Research On Methods Of3D Registration Technique In Augmented Reality

The key of augmented reality is how to employ the computer techniques to effectivelyexpress and integrate the virtual and real worlds. The3D registration technique is the core forAR applications. Under the framework of cognitive science, we resolve the concepts andcharacteristics of augmented reality; then start from research on vision-based registrationmethods and introduction of some theories, we respectively improve three (fiducial marker-based,point cloud model-based, parallel tracking and mapping) AR registration methods, aiming toapply in different scenes. In summary, the main contributions of this thesis are listed as follows:1.Contrast to VR, the concept and characteristics of AR are defined in cognitivesignificance; a summary of the key techniques and applications are then given; we classify the3D registration methods into hardware-based, computer vision-based, and hybrid registration andanalyze the common problems of computer vision-based methods in detail; the main functions inmilitary are generalized at last.2. Fiducial marker-based registration methods are fit for scenes allowed to pre-add markers.Against to some problems such as poor-marked, yawp-sensitive and restriction of intrinsicparameters in traditional methods, we propose a refined fiducial marker-based one. Mainly weintroduce a homographic matrix to solve intrinsic parameters together when camera is initialized,so that the method fits for zoom lens; also we utilizes the edge detection and a distance fieldfrom the marker contour to the bounding box, which improves the recognition; Further, througha nonlinear least squares method to optimize the projection error and a parameter smoothing tooptimize the camera parameters, we can improve the accuracy of parameters calculating andregistration;3. In scenes where markers are not allowed or be turned back, we improve the so-calledpoint cloud model-based method which uses a key-frame and a candidate key-frame recognitionstrategy.When taking use of the structure-from-motion technique to pre-construct the3D pointcloud of the scene, key frames are selected from the input sequence which contain as manystable feature points as possible; during real-time3D tracking, the online image is matched to itscandidate key frames which are determined by fast image recognition method; at last a globalbundle adjustment is used to optimize the whole sequences. This refined method is proved to beeffective and good in real-time performance by experiment, also it is suitable for some scale ofnatural scenes.4. As the point cloud model-based method needs a time-consuming pretreatment and are-building when scenes are changed, which inevitably influence time performance, we proposea parallel tracking and mapping method based on SLAM and online SfM. It doesn’t need a pretreatment for environment, making use of parallel threads to track and map synchronously;Also we build a tracking-recovery strategy when quality is assessed lost to improve therobustness;what’s more, alternative to the global cluster adjustment,a partial one is used to savetime and improve accuracy. Experiments show that it is fit for small-scale desktop ARapplications,but not in open doors.