Figure 2.

Schematic illustrating our novel algorithm for correcting IMU orientation estimates with rotational information from image stitching (Scheme 2). The IMU estimates an orientation matrix O_i for each image i that goes into the panorama, and each O_i provides an orientation estimate relative to absolute world coordinates (defined by magnetic north and gravity), but with too much noise to be solely relied upon to create an aerial view. The image panorama algorithm independently furnishes rotation estimates R_i for each image, but these are not calibrated with respect to absolute world coordinates. However, the rotation estimates can be combined to create highly accurate rotations between image pairs, so that $R_j{R}_i^{-1}$ transforms quantities from image i to image j. These inter-image rotations are combined with the IMU data to produce reconciled orientation estimates (see text).