Research And Application On Monocular Stereo View Reconstruction Technology In Augmented Reality

Augmented reality(AR) aims to superimpose the virtual objects generated by computeron the real scene indistinguishably to enhance the user feeling of the real world. Thistechnology is based on user’s direct perception of the real scene and place the virtual objectsgenerated by computer graphics&image and visualization technology in the real scene byaccrute registering technology. Then a real-world perception fusion environment ispresented to the user with the help of display device. Immersion is one of the main featuresof AR, which can not be fulfilled through traditional2D display technology while doingfusion of virtual and real scene or human-computer interaction. At the same time, applyingbinocular stereo display technology to AR directly will cause a series of problems such ascomplex calibration process an wide baseline. Monocular stereo view reconstructiontechnology converts2D video into3D video, which is an effective way to solve theseproblems. This paper makes a deep research on three aspects of the AR monocular stereoview reconstruction technology which is depth extraction, stereo view reconstruction anddepth based AR occlusion. The main contributions and innovations of the thesis are asfollow:1．Aiming at solving the problem of reducing the depth extraction error of smoothforeground in defocus image, an algorithm to generate the depth map with a single2Dimage based on multi-scale wavelet is presented in this paper, which is suitable for bothdefocus and wide angle image. Pixel classification techniques is used to do depth correctionin this algorithm, which can not only reduce the depth extraction error of smoothforeground in defocus image but also is suitable for wide angle2D image depth extraction.The first step of the algorithm is to extract depth information by proposing multi-scaleadaptive wavelet analysis method to generate depth map and the second is optimizing thedepth map using region grow and edge segmentation. Meanwhile, the fast zerocount methodand multi-scale acceleration method are proposed to accelerate the calculation.2．A new non-hole-filling DIBR method is presented in this paper, which aims toreduce the time complexity of hole filling process of traditional DIBR method. Startingfrom analyzing the causes of hole generation, this paper improve the traditional3D-WARPING algorithm based on pixel interpolation to simplify the process of traditionalDIBR so that can meet the requirements of real time high resolution image processing. Themethod first uses depth expansion algorithm to do depth preprocess, and then a dynamicinverse interpolation algorithm of3D-WARPING method is proposed to generate virtualviewpoint images directly, thus eliminating the time consumption of hole filling.3．The traditional AR occlusion system is restricted by its low efficiency and highconditionality. So this paper design and make an depth-map based AR occlusion system byusing the Microsoft Kinect. The depth information captured by Kinect is applied to the keyscheme of the traditional model-based AR occlusion system to simplified it to a virtual reality based occlusion problem, thus improve the real-time performance and the occlusioneffect. The main idea is to put the real scene and virtual scene together into unified worldcoordinate system, than do occlusion display using computer graphics technology. In thispaper, the3D reconstruction algorithm based on depth map together with the depth aidedhead pose tracking algorithm for view point estimation are proposed. Moreover, the Kinectdepth map preprocessing algorithm is proposed to improve the quality of depth information.Experiments demonstrate that the depth maps obtained by the multi-scale waveletanalysis depth extraction algorithm is more accurate and consistent both in foreground andbackground. Moreover the acceleration optimization algorithm can meet the requirementsof standard video real-time processing. On the other hand, the virtual viewpoint imageobtained by the None-Hole-Filling DIBR algorithm which is proposed as a new solution ofstereo view reconstruction has higher subjective visual quality compared with the traditionalDIBR approach. Meanwhile the algorithm can meet the requirements of high-definitionvideo real-time processing. Finally, the proposed AR occlusion system based on depthinformation has high occlusion accuracy and good real-time performance. And thereasonable price make it a promising application in the future.