Monocular navigation for long-term autonomy

Abstract

We present a reliable and robust monocular navigation system for an autonomous vehicle. The proposed method is computationally efficient, needs off-the-shelf equipment only and does not require any additional infrastructure like radio beacons or GPS. Contrary to traditional localization algorithms, which use advanced mathematical methods to determine vehicle position, our method uses a more practical approach. In our case, an image-feature-based monocular vision technique determines only the heading of the vehicle while the vehicle's odometry is used to estimate the distance traveled. We present a mathematical proof and experimental evidence indicating that the localization error of a robot guided by this principle is bound. The experiments demonstrate that the method can cope with variable illumination, lighting deficiency and both short- and long-term environment changes. This makes the method especially suitable for deployment in scenarios which require long-term autonomous operation.