Abstract: Airborne LiDAR point cloud classification is one of the key steps for three-dimensional reconstruction of urban scenes. To leverage the existing high-performing deep learning network model in image field of image processing, improve classification accuracy and reduce training time and demand for training samples simultaneously, an airborne LiDAR point cloud classification method based on deep residual network is proposed in this paper. Firstly, high discriminative low-level features, i.e. normalized height, point cloud normal vector, intensity and normalized differential vegetation index, are extracted. Secondly, by setting different neighborhood sizes and perspectives, multi-scale and multi-view point cloud feature images are generated via using the proposed point cloud feature image generation strategy. Then, point cloud feature images are input into the pre-trained deep residual network to extract multi-scale and multi-view deep features. Finally, a neural network classifier is constructed and trained, point cloud classification results are obtained by utilizing the trained classifier and postprocessing. Benchmark datasets of ISPRS 3D Semantic Labeling Contest are used, the experimental results show that the proposed method can effectively distinguish 8 types ground objects such as buildings, ground and vehicles etc., and the overall accuracy of the classification result is 87.1%, which can provide reliable information for three-dimensional reconstruction of urban scenes.

Key words: point cloud classification, deep feature, multi-scale and multi-view, transfer learning, deep residual network, airborne LiDAR