Study On Real-time Scene Recognition And Registration Of Markerless Augmented Reality

Augmented reality is a hot research field of computer graphics.It has formed a relatively complete theoretical system by a long period of research and development.With the development of cameras and screen technology and the rapid increase in hardware computing capabilities in recent years,augmented reality technology has been applied to medical,education,military,and other fields,which has a tremendous impact on human life.As one of the research hotspots in the field of augmented reality at present,markerless augmented reality technology maintain the precise alignment relationship between the virtual information generated by the computer and the real scene by recognizing and tracking the feature point information in the scene.Scene recognition and tracking and registration are two issues that need to be studied.In this thesis,we have carried out further research in the filed on the basis of previous study,and the mainly research works are as follows.For the problem that the scene recognition algorithm based on Gaussian scale space is easy to cause boundary blur and detail loss,which resulting in low recognition accuracy,an adaptive fast scene recognition algorithm based on nonlinear scale space is proposed.Firstly the algorithm constructs the nonlinear scale space to solve scale invariability problem,and selects the number of scale space groups adaptively according to different images to ensure the edge details.Then it extracts the FAST feature in the scale space and uses the BRIEF descriptor to describe the feature to improve the speed of feature detection.Finally,the algorithm achieves scene recognition through feature matching and target searching.The experimental results show that the algorithm has strong robustness in view point,scaling rotation and JPEG compression transform and it achieves higher recognition accuracy while achieving fast scene recognition.For the problem that the traditional tracking and registration algorithms cannot adapt to different scene characteristics,which resulting in low tracking accuracy and unstable registration effect,a tracking and registration algorithm based on adaptive parameters and pose confidence is proposed.Firstly the algorithm improves the matching method of ORB algorithm feature descriptors.It adjusts the descriptor matching distance thresholds adaptively according to the image characteristics of different scene images after using binary strings to describe the image feature points,so as to obtain a reasonable number of matching point pairs.Then in the stage of camera pose estimation,the accuracy of the camera registration matrix is determined by quantifying the pose matrix confidence,and the algorithm avoid using the registration matrix with low accuracy to three-dimensional registration.The experimental results show that the algorithm can effectively improve the accuracy of tracking registration and has strong robustness to viewpoint change,illumination change,and rotation scaling.