Light Detection and Ranging (LiDAR) is a popular tool for range sensing applications, with applications including environmental modeling and urban planning. Modern acquisition platforms facilitate data collection rates of over two billion points per hour, enabling the collection of massive datasets with ease. These datasets must typically undergo a large amount of processing before use, making maintenance a difficult issue. Additionally, the presence of transient objects can affect dataset quality, both as unwanted data and by obscuring the underlying surface model.
In this dissertation, we present a framework for enabling automatic change detection between large LIDAR datasets of urban environments. For two large scale datasets, we extract the relevant portions using information about the paths of the acquisition vehicles. However, acquisition inaccuracies can result in subset misalignment of up to 10 meters. To correct for this, we utilize a variety of point cloud alignment techniques, including a novel point descriptor, to bring overlapping pieces of data into alignment. Given the properly aligned data, we then use novel hierarchical and point-based techniques to extract regions of change between the two datasets. These regions can then be extracted and presented for further processing or filtering.
We present the results of our research executed on datasets totaling over 93 billion sample points. At over 1.5 terabytes in size, this represents by far the largest collection of ground-based LIDAR examined in the open literature. We quantify our results with a variety of objective metrics, investigate modes of failure, and recommend directions for future research.