Towards Robust Visual Odometry Systems Against Camera Lens Failures

A Visual Odometry system relies on one or more cameras to estimate the motion of an agent. These systems are mainly applied in the real world (e.g. UAVs, Autonomous Cars, etc.) because systems based on cameras are cheaper and easier to install and operate than other alternatives such as LiDARs, and more informative than IMUs. It is then evident that the camera is a critical component of these agents and malfunctions may lead to system failures, from out of trajectory to collisions. In this paper, we show that problems with the lenses, which are realistic in the operational environment of a camera-bearing agent, can alter the proper behavior of the system. Then, we propose a research roadmap to make the system robust to such failures.