B. Matching
As shown in [7] several methods exist for analyzing the translation between consecutive images. If feature points are extracted, then some kind of matching algorithm is required to calculate translation. In [12] Harris corners are extracted and in the matching are performed with normalized cross correlation over an 11x11 pixel neighborhood. Usually normalized cross correlation is used on small regions. Two reasons are less computational cost and fewer problems caused by different projections of the images. The perpendicular mounting of the cameras and selection of a lens with low distortion reduces these problems.
Since normalized cross correlation can provide a measure of traveled distance directly without any feature extraction, this method is selected. Sub-pixel accuracy is obtained by fitting a paraboloid to the peak in the similarity image [13]. First similarity between the stereo pair is calculated to obtain the height and then between two consecutive images of the left camera to obtain the translation. Fig. 3 shows how the
disparities are calculated.