| Vehicle LiDAR point cloud data has characteristics of high density,complex scenes,rich targets,occlusions,transmission,and noises,which results in the low accuracy in the point cloud classification and the low level of automation in automatic recognition of 3D scene objects.To address the issues raised from the state-of-the-art about the point cloud classification,we in this paper follow the previous strategies,such as semantic rules,machine learning and deep learning,and proposes a deep belief network model,where the extracted features was used as input.Finally,Three-dimensional scenes are classified into six categories: grounds,low vegetations,high vegetations,buildings,power lines,and pole-like objects.The main research content is as follows:First,LiDAR color point cloud data over urban areas are voxelized.On the basis of the analysis of point cloud data characteristics and the summarization of prior work,the geometric features and spectral features of clusters within each voxel are extracted for a 24-dimensional voxel feature vector,which will serve as the input of the subsequent deep learning model.Then,two types of deep belief networks,i.e.,back propagation DBN and associative memory DBN,are constructed for automatically classifying the point cloud data.In our implementation,we investigate the effect of the number of network nodes on classification,the effect of the number of iterations on classification,and the effect of the size of voxels on classification.Moreover,the other frequently-used machine learning models are compared to the proposed method.Experimental results show that both back propagation DBN and associative memory DBN achieved good classification accuracy of more than 90%.In addition,experimental results also suggest that back propagation DBN provided a better classification result than associative memory DBN.Also,compared with the other frequently-used machine learning models,we concluded that both networks obtained higher classification accuracy. |
PDF Full Text Request