Real-time video mosaicking robust to dynamic scenes

This paper presents a system for creating a mosaic image from a sequence of images with moving objects present in the scene. This system first uses SIFT-based image registration on the entire image to obtain the initial global projection matrix. After image segmentation, the global motion model is applied to each region for evaluation. The transformation matrix is refined for best projection in each region, and a more precise global transformation matrix is calculated based upon local projections on majority of coherent regions. As a consequence this method is robust to disturbances to the projection model induced by moving objects and motion parallax. In the image blending stage, pixels in coherent regions are weighted by their distances from the overlapping edges to achieve a seamless panorama, while heterogeneous regions are cut and pasted to avoid ghosting or blurring. The most recent information regarding location, shape, and size of the moving foreground objects is therefore reflected in the panorama. Constructed mosaics are presented to demonstrate the performance and robustness of the proposed algorithm.