Date of Publication :7th July 2016

Abstract: In Feature Based Motion estimation, the most time consuming process is the key point extraction (features) and their similarity measurement. In our paper, we use SIFT based feature extraction in adjacent frames. Based on similarity measurement, few features are selected. The selected features are used to calculate the motion vectors or Global motion between two frames. The computation complexity is higher if we use the whole image to extract points. The complexity of key point extraction is proportional to the area of the images (frames) and the complexity of similarity measurement is proportional to the square of the area. In our paper, we propose block based feature extraction. Using blocks for feature matching, the number of key points is reduced and hence accuracy might be decreased. We use four blocks in the four corners of the frame and use SIFT to extract key points and measure similarity. Now we have several reliable matches from each block, yet interference of moving objects may cause error in global motion estimation. This error can be removed by applying inter block feature matching. Randomly select three points (each from different block) and calculate the area formed by the triangle in both the images. This matching will give us features which are not part of local motion caused by moving objects. Now we have the best matched features obtained from SIFT. Feature matching is very important and crucial step in motion estimation. Mismatch is always inevitable in motion estimation for feature based methods no matter how powerful the feature point locator is. To solve this problem, we propose a novel idea of forming a triangle based on three best features. The pixel values of the triangle formed in two images is accumulated and matched. This matching ensures the removal of faulty features. Though theoretically this may sound good, practically this might not work well if the images are blurred. To avoid this motion de-blurring of frames can be done. Now we have best matched features, using that we calculate the global motion vectors. These global motion vectors are smoothed by using filter. The smoothened motion vectors are used to stabilize the video.

Reference :