Accelerating point cloud cleaning
Thesis
Capturing the geometry of a large heritage site via laser scanning can produce thousands of high resolution range scans. These must be cleaned to remove unwanted artefacts. We identified three areas that can be improved upon in order to accelerate the cleaning process. Firstly the speed at which the a user can navigate to an area of interest has a direct impact on task duration. Secondly, design constraints in generalised point cloud editing software result in inefficient abstraction of layers that may extend a task duration due to memory pressure. Finally, existing semi-automated segmentation tools have difficulty targeting the diverse set of segmentation targets in heritage scans. We present a point cloud cleaning framework that attempts to improve each of these areas. First, we present a novel layering technique aimed at segmentation, rather than generic point cloud editing. This technique represents 'layers' of related points in a way that greatly reduces memory consumption and provides efficient set operations between layers. These set operations (union, difference, intersection) allow the creation of new layers which aid in the segmentation task. Next, we introduce roll-corrected 3D camera navigation that allows a user to look around freely while reducing disorientation. A user study shows that this camera mode significantly reduces a user's navigation time (29.8% to 57.8%) between locations in a large point cloud thus reducing the overhead between point selection operations. Finally, we show how Random Forests can be trained interactively, per scan, to assist users in a point cloud cleaning task. We use a set of features selected for their discriminative power on a set of challenging heritage scans. Interactivity is achieved by down-sampling training data on the fly. A simple map data structure allows us to propagate labels in the down-sampled data back to the input point set. We show that training and classification on down-sampled point clouds can be performed in under 10 seconds with little effect on accuracy. A user study shows that a user's total segmentation time decreases between 8.9% and 20.4% when our Random Forest classifier is used. Although this initial study did not indicate a significant difference in overall task performance when compared to manual segmentation, performance improvement is likely with multi-resolution features or the use of colour range images, which are now commonplace.