A supervoxel-based spectro-spatial approach for 3D urban point cloud labelling

Three-dimensional (3D) point cloud labelling of airborne lidar (light detection and ranging) data has promising applications in urban city modelling. Automatic and efficient methods for semantic labelling of airborne urban point cloud data with multiple classes still remains a challenge. We propose a novel 3D object-based classification framework for labelling urban lidar point cloud using a computer vision technique, supervoxels. The supervoxel approach is promising for representing dense lidar point cloud in a compact manner for 3D segmentation and for improving the computational efficiency. Initially, supervoxels are generated by over-segmenting the coloured point cloud using the voxel-based cloud connectivity algorithm in the geometric space. The local connectivity established between supervoxels has been used to produce meaningful and realistic objects (segments). The segments are classified by different machine learning techniques based on several spectral and geometric features extracted from the segments. All the points within a labelled segment are assigned the same segment label. Furthermore, the effect of different feature vectors and varying point density on the classification accuracy has been studied. Results indicate an accurate labelling of points in realistic 3D space conforming to the boundaries of objects. An overall classification accuracy of 90% is achieved by the proposed method. The labelled 3D points can be used directly for the reconstruction of buildings and other man-made objects.