Error sources and their impact on c-velocity methods

We propose an efficient method for detecting 3D planes from a moving uncalibrated camera. It assumes only a 3D translation for the egomotion. The original idea is to get around the dependence between the 2D motion field, the depth (i.e. scene structure) and the ignored egomotion. Relations are established between predicted iso-velocity curves and 3D surface parameters in eliminating the depth variable from equations. Surface assessment is performed within a simple velocity voting-space: a 3D surface, like a given plane, emerges as a curve, like a parabola. Experimentations validate our approach for several applications including automatic driver assistance, showing a surprising robustness. The aim of the present paper is to analyze this robustness. Five main perturbations are focused on and their influence on the plane detection process is studied in details.