Beyond dominant plane assumption: Moving objects detection in severe dynamic scenes with Multi-Classes RANSAC

We consider the problem of solving moving objects detection in severe dynamic videos captured by a freely moving camera. In complex scenes, especially in indoor scenes, traditional single layer homography and affine transformation model are not strong enough to describe the background motion. This paper utilizes multiple 2D affine transformations to describe the background motion caused by moving camera. Multi-Classes RANSAC is presented to estimate the parameters of the motion model. With an iterative step, it can attempt RANSAC parameters several times(in previous only once), thus fit various data. Background/foreground analysis is also presented, avoiding computing background motion model by foreground motion information. Experiments and comparisons to other motion compensation methods demonstrate the better and more stable performance of the proposed method.