A framework for fully automatic moving video-object segmentation based on graph partitioning

This paper presents an automatic moving video-object segmentation algorithm based on graph partitioning. Initially, each frame is over-partitioned into color homogeneous regions based on the mean shift algorithm and edge detection. This is followed by a low-level morphological region merging step using the so-called flat zone approach to avoid some limitation caused by the watershed-plus-markers technique. Then each segmented frame is represented by a graph in which every region corresponds to one node, and one edge is defined for every neighbouring region of that particular node. Subsequently, a graph partitioning algorithm, namely normalized-cut, is performed on the constructed graph to further merge neighbouring regions. As the final step, the resulting regions are merged to form semantic video objects using the robust optical flow fields. In various parts of the process, Kolmogorov-Smirnov test and the generalized Lloyd algorithm (GLA) are used for clustering. Simulation results at the end of the paper illustrate that the proposed algorithm gives as accurate results as some of the semiautomatic methods introduced in the literature.