EFFECTIVENESS OF VIDEO OBJECT SEGMENTATION BASED ON MPEG LIKE MOTION VECTORS FOR 3D DEPTH ESTIMATION

摘要

This paper discusses the effectiveness of video image segmentation based upon macro blocks associated with the motion vectors defined and embedded in MPEG codec, for the purpose of estimating the 3D structure of a moving object in the video scene. Sequential video frames do not necessarily meet the rigidity assumed in the 3D estimation/reconstruction techniques for stereoscopic pair images of epipolar geometry, thus requiring extraction of the moving objects. To compare the effectiveness of the image segmentation, the robust RANSAC algorithm, which calculates the fundamental matrix of epipolar geometry, was used to statistically examine the deviation of epipoles between two consecutive video images. Test was conducted for a video sequence without image segmentation and for the case where a moving object was extracted based on the macro blocks that have a common specified motion vector. The case with the image segmentation was favored in terms of more consistent values of the epipoles as well as the fundamental matrix.