Segmentation of dynamic objects in video sequences fusing the strengths of a background subtraction model, optical flow and matting algorithms

In this paper we propose to combine some state-of-the-art video segmentation algorithms in an unsupervised fashion to take advantage of its strengths. The proposal is based on a Background Subtraction (BS) model with a Self Organized Map neural network architecture and automatic threshold update that has been proven to be robust to illumination changes and slight shadow problems. An Optical Flow algorithm will analyze the dynamic regions detected by the BS model and will enclose within an ellipse those that have similar velocities and probably define the same dynamic object whose identification was not complete because of camouflage issues. This ellipse will serve as the input needed by a Matting algorithm that will improve the definition of the dynamic object by minimizing a cost function. Final segmentation results demonstrate that the fusion of these two methods produce better recall and false negative rate metrics compared to the BS model alone. Comparing our results against other reported state-of-the-art models demonstrates the effectiveness of our proposal.