Thermal infrared scene reconstruction offers critical capabilities for applications ranging from energy monitoring to search and rescue, yet existing methods remain limited in handling dynamic drone imagery, modeling temporal temperature variation, and mitigating the inherent quality issues of thermal data. In this work, I propose a physics aware 3D Gaussian splatting pipeline for dynamic thermal reconstruction from aerial infrared imagery. The approach introduces three key components: a Viewpoint Aware Attenuation Estimator (VAE) to correct view dependent radiative loss, a Thermal Diffusion Correction Unit (TDCU) to restore fine object boundaries blurred by heat conduction, and a Surface Smoothness Regularizer (SSR) to enforce physically plausible temperature continuity. Preliminary results on the NTR-Guassian dataset show improved edge stability and reduced ghosting compared to baseline splatting methods. I further outline directions for enhancing temporal consistency and extending the framework toward real-time performance and environmental condition modeling.
End-to-End Infrared UAV Mapping: An Integrated Framework for Thermal Scene Reconstruction
By: Lamia Alsalloom | Advised By: Prof. Sebastian Scherer