A Fast Video Stabilization System Based on Speeded-up Robust Features

In this paper,a fast and efficient video stabilization method based on speeded-up robust features (SURF) is presented .We adopted speeded-up robust features as feature descriptor,which are extracted and tracked in each frame .After that, we further refined the matching features through RANSAC, estimating the motion parameters through least squares method and computed the integrated motion. Experimental results illustrate superior performance of the SURF based video stabilization in terms of accuracy and speed when compared with the Scale Invariant Feature Transform (SIFT) based stabilization method. Introduction Video sequences acquired by a camera mounted on a mobile platform are affected by unwanted shakes and jitters. Several digital video stabilization approaches have been proposed to overcome this problem such as block matching[1],optical flow[2],phase correlation[3]and feature matching[4]. Block-matching methods divide a frame into blocks,and compute a motion vector for each one through the searching of the more similar block in the next frame.However,the motion estimation could be biased in low-textured image regions due to the aperture problem.Feature-based methods overcome this problem by computing the motion only in regions that stand out according to a specific image feature.At present,with the feature point extraction technology development,based on feature points matching algorithm to stabilize the image has become the electronic technology to stabilize the image of the main development direction.Commonly used in electronic image Feature points are: Harris corner[5],Smallest Univalue Segment Assimilating Nucleu(SUSAN)[6],Scale Invariant Feature Transform(SIFT)[7],Speed Up Robust Feature(SURF)[8].In this context,a novel digital video stabilization algorithm is proposed, which overcomes the previous problems by computing a robust motion estimation through a variation of the SURF algorithm adapted to video sequences to bediscriminative to scale and orientation. SURF feature extraction and matching In this paper, using the algorithm of SURF as feature point extraction algorithm of image registration.SURF is a translation, rotation and scale invariant feature detector, which is based on Hessian matrix for its good performance in accuracy[8].Compared with SIFT which uses Difference of Gaussian (DoG) to approximate Laplacian of Gaussian (LoG), SURF pushes the approximation even further. It approximates Laplacian of Gaussian by using a box filter to represent the corresponding kernel. The kernel approximation is efficient in computation by using the integral images and hence the time consumption is independent of the filter size.After building the image pyramid, the process continues by traversing the pyramid to remove those points with low contrast and then searching extrema on neighboring scale images. Finally,the points are localized to sub-pixel accuracy through scale space interpolation. SURF descriptor uses Haar wavelet in x and y directions to compute a reproducible orientation. To achieve rotation invariance, a square descriptor region is constructed along the dominant orientation and is divided into 4×4 sub-regions and the descriptor is extracted from it.In addition,SURF computes Haar wavelet through integral images, which decreases the computational complexity.Each wavelet requires only six operations to perform the computation. International Industrial Informatics and Computer Engineering Conference (IIICEC 2015) © 2015. The authors Published by Atlantis Press 1995 Since SURF divides all the feature points into two types by the sign of Laplacian,we can boost the matching speed by comparing the sign of Laplacian.In addition,we drop the unreliable matching by comparing the ratio of distances from the closest neighbor to the distance of the next closest neighbor with a predetermined threshold. Motion estimation The real camera motion between frames is a 3D motion.As a trade off between the complexity and efficiency, we adopt a 2D affine model to describe the motion between frames: