Research On RGB-D Feature Detection And Description Algorithm Of SLAM

Image feature matching is a basic problem of computer vision,which has been widely used in image stitching,SLAM(Simultaneous Localization and Mapping)and so on.Recently,RGB-D cameras have gradually become popular and been improved,making it easier to obtain both texture information and depth information.Two-dimensional features could not effectively express in the absence of texture information,while three-dimensional features perform poorly in scenes with insufficient geometric information.Therefore,there has been a new trend to fuse the texture information of RGB images with the geometric information of depth images to extract robust local features.This thesis proposes TEG(TExture and Geometric feature),an RGB-D feature detection and description algorithm towards SLAM systems,which is improved on RISAS(A Novel Rotation,Illumination,Scale Invariant Appearance and Shape Feature)and significantly improves its com-putational efficiency and performance.The main contents of this thesis are as follows:·TEG utilises a simplified Gaussian image pyramid to perceive the texture information ofdifferent neighbors,then applies the autocorrelation function to the Difference of Gaussian(DOG)maps to extract texture features.Compared to the multi-scale pyramid in SIFT(Scale Invariant Feature Transform),TEG senses the scale by depth images to reduce the number of layers in the image pyramid.·TEG calculates the gradient of point cloud by axes,then superimposes the gradient values asgeometric features.Compared with RISAS which calculats normal vectors of point cloud,TEG only performs difference operations here to improve the efficiency significantly.·TEG couples the plane normal vector of a keypoint to its neighboring points,so that it canperceive the spatial characteristics of the neighbor.·TEG has high computational efficiency and localization accuracy at low thresholds,so itcan satisfy the requirements of SLAM systems and has strong practicability.