Research On Filters And Feature Points Extraction Technology Of 3D Point Cloud

With the developing of computer-aided design technique,target recognition of3 D point cloud has become a hot topic in the field of computer and pattern recognition.But during the process of obtaining point cloud data by 3d scanning equipment,the acquired point cloud is affected by noise,target occlusion and point cloud density to some extent,which will decrease the difficulty and accuracy of target recognition.In order to resolve the problems occurred in the process of target recognition,in this paper,the filtering and key point extraction of 3D point cloud are improved.In terms of filtering,first calculate the k-nearest point field of each point in the model,then calculate the Gaussian curvature of each point in the nearest field and the average Gaussian curvature of all points,and finally compare the Gaussian curvature and the average Gaussian curvature of each point.The points in the model are divided into two categories,one is a point set larger than the average Gaussian curvature,and the other is a point set smaller than the average Gaussian curvature.A set of points smaller than the average Gaussian curvature is regarded as a relatively small change in surface features in the model,that is,a relatively smooth surface of the model,which is filtered by a conventional median filtering method.A point set larger than the average Gaussian curvature is regarded as a region where the feature transformation is more obvious in the point cloud model,and the improved multilateral filtering method is used for filtering.Among them,the improvement of the multilateral filtering is mainly reflected in: in order to make the distance between the sampling point and the point in the neighborhood have a small effect on the noise reduction effect,the weighting function in the multilateralnoise reduction algorithm is weighted.Therefore,the filtering effect of the point-to-point cloud data in the relatively far distance from the sampling point is smaller,and the filtering effect is improved.In terms of key points extraction,SIFT3D、Harris3D、ISS3D、NARF key points methods which have a good recognition are studied in this paper.And through the way of experimental comparison,the advantages and disadvantages of each key point extraction method are analyzed.In order to combine the advantages of various key point extraction methods,this paper proposes an extraction algorithm combining multiple key point extraction methods,which realizes the extraction of key points.Experiments show that the key point extraction method proposed in this paper has good robustness to both noise and model rotation transformation.In this paper,a lot of experimental are carried out on the improved filtering method and key point extraction method.From the experimental results: In terms of filtering,the improved filtering method is very robust to noise in the point cloud;in the key point extraction,the improved key point extraction method is robust to noise and scale transformation of the model.Finally,in order to verify the effect of the improved algorithm in the target recognition process,in the last chapter of this paper,a lot of verifications are carried out under different noise intensities and model rotation transformations.