Fast Scenes Recognition And Tracking For Mobile Augmented Reality Applications

Advances in hardware technology, specially the constant increase in the processingpower of mobile phones have turned smart phones into a powerful platform for augmentedreality applications. Compared to PC, neither computing power nor memory space of theresource-constrained mobile phone can well satisfy the requirements of mobile augmentedreality applications. Currently, the common strategy to solve the above problems is to useoffline reconstruction and registration. Online registration consists of scenes recognitionand online registration where real-time performance and accuracy are two difficulties needto be considered.As the traditional SIFT algorithm cannot meet real-time performance on mobilephone, with the use of fast corner detector, lookup table and descriptor compressiontechniques to accelerate the descriptor generating and matching process, a method forfast descriptors generating and matching is realized for mobile devices. In order toreduce the massive distance computation resulted by the exhaustive process and theasymmetric distance computation in the origina l transform coding method, we dividethe original vector into sub vectors to accelerate the asymmetric distance computationprocess. Moreover, the original exhaustive search process is divided into three parts:coarse search, fine search and re-ranking, which can reduce the number of retrievalvector. In order to realize real-time registration, the registration is divided into two parts:camera pose initialization and camera tracking. In the stage of camera poseinitialization, improved sift and guided matching is used to get initial camera pose. Inthe stage of camera tracking, optical flow tracing is used to track natural scenes.Experiments show that the proposed scenes recognition algorithm can providereasonable search performance and reduce the memory usage on mobile phonessignificantly, which facilitates the vision based scenes recognition in mobile augmentedreality applications to a degree, meanwhile, the improved registration method cantrack non planar scenes at frame rates of up to10Hz on current generationphones(1GHz), which solves the real-time problems of registration in mobileaugmented reality to a degree.