Determination of the essential matrix using discrete and differential matching constraints

We present a method to determine the essential matrix using both discrete and differential matching constraints. Differential constraints, derived from optical flow, are abundant in contrast to the discrete constraints, derived from feature correspondences, which are scarce when just a limited number of salient features are available. We formulate a likelihood of the camera motion given the correspondences of a set of features and the image velocities of these features. We show how this likelihood can be used to determine the essential matrix both in a robust hypothesize-and-test framework, and then in non-linear iterative refinement. Our results show that the use of the extra optical flow constraints gives better estimates of the essential matrix, when compared to using the discrete data alone.