Study Of Panorama Image Mosaics Based On Features

Image-Based Rendering technology constructs lifelike virtual 3D environment with the 2D images of real scene by computer, which avoids complex calculation during scene construction and rendering, and has many advantages, such as low hardware requirement and strong practicability. Panorama technique is a kind of typical IBR technology. Panoramic image has been widely used in tourism, education, exhibition, intelligent monitoring and other fields. Panoramic image mosaicing method based on features has become the hot topic in the study of image stitching because of the small amount of calculation, good robustness and extensive adaptability.In this paper, cylindrical panoramic image mosaicing technique based on point features is studied. The research is focused on two key technologies of image mosaic including image registration and image fusion. In order to improve the speed of algorithm, some improvements are proposed on the basis of in-depth research of the existing algorithm.In image registration phase, this paper presents an algorithm to extract feature points which uses an improved FAST (Features from Accelerated Segment Test) algorithm to detect corners and SURF (Speed-Up Robust Features) algorithm to compute the direction and feature vector for each point. Then using randomized KD-tree searching tactics to find two approximate nearest neighbor, the best matching points set is determined through comparing whether the ratio between nearest neighbor and second nearest neighbor is less than a given threshold. The match procedure is two-way between reference image and the image to be registrated. The RANdom SAmple Consensus (RANSAC) algorithm is used to remove false matching points, and the transformation parameters are calculated. Compared with the conventional SIFT (Scale Invariant Features Transform) and SURF algorithm, the speed of feature detection and image matching is faster. Experimental result shows that this method can effectively match the images which have a certain degree of scaling, rotation, illumination changes and viewpoint changes and so on.In image stitching and fusion phase, the effect of the weighted average fusion is compared with the multiresolution spline fusion through experiment.. The best seam method based on graph cut with multiresolution spline fusion is used to remove the ghost caused by moving object and/or registration deviation, and obtains good fusion effect. Finally using the above algorithms combined with bundle adjustment technique as a whole alignment of panorama, a seamless cylindrical panorama is successfully achieved.