Camera-Based State Estimation for Surface Vessels

Simultaneous localization and mapping (SLAM) is the twofolded problem of locating a vehicle starting at an unknown location and mapping the surrounding environment simultaneously. This thesis considers the SLAM problem using a sensor package consisting solely of a monocular camera. Additionally, the application of visual SLAM algorithms to autonomous surface vehicles (ASVs) is considered. Different SLAM paradigms are discussed, today’s state-of-the-art visual SLAM systems are compared, and visual SLAM is tested using CS Saucer in MCLab at NTNU. While filter-based methods for a long time were ahead of the game, the future of visual SLAM is considered graph-based where SLAM is formulated as an optimization problem. This thesis goes into great detail in outlining the architecture of a modern visual SLAM system, diving into the fields of computer vision and photogrammetry. The current state-of-the-art algorithm for visual SLAM is ORB-SLAM developed by Mur-Artal et al. (2015). This is a local feature- and graph-based system outperforming all of its predecessors. The mapping performed by ORB-SLAM contains sparse metric information. Dense metric information is required for path planning and collision avoidance. This can be achieved either by augmenting the sensor package, or incorporating global descriptors in the visual SLAM algorithms providing dense metric mapping. Autonomy and safe guidance at sea are discussed, and visual SLAM is seen as a key element of an autonomous future. State estimation based on visual SLAM is compared to that of the Qualisys Motion Tracking System. The accuracy is considered inferior at its current state, but the rich information provided by cameras should be taken advantage of in an autonomous future. A method for determining scale, orientation and translation of a visual SLAM system is developed and tested with great results.