Multi-temporal point cloud data: Detecting 3D change from amongst the noise

Advancements permitting the rapid extraction of 3D point clouds from stereo imagery covering large portions of the landscape have provided a vast collection of high-fidelity digital surface models (see figure to the right) of the planetary surface. These models should be acquired rapidly in time, presenting an opportunity for surface model change detection on a quasi-global scale (see figure to the lower right). Although two-dimensional change detection from remotely sensed imagery is fairly common, leveraging 3D data derived from stereo pairs present new challenges. Some of these challenges include complex propagation of error and uncertainty associated with the tasks of photogrammetric processing, ideal stereo pair matching, and designing algorithms to be semantically effective with this modality of data. Additionally, topographic data are often represented using traditional computer vision techniques which either compromise the data fidelity or the ability of a change detection model to perform. This work presents and explores solutions to these challenges, while exploring the state of the art possible with change detection. An end to end analysis of the 3D change detection task is made as the author presents novel approaches to data preprocessing, error localization and change detection. Together, these efforts will establish the computational tools necessary to rapidly evaluate 3D change from overhead imagery, with applications to various domains including humanitarian assistance, disaster response, and environmental monitoring.