Rapid Creation of Large-Scale 3D Models

Abstract : This STIR project focuses on the development of 3D shape matching and recognition techniques specially targeted for 3D modeling. The ultimate objective of our research is to develop effective methods for rapid creation of large-scale 3D models from 3D geospatial sensor data, and specifically LiDAR point clouds. Our technical approach is a novel alternative to traditional modeling approaches. The novelty arises from using the strategy of Modeling by Recognition (MBR) to rapidly identify objects from a 3D library of objects within point-cloud data. The recognized-object point clouds are then replaced with library data, such as polygon surface models, thereby constructing accurate and complete 3D scene models. Our research foci are the key components in the proposed modeling approach: the 3D shape matching algorithms that are used to detect objects of interest from point-cloud inputs and match them to model-library elements. We pursued shape matching algorithms in two ways: (1) Global primitive analysis for automatically detecting and extracting primitive shape geometry such as planes, cylinders and cuboids from point-cloud data, and (2) Local-feature techniques for representation of point-cloud features to produce unique 3D geometric descriptions for general 3D shape matching. We designed and implemented the approaches, and then evaluated them extensively with various datasets. Ultimately, these methods can become the core of a unified framework for automatic matching of point cloud data to a library of model components for creating both 3D models and object recognition/labeling.