Motion field estimation in images based on segmentation

Optical flow estimation in videos is useful in intelligent video analysis. To realize this, we need to combine segmentation algorithms with feature matching methods. In this paper, we originally combine machine learning methods (Adaboost) with feature extracting and block matching algorithms. We first use a classifier to divide an image into different regions, then match 12-by-12 blocks in these regions to corresponding ones in another image. Finally, global motion estimation and refining algorithms are proposed to correct the wrong segmentations and wrong matches, accurate matches as well as optical flow estimation can be obtained. Experiments are used to testify the usefulness.