Design And Implementation Of Tracking And Registration Algorithm In Augmented Reality Based On The Improved Ferns

An augmented reality (AR) system supplements with computer-generatedvirtual objects or digital information into the real world, which will enhance theperception of the real world and will bring a more intuitive visual experience. ARtechnology is one of the most hot research topic in the computer vision, and hasbeen widely used in the field of media, medical, military, education and others. Inthis paper, two parts will be included in our research, namely features trackingand3d registration, the main content is as follows:1. A detailed introduction is presented by studying and investigating intooverseas and domestic research status and the future direction of AR technology.2. This paper elaborates on the background theoretical knowledge involvedin AR technology, such as related technologies of digital image processing,features matching and computer vision.3. Improve the existing feature matching algorithm. The Random Fernsalgorithm, which is more mature, has been widely used in feature matchingalgorithm, and it is famous for its simplicity and high speed. However, due to itslarge volume obtained from training the classifier, it is unbearable for the lowmemory device and limits the range of application. To reduce the volume of aclassifier, this paper proposed a feature matching algorithm which is an improvedrandom ferns based on oriented information, named Oriented Ferns, OFerns.Using intensity centroid algorithm, Import the direction information into theFeature attribute set, so that the affect resulting by Rotation can be reduced.Through experiments prove, under the similar recognition accuracy, the size ofthe classifier obtained by training is reduced to the1/8to1/10compared with theoriginal algorithm, and it does not affect the speed of the algorithm. 4. Implement an AR system based on OFerns algorithm. The part of featuredetection adopts the Features from Accelerated the Segment Test (FAST)algorithm and image pyramid algorithm to quickly get multi-scale features.Thepart of feature matching adopts OFerns algorithm to obtain the correspondence offeatures, furthermore, combines with RANSAC algorithm on the disposition ofthe Mismatching problem. According to the correspondence, the homographicmatrix of key frames is computed, then process the three-dimensional registration,and be overlaid with virtual objects to a real-world scenario. The experimentalresults show that, this algorithm can identify the mark quickly and accurately, inthe case of partial occluding.