Real time video segmentation optimization with a modified Normalized Cut

The low-level task of foreground-background segregation is an important foundation for many high-level computer vision tasks and has been intensively researched in the past. Nonetheless, unregulated environments usually impose challenging problems and often particular difficulties arise from real time requirements. In this paper we propose a new energy function to evaluate the spatial relations in a segmentation. It is based on the Normalized Cut but adapted these principles to the usage of videos instead of single images. This makes it possible to get a comparable spatial-accuracy as in state of the art approaches (e.g. Markov Random Fields). However, the optimized hierarchical local minimization process for our energy function is at least two orders of magnitude faster. In combination with an efficient Background Subtraction this results in an accurate real time video segmentation algorithm even for high definition videos.