Information extraction from large point cloud data : a deep learning approach
Hasan Asyari Arief · NORA - Norwegian Open Research Archives · 2020
Recent advances in Light Detection and Ranging (LiDAR) sensors have led to an increasing amount of large scale point cloud data collections. The LiDAR sensors can capture the fine spatial details of a remote environment in a full three-dimensional perspective, thus providing huge potentials for better machine understanding of a 3D scene. This thesis explores these potentials by providing robust and effective ways to extract information from large scale point cloud data. The study focuses on the utilization of deep learning techniques for the 3D scene understanding tasks, i.e semantic segmentation and object detection. It should be noted that the deep learning techniques were chosen mainly because the techniques simplify the generation of representative and robust features taking into account the spatial autocorrelation of input data, while often resulting in the highest prediction accuracies. As the backbone of this thesis, the deep learning approach has shown remarkable progress in generating the highest classification accuracy for several benchmark datasets, including our in-house dataset. Our contributions to improve the quality of point cloud annotation is closely related to the improvement of the deep learning models, i.e improving the deep learning preprocessing step by using a better density sampling approach, restructuring the deep learning modules by developing our Stochastic Atrous Network (SA-NET) architecture, and refining the post-processing step of deep learning prediction by invoking spatial and spectral similarities of point cloud data, using our Atrous X Conditional Random Field (A-XCRF) algorithm. The present PhD-work started by addressing some challenging problems regarding the modelling of the 3D point cloud data, and it was completed by providing a deliverable prototype capable of generating fast and accurate point cloud annotation labels. During the research process, we have managed to develop a better solution for extracting information in the form of semantic labelling from 2D projected point cloud data. We also developed a post-processing module refining point-level classifications directly generated from raw point cloud data. Finally, we developed an open-source and robust semi-automatic point cloud annotation tool, called Smart Annotation and Evaluation (SAnE). The SAnE speeds up the point cloud annotation process while also offering significantly better annotation accuracy than the baseline annotation approaches.