A Feature Point Tracking Method Based On The Combination Of ORB And Klt Algorithm About The Augment Reality

AR systems support the coexistence of real elements (that are part of users’ world) and synthetic ones (computer generated) in the same environment. Nowadays, this kind of user interface has obtained more attention due to the fact that it allows users performing tasks in a more intuitive, efficient and effective way. Interactive AR inter-faces may augment users’ perception of the real world by adding virtual information to it. In addition, users’ interaction with the system may be augmented, making the computer as seamlessly as possible, by exploring the use of real objects for interaction with the application.In AR the technical challenges lie in determining, in real-time, what should be shown where, and how.Ideally, AR proposes that the user must not be able to distin-guish between real and virtual information, demanding that the virtual elements show both geometric and photometric consistency.The problem related to correctly position-ing virtual information relative to the real environment, called registration, is solved by tracking the environment so that the synthetic elements can be adequately registered with the real scene.They capture features from the real world, and based on this infor-mation the AR system determines when, where and how the virtual scene should be exhibited.Currently there exists several schemes for the implementation of augment reality. This article explains the theory and implements the augment reality system in two ways:recursive tracking and tracking by detection, and we compare the systems from two aspects:the processing time and the stability. At present, most of the schemes implementing the augment reality is based on tracking by detection:extracting the feature descriptors from every frame, and match-ing the descriptors with feature descriptors of reference picture. Then we could get transform matrix from these corresponding points using algorithms such as Pnp. The coordinates of the virtual objects can be calculated at last. This type of implementation is simple and we use the same solution to each frame in the video stream. So there is no cumulative error, and each time the results are independent. But there are some flaws in system stability due to the descriptors can not match correctly.This article implements the system using the method of matching ORB feature descriptors, and optimizes the match results using RANSAC algorithm. This system could recover from error situa-tion, but is worse than systems based on recursive tracking in the stability.Based on the above facts, this article presents an algorithm based on recursive tracking.The recursive tracking, where the previous pose is utilized as an estimate to calculate the current pose. This article implements the system using the method based on the combination of ORB and KLT algorithm. First, we extract the feature descrip-tors from the first frame and the reference picture. Then we tracking these descriptors in the following frames to get the transform matrix between the frame and reference picture, we can get the coordinates of virtual objects more convenient. This algorithm is to track the specified feature points between frames, so there is no feature points matching problem, which also eliminates the jitter of virtual objects. This solution based on tracking is more quick than solution based on extracting and matching. But KLT tracking algorithm requires a high relatively inter-frame correlation, if the differ-ence between frames is big, when the camera perspective moves fast, then the tracking of the feature descriptor will be lost. This issue needs to be improved in future work.Based on these algorithms, this paper tests these systems and compares the test results. The result shows that the proposed algorithm has better performance and sta-bility than the systems based on the tracking by detection. Finally, we summarizes several aspects to improve the proposed algorithm:the processing speed of the KLT tracking algorithm, optimizing the tracking window so that it can handle the frames of the lower correlation, in order to adapt to the fast-moving perspective of the camera.