Monocular thermal SLAM with neural radiance fields for 3D scene reconstruction

Abstract

Visual simultaneous localization and mapping (SLAM) faces significant challenges in environments with variable lighting and smoke. Excelling in such visually degraded settings, thermal imaging captures scene radiance effectively. To address the limitations of traditional thermal SLAM in 3D scene reconstruction, we propose ThermalSLAM-NeRF, a novel integration of thermal SLAM with neural radiance fields (NeRF). This method significantly enhances the quality of high dynamic range thermal images by improving their signal-to-noise ratio, contrast, and detail. It also employs online photometric calibration to ensure grayscale consistency between consecutive frames. We utilize a sparse direct method for pose estimation, selecting keyframes based on photometric error and tracking quality. The NeRF map is reconstructed using a multi-view keyframe sequence. Our evaluations on datasets containing over 15,000 thermal images show that ThermalSLAM-NeRF achieves an average improvement of 59.30% in trajectory accuracy over existing state-of-the-art SLAM methods. This approach uniquely tracks all sequences and constructs comprehensive NeRF maps, enabling robust and precise pose estimation without the need for extensive pre-training.