Estimation of depth-from-motion combining iterative prediction scheme and regularization framework

The purpose of depth-from-motion estimation is to determine the depth function of a scene using the image sequence acquired by a mobile sensor whose motion is known. An elementary depth estimation can be deduced from the apparent motion in the image plane for each pair of consecutive images. By using successive estimations over time, it should be possible to get more complete, accurate and reliable depth information. We propose a scheme implementing such a temporal integration. The process recursively updates a depth value map and a confidence map. Both maps are predicted at each time interval using camera motion. The update step is performed within a regularization framework, which allows one to combine predicted depth values with a current depth observation associated to corresponding confidence measures at each pixel