Bundle adjustment in large-scale 3D reconstructions based on underwater robotic surveys

In this paper we present a technique to generate highly accurate reconstructions of underwater structures by employing bundle adjustment on visual features, rather than relying on a filtering approach using navigational sensor data alone. This system improves upon previous work where an extended information filter was used to estimate the vehicle trajectory. This filtering technique, while very efficient, suffers from the shortcoming that linearization errors are irreversibly incorporated into the vehicle trajectory estimate. This drawback is overcome by applying smoothing and mapping to the full problem. In contrast to the filtering approach, smoothing and mapping techniques solve for the entire vehicle trajectory and landmark positions at once by performing bundle adjustment on all the visual measurements taken at each frame. We formulate a large nonlinear least-squares problem where we minimize the pixel projection error of each of the landmark measurements. The technique is demonstrated on a large-scale underwater dataset, and it is also shown that superior results are achieved with smoothing and mapping as compared to the filtering approach.