Minimum-error active matching for real-time vision

As an integral part of real-time vision system, there are two most important requirements for feature matching mechanisms: high computational efficiency for meeting the real-time demands, and high correct matching rate for ensuring the convergence and consistency of state estimation. Both of these are addressed and solved as an integrated whole by the efficient minimum-error active matching scheme proposed in this paper. Image processing is performed in a dynamically guided fashion by checking only parts of the image where positive matches are most probable. For achieving the global consensus matchings, rigorous analysis on how to minimize the matching errors in active matching by choosing an optimal search order is made. After that, practical feature matching algorithms are given, which have naturally absorbed the ideas of nearest neighbor (NN) and joint compatibility branch and bound (JCBB) approaches. Both statistical simulations and real-world experimental results have verified the proposed methods can perform better than the state-of-the-art algorithms, i.e. being able to obtain the best global consensus matchings with much lower computational cost.