More Accurate Plane Recognition In Augmented Reality

Augmented reality can superimpose virtual objects into the real environment,enhancing consumers' perception and understanding of the real world.Augmented reality systems based on natural features have become the focus of research compared with fixed landmark maps in early recognition scenarios.There are many natural features,among which the plane is the most common and easiest to use.When the plane is recognized,the virtual object can be rendered to the plane through the three-dimensional registration technique.However,the plane recognition is not accurate enough,the plane feature points are less,and it takes a long time to divide the object on the plane.Therefore,this paper has done the following three tasks for the above points:(1)Based on RANSAC,this paper proposes a multi-planar fitting algorithn based on RANSAC.With the aid of color features,the contour information is used to find the plane range and filter the noise.The filtered points are clustered,and then the categories are mapped into the three-dimensional space.The plane is fitted to each class and finally divided according to the error rate.Weights,and then weighted the patches according to the normal vector and distance to obtain accurate plane parameters.The algorithm cannot only fit multiple models,but also get more accurate planes to improve the effect of augmented reality.(2)Based on the projection idea,this paper proposes an algorithm based on projection to generate feature points on the plane.The algorithm considers two adjacent frames,assuming that there is no expansion and contraction on the Z-axis,and because the adjacent two frames are quite close,the algorithm firstly projects the adjacent two frames according to the pose to the XOY plane to obtain a projection map,and calculates ORB feature int the projected image.And matching,according to the RANSAC algorithm,the mismatched pairs are filtered to obtain the pure matching,and the displacement of each point pair in the X and Y axes is calculated.Since there are still errors,we use the quartile to calculate the more accurate displacement.Points generate many new feature points.(3)Based on the idea of multi-threading,this paper is divided into front-end threads.The front-end thread is responsible for calculating the pose and rendering image.The back-end thread uses MaskRCNN to detect the pixel points of the object.Because the detection effect is poor when the image is blurred,the Laplacian gradient function is used to estimate the image.If the image is clear,the frame is stored in the array,and the adjacent clear key frame is selected.Since only one frame cannot obtain the three-dimensional point through triangulation,this paper uses the projection idea to infer the object pixel of another frame.And the world coordinates are obtained by triangulation,and then the object world coordinates are mapped to the pixel coordinate system of the current frame in real time,thereby real-time detection.