Integration of stereo vision and optical flow using Markov random fields

A Markov random filed (MRF) model is used to derive the maximum a posteriori stereo-matched solution for an intensity matching and an edge-matching algorithm. The MRF-Gibbs distribution equivalence reduces the problem to finding appropriate energy functions that describe the constraints on the solution. If a simple MRF is used to model the data, the energy function will yield poor disparity results by smoothing across object boundaries. This problem is particularly apparent with occluded objects. The authors utilize edge and optical flow information to indicate object boundaries and improve the disparity solution in occluded regions. The similarity in intensity (or edge orientation) between corresponding sites of the left and right images is also used in the energy function. Although the authors use simulated annealing to find disparity solutions, a neural-network implementation with analog resistive networks can also be used, as shown by C. Kotch et. al. (1986).<>