Local features can be used in many vision applications such as image matching & indexing, object recognition, mobile robot localization, etc. Robustness and Invariance are the two most important characteristics of local features which influence entire feature performance for matching and recognition. Recently, Local scale & affine invariant features have gained popularity in the local feature research due to its ability to handle large image transformations.
Local feature detection is composed of interest point detection, invariant region extraction and region description. The first step extracts a set of characteristic points from an image. In the second step, invariant regions are extracted for these detected points. Finally, these invariant regions are characterized by descriptor vectors to provide invariance for matching.
In this research, we developed a novel robust local invariant feature detector for image matching and recognition. The proposed feature detector is based on the key idea of tracking and grouping scale space interest points to extract reliable scale invariant features. The algorithm significantly improved the repeatability rate over the existing feature detectors. In addition, based on the same tracking and grouping method, we further generalized the proposed scale invariant detector to the affine invariant one for handling large viewpoint changes.
Our proposed feature detectors have high repeatability and invariance to the various geometric and photometric transformations. The efficiency and usefulness of the proposed feature detection methods are confirmed by the excellent experimental results and performance evaluation. Finally, we applied the developed feature detectors to the object recognition framework. The results showed that our feature detectors can be robustly and efficiently used in the real-time object recognition system.