Robust thermal–inertial SLAM with online photometric calibration and point-edge flow tracking

Thermal cameras can directly capture the thermal radiation in the environment and realize effective perception in low illumination environment. However, due to the complex temporal and spatial photometric changes in low-contrast thermal images, it is difficult to construct robust data association to realize pose estimation. To solve the above problems, a new thermal–inertial simultaneous localization and mapping (SLAM) system is proposed in this paper. This method carefully analyzes the causes of photometric change, and calibrates the temporal and spatial parameters of thermal images online, which effectively maintains the cross-frame photometric consistency of thermal images, thus establishing a robust point flow tracking based on appearance. Meanwhile, this paper constructs proxy textures representing geometric structures using a limited distance transform based on thermally salient edges. Furthermore, we propose a limited distance transform-based Kanade–Lucas–Tomasi tracker to realize edge flow tracking. The Edge flow tracking not only provides reliable feature matching results for photometric calibration, but also completes accurate pose estimation together with point flow tracking. The experimental results demonstrate that the proposed SLAM system achieves precise localization in various indoor and outdoor environments. For indoor scenarios, the method attains localization accuracies of 0.06 m and 0.04 m on public and self-collected datasets, respectively, reaching centimeter-level precision. In outdoor scenarios, the method maintains robust performance across all sequences, particularly in unstructured environments like the Forest scene, where it achieves a 23.4% improvement over baseline methods. Additionally, the method retains high real-time performance, operating at a stable frequency exceeding 30 Hz.