Robust Thermal-Inertial Localization for Aerial Robots: A Case for Direct Methods

Abstract

GPS-denied localization of aerial robots in sensing-degraded environments and especially in dark, textureless and dustor smoked-filled settings is rendered particularly hard. An alternative is to utilize Long Wave Infrared thermal vision which is unaffected by darkness and can penetrate most obscurants. However, utilization of thermal vision is mostly considered in combination with traditional visual-inertial odometry frameworks which are made for grayscale visible-light camera images and fail to utilize the full (e.g. 14-bit) radiometric information of a thermal camera system. Using rescaled thermal images makes the data compatible with existing vision-based odometry solutions yet comes at a cost of information loss which in turn can lead to unreliable tracking of image correspondences and thus to odometry estimation failure. An alternative is to use direct methods exploiting the full radiometric information provided by a thermal camera. This paper presents a comparative study between state-of-the-art visual-inertial odometry frameworks applied on rescaled thermal camera data against full radiometric information-based methods using a modified filter-based estimator and a recently proposed keyframe-based direct technique specifically designed for thermal-inertial localization. In a data-driven manner we demonstrate that direct full radiometric information-based thermal-inertial localization should be the prime selection especially in hard environments with weak thermal gradients. The presented results include comparison both inside a motion captured-controlled environment, and in an underground mine.