Maximum-likelihood optical flow estimation using differential constraints

Many optical flow estimation techniques are based on the differential optical flow equation. For these techniques, a locally constant flow model is typically used to allow the construction of an over-determined system of constraint equations. In this paper, the problem of solving the system of optical flow equations using a constraint total least squares (CTLS) approach is investigated. It is shown that by modifying the CTLS approach it becomes identical to a maximum likelihood (ML) approach to the problem. This modification improves the CTLS estimates especially when the estimation window size is small, as is demonstrated experimentally.