Motion Segmentation Based on Integrated Region Layering and Motion Assignment

We describe a novel algorithm for segmenting image sequence frames into regions of pixels moving with coherent motion. It is based on fusing local grey level segmentations with motion estimates obtained using block and partial correlations. The key innovation is the method employed to assign motion labels to the grey level regions. This uses an explicit model of motion occlusion and uncovering based on boundary ownership which predicts the location of motion-compensated difference energy for a given labelling and depth ordering of adjacent regions. A significant advantage of the approach is that region layering is automatically generated with the best assignment. We incorporate the scheme into a global segmentation algorithm in which the local motion assignments are combined using a consistency criterion. This leads to layered sets of connected sub-regions representing the segmented motion regions within the frame. Results of experiments demonstrate that the approach is effective.