A framework for automatic classification of mobile LiDAR data using multiple regions and 3D CNN architecture

Abstract

ABSTRACT This paper proposes a framework for automatic classification of mobile laser scanner (MLS) point cloud using multi-faceted multi-object convolutional neural network (MMCN). The proposed method takes a full three-dimensional (3D) point cloud as input and outputs a class label for each point. Unlike other existing classification methods for MLS data, the proposed method is not dependent on any parameter or its tuning. The proposed MMCN uses multiple objects of a sample, defined by different sizes of the sample, in addition to the different facets obtained by rotating about the various axes, thus adding more information during the training and testing stages. The proposed framework uses manually extracted samples for training the MMCN. Automatically extracted multiple regions based on the various radii of spherical neighbourhoods around MLS points are passed through the trained MMCN for determining their probabilities of belonging to different classes. The class probabilities of different sized regions are then used as a feature vector to train a support vector machine (SVM), and the final decision for the class of a point is based on the SVM output. The proposed framework has been trained for five classes, viz., Ground, House, Pole, Tree, and Car and has been tested on Oakland and Paris-Lille 3D MLS datasets. The total accuracy and kappa coefficient (κ) reach up to 96.5% and 93.8%, respectively, for the framework. The MMCN together with the SVM is able to achieve parameter-free classification of MLS data, thereby eliminating the need for manual parameter tuning as in the existing methods. Therefore, besides the use for classification of MLS data for mapping purpose, the approach is also suitable for classification of light detection and ranging (LiDAR) data resulting from autonomous vehicle sensors. The accuracy of this work can be further improved by incorporating more and varied training samples and deeper convolutional neural network (CNN) with better hardware resources.