Self-TIO: Thermal-Inertial Odometry via Self-Supervised 16-Bit Feature Extractor and Tracker

Abstract

In recent years, thermal odometry has gained significant attention in mobile robotics for addressing visually degraded scenes. To achieve reasonable robustness and accuracy of thermal odometry, a learning-based image feature extractor and tracker has been proposed. While learning-based methods generally provide better feature tracking results in thermal images compared to classical methods, they still require labeled data for training and struggle with real-time execution. To deal with these issues, this letter presents a robust and accurate thermal-inertial odometry (TIO) system, Self-TIO equipped with a self-supervised feature extractor and tracker designed for the 16-bit radiometric image domain. Moreover, Self-TIO employs a hybrid tracker, combining the Kanade–Lucas–Tomasi (KLT) tracker and learning-based optical flow, to achieve high robustness and sub-pixel accuracy, even in scenes affected by non-uniformity correction (NUC) and aggressive motion. Experimental results demonstrate that our method outperforms state-of-the-art methods in both feature tracking and thermal-inertial odometry.