Research On Point Cloud Registration Technology Based On Shape And Texture Fusion

With the continuous development of digital technology,three-dimensional point clouds have attracted more and more researchers’ attention due to their dimensional advantages,and have been widely used in many fields such as terrain survey,industrial measurement,virtual reality,and robotics.At present,due to the rapid development and popularization of threedimensional laser scanning technology,the acquisition of point cloud color information has become very convenient.The information contained in it,such as texture or surface reflection intensity,provides space for the further development of point cloud technology.Due to the current point cloud registration algorithms mostly focus on geometric information while ignoring the use of color and texture information,there are shortcomings in registration accuracy or efficiency,and they are easily affected by the initial position and posture of the point cloud.In response to the above issues,this paper introduces color and texture information to improve the effect of point cloud registration,mainly studying keypoints extraction,local feature description,and registration algorithms for point clouds that integrate shape and texture,in order to improve the accuracy of point cloud registration and maintain a certain processing speed.The main research work of this article is as follows:(1)Aiming at the problem that low curvature point cloud models are difficult to detect keypoints,a covariance matrix based point cloud keypoint extraction method is proposed,which simultaneously extracts geometric and texture significance keypoints of point clouds.Firstly,using the ISS(Intrinsic Shape Signatures)algorithm with good efficiency and repetition rate to extract keypoints based only on geometric features;Secondly,using less relevant point cloud HSV color space information,combined with point cloud spatial distance weights,a color covariance matrix is constructed within the local neighborhood,and texture saliency is measured based on generalized variance.Candidate keypoints are extracted.Finally,the Non Maximum Suppression(NMS)algorithm is used to filter out the most prominent keypoints in the local region.Experiments show that the algorithm can extract geometric salient points and texture salient points simultaneously,with better repeatability.(2)To solve the problem of insufficient feature description ability and time-consuming due to the lack of color and texture information in existing point cloud local descriptors,a multi feature fusion descriptor based on FPFH(Fast Point Feature Histogram)is proposed.Firstly,FPFH algorithm is used to extract shape features;Secondly,a topological structure is established for the point cloud within the feature neighborhood,and texture features are extracted using the ratio of HSV color channels between point pairs;Finally,a descriptor is constructed through feature fusion,and feature matching is performed based on the nearest neighbor ratio strategy to evaluate performance.Experiments show that this algorithm effectively reduces the computational complexity compared to other descriptors,and can improve the efficiency and accuracy of color point cloud feature matching.(3)Aiming at the problems of poor registration effect and low accuracy of existing point cloud registration algorithms when the geometric features of point clouds are not significant,a point cloud registration method that combines shape and texture features is proposed based on improved keypoint extraction and improved feature description algorithms.Firstly,keypoints with geometric and texture significance on the surface of a point cloud are extracted,and then the shape and texture of the keypoints are characterized.Keypoints are matched based on feature similarity.Then,Random Sample Consensus(RANSAC)algorithm is used to eliminate mismatches and estimate the pose matrix to achieve coarse registration,providing good initial values for subsequent fine registration,Finally,a color ICP registration algorithm is used for fine registration.Experiments show that the algorithm has good registration results on color point cloud models with clutter,low overlap rate,and insignificant shape characteristics.