3-D modeling of indoor scenes by fusion of noisy range and stereo data

The authors present a method that merges and fuses two kinds of 3-D data, surface data obtained with a laser rangefinder and photometric data produced by a stereovision system. It describes indoor scenes with a set of planar faces, a first step toward polyhedral modeling. Noisy geometrical features are represented by random variables whose variance is known. Extended Kalman filtering techniques are used for numeric fusion of features into higher level ones (from points or pixels to 3-D lines, from 3-D lines to planes) and for identification of the transformation linking the sensors´ reference frames (calibration). This latter issue receives particular attention, and a methodology is presented for iterative and automatic relative calibration. Experimental results show that this description is sufficiently accurate and reliable for feeding higher level processes, such as mobile robot localization or scene interpretation