Graph-based segmentation and scene understanding for context-free point clouds

The acquisition of 3D point clouds representing the surface structure of real-world scenes has become common practice in many areas including architecture, cultural heritage and urban planning. Improvements in sample acquisition rates and precision are contributing to an increase in size and quality of point cloud data. The management of these large volumes of data is quickly becoming a challenge, leading to the design of algorithms intended to analyse and decrease the complexity of this data. Point cloud segmentation algorithms partition point clouds for better management, and scene understanding algorithms identify the components of a scene in the presence of considerable clutter and noise. In many cases, segmentation algorithms operate within the remit of a specific context, wherein their effectiveness is measured. Similarly, scene understanding algorithms depend on specific scene properties and fail to identify objects in a number of situations. This work addresses this lack of generality in current segmentation and scene understanding processes, and proposes methods for point clouds acquired using diverse scanning technologies in a wide spectrum of contexts. The approach to segmentation proposed by this work partitions a point cloud with minimal information, abstracting the data into a set of connected segment primitives to support efficient manipulation. A graph-based query mechanism is used to express further relations between segments and provide the building blocks for scene understanding. The presented method for scene understanding is agnostic of scene specific context and supports both supervised and unsupervised approaches. In the former, a graph-based object descriptor is derived from a training process and used in object identification. The latter approach applies pattern matching to identify regular structures. A novel external memory algorithm based on a hybrid spatial subdivision technique is introduced to handle very large point clouds and accelerate the computation of the k-nearest neighbour function. Segmentation has been successfully applied to extract segments representing geographic landmarks and architectural features from a variety of point clouds, whereas scene understanding has been successfully applied to indoor scenes on which other methods fail. The overall results demonstrate that the context-agnostic methods presented in this work can be successfully employed to manage the complexity of ever growing repositories