Broad Learning Based 3D Point Cloud Object Classification

As 3D point cloud data were wildly used in various fields,how to recognize 3D point cloud objects efficiently and accurately had become one of the research focuses,especially in fields like unmanned vehicle which highly demanded for real-time object classification.Due to the characteristics of uneven distribution and unstructured data format of point clouds,it was difficult for traditional deep learning networks to consume point cloud data directly.Besides,the complexity of network structure and the huge parameters in the network also had negative effects on the efficiency of deep learning networks.Although the broad learning system had reduced the time cost,it cloud not process unstructured point cloud data as well.In addition,there are few researches about its application on 3D point cloud area.In this paper,a 3D object classification network was developed which consisted by two deep learning networks as feature extractors and a broad learning network as the classifier.The main contributions of this paper are as follow:(1)In this paper,a fast 3D object classification network architecture was developed.The proposed network introduced two deep learning networks as feature extractors to extract voxel features and normal features from point cloud data.Then,those structured features were inputted into the broad learning classifier for 3D object classification.The proposed method solved the problem that broad learning network cloud not process point cloud data directly;(2)This paper analyzed the relationship between object classification accuracy and the network architecture of broad learning network.The effects of three strategies for adjusting the broad learning network structure,increasing nodes in the enhancement layer,increasing window number in the feature layer,and increasing nodes in each window in the feature layer,on classification accuracy were tested and discussed.Besides,this paper also analyzed the relationship between the time consumption and the number of nodes in the broad learning network.Based on above experiment results,this paper gave some suggestions about adjusting the number of nodes in the broad learning network,which laid a certain foundation for the future research of broad learning network;(3)The proposed 3D object classification network was robust to noise.This paper proved three adjusting strategies,including increasing nodes in the enhancement layer,increasing window number in the feature layer,and increasing nodes in each window in the feature layer,were robust to Gaussian noise and white noise.This paper gave advice about adjusting network structure to improve the performance when dealing with noise data.