Scene Perception And Localization For Augmented Reality

With the development of technology and hardware,augmented reality has become more and more widely used in life,especially for mobile augmented reality applications.In order to meet the needs of various augmented reality scenarios,vision based scene perception and localization has become the core of augmented reality R&D,more and more researchers are investing in.This thesis is mainly based on the mobile platform,focusing on the construction of a set of synchronized rendering Unity3 D augmented reality framework and the research of three different forms of perceptual tracking and localization methods to meet the needs of different augmented reality scenarios.The main contributions of this thesis are as follows:1.This paper presents a fusion detection and tracking method for marker localization.This method not only guarantees good real-time performance at the mobile end,but also guarantees the accuracy and stability of the localization.It can also effectively perform tracking and localization in the case of severe tilt,occlusion,and scale changes.Compared to the feature point matching method,our method adds the information of the time context,which makes the localization result smoother and more stable.By comparing experiments,the robustness of the proposed algorithm is also verified.2.This paper proposes a multi-plane sensing extended tracking method with a fast initialization of SLAM based on marker images.This method not only can directly initialize the SLAM in a single frame,but also can introduce real scale information,so that the position and posture of the localization have absolute scale reference value,and the accuracy can achieve centimeter-level.The multi-plane perception can detect horizontal planes,vertical planes,and slopes in space,enhancing the spatial perception of augmented reality.Through rich scene testing and comparative experiments,the augmented reality of this thesis has achieved very good results and experiences.3.This paper presents a natural contour tracking method and uses the contours to accurately locate the position and attitude of a 3D object.This method can accurately extract the 2D natural contour of object in videos.Through the comparison experiment of public data sets,the recall and the precision of the algorithm are verified.At the same time,the contour information can be used to effectively perform high-precision pose estimation on textureless 3D objects in augmented reality scenes,for satisfying the requirements of augmented reality manuals.In order to achieve the final augmented reality effect,this thesis builds a framework for augmented reality development based on the Unity3 D cross-platform game engine.The framework guarantees synchronous rendering,particle effects,background video rendering,and real-time shadow rendering.At the same time,the interface is encapsulated,and it can access the perceptual and localization algorithm libraries under different platforms.