Realistic 3D rendering and viewpoint synthesis have become essential in industries such as robotics, healthcare. The development of techniques like NeRF and Gaussian Splatting has greatly advanced rendering capabilities. However, their performance is satisfactory only under ideal conditions, such as perfect lighting and the absence of occlusions, which are rarely encountered in real-world scenarios. To address this limitation, we propose the integration of thermal imaging sensors into Gaussian Splatting. Thermal imaging, which captures infrared radiation emitted by objects, offers robustness in low-light and occluded environments. Additionally, we introduce a thermal mapping process combined with a keyframe selection method that accounts for non-uniformity correction, effectively extracting feature points from low-texture and noisy thermal images, thereby enhancing structure-from-motion outcomes. Experimental results demonstrate that our proposed approach outperforms existing methods under challenging lighting conditions, improving the accuracy of 3D rendering across various techniques.
- Book : 20(1)
- Pub. Date : 2025
- Page : pp.104-111
- Keyword :