Thermal stereo odometry for UAVs

In the last decade, visual odometry (VO) has attracted significant research attention within the computer vision community. Most of the works have been carried out using standard visible-band cameras. These sensors offer numerous advantages but also suffer from some drawbacks such as illumination variations and limited operational time (i.e., daytime only). In this paper, we explore techniques that allow us to extend the concepts beyond the visible spectrum. We introduce a localization solution based on a pair of thermal cameras. We focus on VO and demonstrate the accuracy of the proposed solution in daytime as well as night-time. The first challenge with thermal cameras is their geometric calibration. Here, we propose a solution to overcome this issue and enable stereopsis. VO requires a good set of feature correspondences. We use a combination of Fast-Hessian detector with for Fast Retina Keypoint descriptor for that purpose. A range of optimization techniques can be used to compute the incremental motion. Here, we propose the double dogleg algorithm and show that it presents an interesting alternative to the commonly used Levenberg-Marquadt approach. In addition, we explore thermal 3-D reconstruction and show that similar performance to the visible-band can be achieved. In order to validate the proposed solution, we build an innovative experimental setup to capture various data sets, where different weather and time conditions are considered.