3D Point Cloud Object Recognition Method Based On B-spline Surface

With the opening and development of the artificial intelligence era,artificial intelligence technology has greatly promoted the development of technologies such as speech recognition,image recognition,and NLP,which have extensively and profoundly affected industries,scientific research,medicine,and various social life-related issues.Many fields.As the core technology of the current artificial intelligence — Computer vision technology,the main research problems and methods proposed mainly revolve around the reconstruction,segmentation and classification of graphics and images and identification problems.At present,with the trend of deep learning as the standard in the field of image,the research and application requirements of 3d point cloud related technologies are also gradually strengthened,and the classification and recognition based on 3d point cloud has gradually become the key research direction to be solved in the field of computer vision.Based on the research status and development trend of computer vision research in home and abroad,this paper proposes a three-dimensional point cloud object recognition method based on B-spline surface(BSR)for object recognition in complex scenes,the specific research work is as follows:Firstly,point cloud fitting based on b-spline parametric surface was used to describe the geometric shape of the object,and parameter resampling was carried out on the fitted surface to obtain the point cloud with regular distribution.Secondly,a local geometric feature description consisting of principal curvature,principal direction and normal vector is defined,and point cloud features are extracted and matched.Then,isometric calculation and classification of matched point pairs are carried out.By comparing isometric difference of features of similar point pairs under isometric,unsupervised spectral clustering algorithm is applied to point pairs with approximately same isometric transformation to obtain similar regions constructed by similar point pairs.Based on this,similar parts between objects are detected.Finally,similarity measurement is carried out and the proportion of similarity point cloud and sampling point cloud under different isometric classification is counted to obtain a measurement value of similarity degree so as to complete object shape recognition.The point cloud surface fitting reduces the tedious preprocessing steps,and the local geometric features defined have high efficiency and robustness,which improves the recognition efficiency of the algorithm as a whole.The calculation of space transformation such as isometric classification of point-to-features improves the classification effect and recognition accuracy of spectral clustering algorithm.The experimental results show that the proposed algorithm is tested under the public datasets of Princeton,Mcgill and Tosca.The average recognition rate of object shapes can reach 96%,94% and 93% respectively,throughthe isometric transformation and noise of the original point cloud data.After the downsampling process,the effects of real scene and perspective changes are simulated,and the effectiveness and robustness of the proposed method are fully verified.Through a large number of tests and performance comparisons of these three data sets,it can be concluded that the recognition method(BSR)of this paper can not only solve the problem of point cloud classification and recognition of objects,but also change the angle of view after the selection and adjustment of the optimal threshold.The high-precision recognition problem of the shape difference of the objects within the class also has a good discrimination.