often fail under poor initialization, textureless areas or reflective surfaces. To overcome these issues, this work proposes a hybrid transformer-geometric framework that fuses learned scene priors with keyframe-based SLAM. A TinyViT encoder and lightweight multi-task decoder jointly estimate inverse depth, surface normals and semantic segmentation, providing dense geometric and semantic cues that stabilize localization and mapping. These priors are incorporated into the SLAM optimization to enhance convergence, reject dynamic objects and improve relocalization. The system operates near real-time (~1 FPS) on a Raspberry Pi 5 CPU, suitable for keyframe-level inference. Experiments show robust localization and consistent mapping in cluttered, reflective and dynamic industrial scenes, confirming that transformer-based dense perception effectively complements classical SLAM for resource-efficient UAV navigation.