Stereoscopic analysis of multiple images
Sébastien Roy, Jean-Guy Meunier · 1993
A new algorithm to recover depth from a sequence of two or more stereoscopic images is presented. The algorithm, which uses a dynamic programming approach, builds a dense depth map and allows the camera displacement between each image to be any combination of rotation and translation. Since no smoothing constraint on depth is used, occlusions and depth discontinuities along the border of objects are preserved and easy to identify. For a given cost function, the algorithm finds the optimal correspondence along epipolar lines. In our case, this function is the difference of intensities between corresponding points, adjusted with a factor accounting for occlusions. Tested on nontrivial synthetic image sequences with true depth map available, we obtain a mean disparity error of less than one pixel. submitted august 7 th 1993 to the International Journal of Computer Vision Copyright 1993 by Sbastien Roy and Jean Meunier 2 1 Introduction Depth estimation of objects in a scene is very u...