Performance and accuracy in visual motion computation from FS sonar video sequences

In addition to its numerous terrestrial applications, the capability to compute self (or a target's) 3-D motion is important for many underwater operations. While optical systems are the most common imaging modality above the sea, serious complexities arise when the mission is to be carried out beneath the sea surface, with turbidity being the most prohibiting factor. Acoustic-based imaging and methods, because of the ability to penetrate silt and other sources of turbidity, offer a more effective solution. It is thus desired to devise robust motion vision techniques comparable to those realized for optical cameras. In this paper, we explore the impact of selected key factors on the accuracy of motion estimation from 2-D sonar cameras. These factors comprise the field of view of the camera, the number of feature matches in a motion sequence, and the inaccuracy in the knowledge of their image positions. The significance of each factor is analyzed based on the variances of the sought after motion parameters. We also present various results from an experiment with real data, where we apply different measures that provide us with ground truth for the quantitative assessment of the motion estimation accuracy.