An Empirical Analysis of Errors in Structure from Motion

The process of registering airborne video to some reference imagery described by Cooke et al. [2] involves feature detection and tracking, structure from motion with missing data, georeferencing, dense matching and 2D image registration. This paper considers two aspects of this process. Firstly, it compares the errors produced by a range of methods for extracting the camera parameters and the scene structure from an incomplete set of feature measurements. Secondly, it examines the effect of this error on the overall system performance by describing a framework in which the error in the final registration may be automatically estimated at any point in a video sequence. The quality of the estimated errors in some of the intermediate results is then checked numerically using an airborne video sequence of an airport control tower.