Modelling the optical-flow with projective-transform approximation for large camera movements

The camera movement estimation targets the finding of a projective transformation which links consecutive images. As far as small motion is considered, an approximation of the vector displacement field is calculated from local optical flow following Longuet-Higgins-Prazdny (LHP) model; the camera large movements are usually estimated from pyramidal implementation of optical flow and LHP model. An alternative to this latter-addressed in this paper — aims to model globally the large optical flow vector field in order to obtain a non approximated homographic model which links two images. The projective parameters are obtained using Bellman's 1D dynamic programming (1D DP) extended to 2D (2D DP) and its pyramidal implementation. The experimental results confirm the better correlation between images matched through the proposed non-approximated homography when compared to other existing local (Lucas-Kanade for example) or semi-global (Horn-Schunck for example) approaches.