关于稳定特征路标的一点描述

When it is necessary to find a particular feature in a subsequent image, a search is carried out using normalised sum-of-squared-difference: the
patch detection algorithm is not used on the new image. This differs from many structure from motion systems where the feature detector is run on each new image: features are found as local maxima of the particular feature-finding operator, and then correlation matching is carried out between the features found in the current and previous images, all of which
are local maxima. Some authors [21] have found that requiring all patches matched to be maxima like this is restrictive, and features can be missed which still give good correlation matches. This is a good point when applied to our active system, since it is important to keep track of the same features for as long as possible. Since a search in a particular image is only for one feature at a time, it makes sense to move straight to correlation matching.