Directed correspondence search: Finding feature correspondences in images using the Harmony Search algorithm

A robot's ability to correctly recognise landmarks as ones previously seen is a key feature in visual navigation algorithms. Each newly captured scene contains multiple visual features with an unknown number of new features that have never been observed and old features that need to be correctly recognised and identified. This challenge of correctly identifying common features between different frames is also known as the visual correspondence problem. In this article a new algorithm is presented that attempts to solve this problem by using an adapted version of the Harmony Search algorithm. All features are assumed to be static and consistency in the predicted camera movement is taken into account to remove outliers. Initial results show rapid convergence early in the optimisation process. This suggests that DCS could be effectively used as a rapid way of initialising other algorithms like RANSAC.