Research On360Degree Panoramic Image Stitching

Recently,360-degree panorama is a new hotspot in virtual reality and computer vision. It is a low-cost virtual reality technology based on image stitching. This technology can enhance the interaction with the user and enhancing the rendering, allowing users to have immersive feel. It achieve a pseudo-3D effect in user’s senses, the traditional three-dimensional modeling using three-dimensional laser point cloud to scanning system to reconstruction. This method is widely used in reverse design, product testing, etc., although this method has the advantages of high accuracy, portable, automatic stitching, etc., but the equipment is expensive, slowing in data acquisition and poor texture information. In this paper, we generate a panorama based on image based rendering techniques. It is not only low-cost but also more realistic scenes.This paper firstly introduce the panorama’s concepts, classification, application, the camera imaging model, image transformation model and the process of generating panorama. I have improved the classical algorithm and propose some new algorithms in my work. For example, in Feature matching stage, firstly, we sort the descriptors of SIFT then use the rank number instead of old descriptors achieved better results; in exact feature match stage, this paper presents a new exact matching algorithm based on distance constraints compared to the traditional RANSAC algorithm is more simple and efficient; the image will appear "bending" phenomenon after several stitching, this paper presents a new global correction algorithm to solve the problem; Panorama’s end is start, we use a combination of image texture information and a global coordinate constraints algorithm to achieve this effect; in this paper presents a high-definition image stitching solutions based on use thumbnail to extract SIFT feature points.Panoramic technology is the most intuitive and most inexpensive way to achieve virtual reality and a wide range of applications in many areas. Street services is the application of the most influential, there are many service providers at home and abroad, compared to a traditional map service, in addition to text and two-dimensional map information, also joined the current point of view corresponding to the real scene, also joined in the street in the license plate obscured, face masking, better protection of the public privacy. And panoramic technology still has been widely used in the field of real estate, tourist attractions, virtual campus, project report, etc. The panorama is generally divided into cylindrical panoramas, spherical panorama and cube panorama, cylindrical panorama including the top and bottom information, spherical panorama and cube panorama are full range. Cube panorama is usually converted from spherical panorama, this article introduces the generation algorithm of the cylinder panorama, cylindrical projection transformation is required for all images to be spliced to generate cylindrical panorama, and then extract and match feature, we can find a image transformation model (affine transformation model, perspective transformation model, polynomial transformation, etc.) based on matching feature points, and use the image transformation model to transform the image and mosaic the overlap area. We do it for all images and a panorama is generated.The article uses image stitching to generate panorama, in image stitching, the basic stage is image feature extraction and matching, in feature extraction, using the classic SIFT algorithm, to extract feature points, SIFT feature points remain the same on scale changes, rotation changes, illumination changes and certain invariance to affine changes; in the matching process, using the ratio of the nearest distance and second nearest between the feature vector to determine the which feature points to match. On the basis of this classical algorithm we sort the descriptors of SIFT then use the rank number instead of old descriptors achieved better results, experimental results show that the method is more robust. Nevertheless, can not guarantee that all feature points matching is correct match, but most of the matching points are correct matching points, based on this, this paper presents a exact match method use distance constraints to find all the correct match from all matches, and removed the wrong matching pairs.In the process of generate panoramic often require multiple splicing, with the increase in the number of splicing after stitching the images tend to exhibit substantial "bent", for this problem, we make each column of the image fitted to the same central location to achieve the correction of the image. Panorama must be an end to end, we calculate the texture information of the central region for each and choose one image cut into two images which has least texture information, the two images respectively as panorama’s head and tail, then we alignment the two images, whether the lighting effects, or the continuity of the scene show perfect effect because of they come from the same image.The general situation is that the camera fixed on a tripod when we get the get panorama’s original images, and then clockwise or counterclockwise rotation at an angle to take a photograph and save it. All the saved image’s filename is regular and it is a basis of image adjacent. In feature matching stage, we just judge its adjacent image; With the enhancement of the resolution of the camera, the resolution of the panorama is also proposed new requirements, according to the scale invariance of the SIFT algorithm we proposed a high-definition image stitching program, firstly, we down-sampling the original image and record the sampling scale, and use downsampled images to extract feature and match them, then feature points revert to its original scale. This program can rapid achieve high-definition images stitching.