Moving region segmentation from compressed video using Global Motion Estimation by macroblock classification and Markov Random field model

Abstract

In this paper, we introduce new method to segment the moving regions from compressed video by incorporating more features from different previous segmentation methods. Briefly, our method proceeds as follows. First we classify the macroblocks of the compressed video frames into different classes and we perform Global Motion Estimation and Global motion Compensation techniques to remove the influence of camera motion on the Motion Vector field from the compressed video. Then Motion vector quantization (VQ) based on similarity of local motion is used to find the likely number of moving regions. The inferred statistics are used to initialize prior probabilities for subsequent Markov Random field (MRF) classification, which produces coarse segmentation map. Finally, coarse to fine strategy is utilized to refine region boundaries. This proposed approach produces accuracy in segmentation. While each of these components has been employed in previous segmentation approaches, we believe that complete solution incorporating all of the listed components is novel and represents the main contribution of this work.