Densifying The Point Cloud For 3D Reconstruction Based On Iterative Derivation

3D reconstruction based on image sequence is a classic research problem.The accuracy of the final model obtained by this method depends largely on the quality of the point cloud.The quality of the point cloud is determined by the accuracy and quantity of the points.The method widely used in image sequence reconstruction is to restore the three-dimensional scene structure from motion information.The method is based on the feature extraction of the image,and then the adjacent images are matched to find the corresponding matching feature point pairs.According to the matching point pairs to estimate the conversion relationship between the two adjacent cameras.The matching feature point pairs can restore the point in three-dimensional space.This method is severely limited by the matching operation and the number of the feature point pairs.If there is no matching point pair on the target object,the surface information of the corresponding region will not be obtained in the point cloud.In the three-dimensional reconstruction based on image sequence,a large number of point cloud densification algorithm increase the number of point cloud by analyzing the characteristics of the image.But the focus of these densification algorithm is to propose a better feature extraction method or feature matching algorithm.Reconstruction of the very sparse target is still a problem that needs to be overcome in 3D reconstruction.In this paper,we mainly consider the reconstruction of the target object with sparse characteristics,and propose an algorithm to improve the density of point cloud.The goal of the algorithm is to reduce the dependency of the image feature points in the process of densifying point cloud.Improve the density of the point cloud by analyzing the information such as the contour provided by the image,the algorithm can get a better point cloud for the sparse target.The algorithm has better universality.The main work of this paper contains the following five aspects.First,the algorithm tries to derive the initial sparse point cloud of the input.The process of derivation takes the point in initial point cloud as the center of a cube,and derive new points around the corners of the cube.Deriving new points for each point in initial point cloud,and obtaining a new derived point cloud.Second,the algorithm proposed a point-based tangent plane constrain to filtered the new point cloud.The filtering operation filters out the outliers in the new point cloud and error points.Thirdly,the algorithm proposes a filtering operation based on contour constrain to filter the invalid derived points which locate outside of the target surface.Fourthly,on the basis of the first two filtering operations,an algorithm is proposed to analyze the outer envelope of the new points.The filtering operation will filter out the invalid new points which locate inside of the surface.Fifth,The derivation and filtering operations can be performed iteratively,and the derived point cloud can be used as the initial point cloud to derive again.After several iterations,the final point cloud is obtained.In summary,this paper divides the process of densifying point cloud into two steps,the derivation operator and filtering operator.By trying to derive the new points and filtered the invalid points according to the image contour,the algorithm is effectively separated from the dependence of the feature points in the image.The algorithm can obtain the dense point cloud of the sparse target object.The approach has been tested on several datasets,and the results demonstrate the good performance of proposed method.