Visual motion estimation via second order cone programming

The visual motion, induced by ego-motion of the camera, can be estimated through resection, intersection and transfer processes. Under the assumption of affine camera, the intersection/transfer process can be formulated as a system of 5 linear equations, so that any correspondence of an image point and its affine coordinates can be obtained by solving the equations using least squares (LS) techniques. However it produces sometimes a poor estimation result, due to the singularity of the coefficient matrix. In order to solve the problem, instead of trying to find an exact solution of the equations, we tried to obtain a robust least squares (RLS) solution via a second order cone programming technique. The superiority of RLS over LS is demonstrated by the experimental results