Automatic design of morphological operators for motion segmentation

A problem of interest in digital video editing is the elimination of moving objects from one video and the introduction of pieces of other videos in their places. A fundamental problem to build computational tools for this purpose is the segmentation of moving objects. The paper approaches this problem with a novel technique based on Beucher-Meyer's paradigm (F. Meyer and S. Beucher, 1990), with markers detected by morphological operators designed by computational learning techniques. The objects in the first frames of the video are marked manually and used to train the marker detector. Then, the operator designed is used to mark the objects in the other frames and the paradigm is applied to all frames marked by the detector. Some synthetic and real world examples illustrate the application of the technique proposed. Complex situations such as occlusion are examined.