Research On Image Registration Methods For Augmented Reality

Augmented reality is a research hotspot that develops on the base of virtual reality. Better man-machine interactive interface gives augmented reality a broader application prospect. Tracking technology is very important in augmented reality. Compared with tracking technology based on sensors, vision-based method is easy to realize, costs less and has a higher accuracy. The core work of tracking technology based on machine vision is image registration, which is the main concern of this dissertation.According to the application features in augmented reality, two novel methods of image registration are proposed in this dissertation: they are the image registration method based on artificial mark and image registration method based on natural features.In the image registration method based on artificial mark, a designed mark is placed in the scene in advance. The location information of a mark is obtained by iterative approximation of the inconsistency between the synthetic gray-scale image and authentic binary image. Finally different images are matched according to the location and direction of the mark as well as the selected points of it in each image. Compared with other systems, such as ARToolkit, this method is simple and solves the problem of location and recognition in the situation of partial occlusion and image blurring as well.Two classic algorithms SIFT and SURF are deeply analyzed. The feature points detection algorithm and feature points matching algorithm are improved for the need of image registration. We measure the robustness of feature points with their information amount, and proceed feature matching by feature subsets. Finally deny the false matches with RANSAC and least square method. The results show that the improved algorithm costs less time for calculating and achieves higher matching accuracy rate.