Vision-based state estimation of autonomous boats

Abstract

For any autonomous vehicle, such as a self-driving boat, it is essential to estimate its localisation accurately. One approach to this problem is to use visual odometry, which is a purely vision-based state estimation. Today, autonomous boats mainly use global navigation satellite systems (GNSSs) or inertial measurements units (IMUs) and are commonly only partly self-driving. In contrast, a camera-based system would be more cost-effective and function in areas where there are no signals from the GNSS. However, a vision-based state estimation tends to be not as accurate. This project implemented the algorithm called direct sparse odometry to investigate how such a monocular localisation system could replace an IMU and a GNSS. At the same time, the work addressed how this method and similar kinds of algorithms could be automatically evaluated at a future web-based platform. We could show that the algorithm did not perform so well on the chosen sequences. However, there are indications that a direct method could attain better performance than a feature-based visual odometry method. The project’s results demonstrate how the architecture of an algorithm running on the platform can be designed and showed directions for research of more accurate performance. For example, to use several monocular cameras or use a full simultaneous localisation and mapping (SLAM) system instead would probably result in a more precise vision-based state-estimation of a boat. The resulting algorithm will hopefully work as a reference algorithm for future localisation algorithms on the mentioned platform. Moreover, the conclusions drawn from what requirements are put on such algorithms can facilitate the platform’s design and implementation.