3D terrestrial LIDAR classifications with super-voxels and multi-scale Conditional Random Fields

In this paper, we propose a new method for 3D terrestrial laser range data classifications. This functions as the first step towards virtual city model reconstructions from range data and is particularly useful for scene understanding. Classification of the outdoor terrestrial range data into different data types (for example, building surface, vegetation and terrain) is challenging due to certain properties of the data: occlusions due to obstructions, density variation due to different distances of the scanned object from the laser scanner, multiple multi-structure objects and cluttered vegetation. Also, the range data acquired are massive in size and require a lot of computation and memory. Recognizing the redundancy of labeling every individual data, we propose over-segmenting the raw data into adaptive support regions: super-voxels. The super-voxels are computed using 3D scale theory and adapt to the above-mentioned range data properties. Colors and reflectance intensity acquired from the scanner system are combined with geometry features (saliency features and normals) that are extracted from the super-voxels, to form the feature descriptors for the supervised learning model. We proposed using the discriminative Conditional Random Fields for the classification problem and modified the model to incorporate multi-scales for super-voxel labeling. We validated our proposed strategy with synthetic data and real-world outdoor LIDAR (Light Detection and Ranging) data acquired from a Riegl LMS-Z420i terrestrial laser scanner. The results showed great improvement in the training and inference rate while maintaining comparable classification accuracy with previous approaches.