Register Based on Large Scene for Augmented Reality System

Abstract

Register is steadily gaining in importance due to the drive from various computer vision applications, such as augmented reality (AR), mobile computing, and human-machine interface. Efficient keypoint-based approachs are widely used in scene register. These approaches often model a scene as a collection of keypoints and associated descriptors, and then construct a set of correspondences between scene and image keypoints via descriptor matching . Finally, these correspondences are used as input to a robust geometric estimation algorithm such as RANSAC to find the transformation of the scene in the image. This paper focuses on designing a robust and flexible registration method for wide-area augmented reality applications. Firstly, we propose to partition the whole scene into several sub-maps according to the user’s preference or the requirements of the AR applications instead of building a global map of the wide-area scene.
Secondly a linear structured SVM classifier is used to perform scene learning online, which allows us to quickly adapt our model to a given environment. Finally, a hybrid tracking strategy is implemented by combining both wide and narrow baseline techniques. Some experiments have been conducted to demonstrate the validity of our methods.