Progressive Multi-image Registration based on Feature Tracking

Many previous multi-image registration methods perform their global registration process when all images of the input sequence are available. However, these methods are not suitable for real-time applications that aim to progressively register the current incoming image with respect to the previous images. In this paper, we present an innovative progressive multi-image registration method that uses an efficient feature-based pairwise local registration and feature tracking. First, transformations for temporally successive frames are estimated and concatenated to relate non-successive images. This pairwise registration recovers 2D transformation parameters using feature correspondences and an iterative least squares method. Then, overlapping images are detected among the previous and current image by mapping all images to the reference image coordinate space. When multiple overlaps exist between the current image and previous images, multi-image registration is performed. To do so, features in the overlapping regions are tracked, i.e. same features are identified from non-successive frames within the overlaps and linked together. These are the same features used earlier for pairwise local registration. Tracked features contribute to compute geometric errors caused by the concatenation of pairwise transformations and are used to correct misalignment of the current image. Our global registration method extends methods used for local registration (feature matching and parameter estimation) to feature tracking and parameter modification for the global registration in a unified way. The accuracy of our method is demonstrated by creating mosaic images from sets of images