Robust motion segmentation for on-line application

This paper presents a novel approach for on-line video motion segmentation. Common methods were designed for off-line processing, where time to process one frame is not so important and varies from minutes to hours. The motivation of our work was an application in robotic perception, where a high computational speed is required. The main contribution of this work is an adaptation of existing methods to a higher computational speed and on-line processing. The proposed approach is based on sparse features, we utilized the KLT tracker to obtain their trajectories. A RANSAC-based method is used for initial motion segmentation, resulting motion groups are partitioned by a spatial-proximity constraints. The correspondence of motion groups across frames is solved by one-frame label propagation in forward and backward directions. Finally, an approximation of dense image segmentation is obtained by using the Voronoi tessellation.