Panoramic background model under free moving camera

segmentation of moving regions in outdoor environment under a moving camera is a fundamental step in many vision systems including automated visual surveillance, human-machine interface, tracking etc. It is also a challenging task due to camera motion, object motion, and outdoor scene challenges i.e. periodic motions of swaying of trees, gradual illumination changes, etc. In this paper, a wide area scene modeling approach for object segmentation under a moving camera is proposed. This approach suffers due to parallax effect, misallignment errors etc and needs their concurrent removal for its success, we explicitly model the dense correspondence between input image and panoramic background model. Foreground segmentation and correspondence estimation are expressed as a unified labeling problem, and are solved efficiently via tree dynamic programming (TDP). Lucas-Kanade method is used to find sparse correspondence between image and model, and robust M-estimator is then applied to find the projective transformation for initialization of TDP's window search. Optimal dense correspondences are achieved and are used to update panoramic background model and as a byproduct, online refined panoramic image is generated which is empty in the beginning and is filled step by step. We test our algorithm with hand-held camera and also with a camera mounted on a moving platform. Experiments proved our algorithm to be robust in performance.