Practical Infrared Visual Odometry

Abstract

The use of cameras as a sensor for odometry estimation is an active research topic that has seen significant growth in recent years. Most methods, however, are only suitable for standard cameras that rely on reasonable lighting. An alternative to overcome low-light conditions is the use of thermal or long-wave infrared imaging. Although visible spectrum and thermal imaging share many characteristics, it is not straightforward to apply standard visual odometry algorithms to thermal imaging data. In this paper, we propose a practical visual odometry system based on a monocular thermal camera. As monocular odometry suffers from an unknown scale factor, the system performs efficient ground plane detection for targeted feature extraction, such that the scale factor can be reliably estimated if the camera height and pitch are known. We also address the problem of periodic nonuniformity correction, which is a necessary characteristic of thermal cameras that freezes the output potentially for over a second and can severely affect motion estimation. In this sense, we automatically determine appropriate times to perform nonuniformity correction based on the current and predicted camera rotations. Experiments illustrate the applicability of the system and compare it with other state estimation approaches.