Development of deep learning architecture for automatic classification of outdoor mobile LiDAR data

Abstract

ABSTRACTThis paper proposes a deep convolutional neural network (CNN) architecture for automatic classification of mobile laser scanning (MLS) data obtained for outdoor environment, which are characterized by noise, clutter, large size and larger quantum of information. The developed architecture introduces a look up table (LUT) based approach, which retains the geometry of the input MLS point cloud while rescaling. Further, with the voxelisation of the input MLS sample, the ambiguity of selecting one out of multiple point values within a voxel is resolved. The performance of the architecture is evaluated on MLS data of outdoor environment in two instances, first using tree and non-tree classes (non-tree class has objects like electric pole, wire, low vegetation, wall, house and ground) and then with tree and electric pole classes. Additional testing is carried out by mixing the outdoor MLS data of tree and electric pole classes with three classes of indoor objects, taken from Modelnet dataset, thereby assessing the architecture efficacy over an ensemble of three-dimensional (3D) datasets. Classification of tree and non-tree classes, followed by tree and electric pole classes from MLS samples result in total accuracies of 86.0%, 90.0% respectively and kappa values of 72.0%, 78.7% respectively. Moreover, for the combinations of MLS and Modelnet classes, the classification results are promising, reaching a total accuracy of 95.2% and kappa of 92.5%. The LUT based approach has shown better classification over the traditional rescaling approach for the MLS dataset, resulting in an enhancement up to 9.0% and 18.0% in total accuracy and kappa, respectively. With different varieties of tree, non-tree and electric pole samples, the proposed architecture has shown its potential for automatic classification of MLS data with high accuracy. This study further reveals that the accuracy of classification is improved by introducing more spatial features in the input layer. The accuracies produced in this work can be further improved with the availability of better hardware resources.