Super-resolution Mosaicing using Embedded Hybrid Recursive Folow-based Segmentation

We present a new strategy for the generation of background super-resolution mosaics from videos with arbitrary camera pan, tilt, and zoom including freely moving foreground. Our main focus is directed to the automatic, embedded pre-segmentation of foreground objects. The segmentation technique is based on efficient and robust computation of the optical flow between neighboring frames in a video scene using a hybrid recursive approach, i.e. a combination of block-flow methods and spatial-temporal anisotropic diffusion-based flow field regularization. Unlike in other related publications we are able to segment moving foreground objects before the actual image-to-mosaic-registration is proceeded even if the foreground objects do not move relatively to the camera motion. Thus, every segmented background frame can be used to enhance the resolution of the composed mosaic due to an effective blending process. Additionally, the appearance of disturbing ghost objects is prevented