A Constraint Based Algorithm for Dense Disparity Estimation in Stereo Vision

This article presents a new algorithm for dense stereo correspondence using Constrained Non-Linear Optimisation. The disparity is modelled parametrically as a continuous surface and uses a block decomposition scheme to allow distributed parallel execution. Inter-block continuity is modelled explicitly to eliminate surface jaggedness, while intra-block smoothness is orientated with the image gradient to preserve depth discontinuities. The use of arbitrary nonlinear constraints allows more powerful enforcement of regularisation and also, higher flexibility in embedding scene constraints.