AR Based on ORB Feature and KLT Tracking

Currently Augmented Reality (AR) simply based on natural features is not efficient. In this paper, we propose a novel simple method of AR registration using Oriented FAST and Rotated BRIEF (ORB) features and Kanade-Lucas-Tracker (KLT) tracking algorithm. First we extract ORB natural features from a reference image and the first video frame. Select the matching points using appropriate Hamming distance and Random Sample Consensus (RANSAC) algorithm. Then we track the matched ORB features of the next video frames with an updating KLT strategy. Finally we calculate and determine the camera pose using the matched feature point sets. Our experiments show that, in the cases of different scales, different angles, some changes in ambient light, complex backgrounds or part of the reference image blocked, our AR system performs well to track and position accurately. And the processing speed of our system approximately meets the requirement of real-time system.