Research On Outlier Detection For Reconstructed Point Clouds Based On Images

Compared with the traditional modeling method or reconstruction with a scanner,modeling based on image is not only simple and fast,but also extremely low cost,so it has a wide application prospect.Because the image contains a lot of information about the geometric scene,it is possible to recover the 3D geometry according to a set of images.Image based modeling method has a wide range of applications in many fields,such as games,medicine,archaeology and so on.However,due to complex surrounding environment,the number and angles of input images,and the changes of reflectivity or color from the surface,it is inevitable for reconstructed point clouds to contain many outliers which don't belong to the surface of modeling object.Since the formation of outliers is influenced by many factors,the distribution of outliers is not a specific state,which is random and scattered.From the point of view of local density,outliers in reconstructed point clouds can be sparsely distributed or densely clustered.From the location of distribution,some outliers locate in a far distance from the main body of model,and others surround the model or connect to the surface.Such erroneous data points pose some problematic issues to the applications of the reconstructed point clouds.For instance,it will cause roughness surface or model deformation when taking the surface reconstruction directly with these original point clouds.In order to obtain a more accurate 3D model,the original point clouds need to be processed before surface reconstruction.Common manual denoising process is not only time consuming,but also heavily dependent on personal experience.Therefore,it is urgent to find a robust outlier detection and filtering method.This paper focuses on the outlier detection of point cloud which is reconstructed based on images.The main work contains the following three parts:1.In this paper,the method of 3D reconstructed point clouds based on images,the forming process of the point cloud model,the forming reason and the distribution characteristics of the outliers have been discussed.The existing outlier detection algorithms at home and abroad are summarized,and the advantages and disadvantages of the existing algorithms are analysed.2.Aimed at the outliers around the main body of model,a new denoisingalgorithm based on neighborhood expansion clustering is proposed.According to the Euclidean distance between data points and the transitivity of the neighborhood location relation,the algorithm searches for neighboring points of each data point,and expands them to find a largest cluster,which can detect and filter the outliers in the point cloud model.The concept of neighborhood expansion clustering and the fast search algorithm based on dynamic grids division are mainly discussed.Simulation results show that the proposed algorithm can effectively filter the outliers in the isolated and dense distribution of the point cloud model,and improve the denoising efficiency of traditianal k-nearest neighbor algorithm for the point clouds3.Based on the distribution characteristics of outliers,an integrated outlier detection method is proposed.Firstly,the spatial topological relation of point cloud data is established by dividing cell method,then the maximum connected domain is sought,hence the outliers far away from the main body of model be detected and removed.When the model surface sampling is insufficient,it can form a smaller cluster which is separated from the main body of the model surface,the boundary matching method can be used to preserve the effective points and remove outlier clusters.Finally,the improved K-means algorithm is adopted to cluster analysis for all point data,and removes the outliers which are connected with the model surface.Simulation results show that the proposed algorithm can effectively detect the multiple distribution state outliers in the reconstructed point cloud.