Abstract for HONS 02/22
Semantic Segmentation of Trees in Aerial Point Clouds of New Zealand Urban Areas
Grant Wong
Department of Computer Science and Software Engineering
University of Canterbury
Abstract
Urban trees are critical in mitigating many negative impacts of urban areas, and understanding the extent to which they do this requires accurately segmenting and identifying the precise area of tree regions in a city. For urban areas represented by aerial point cloud data, existing methods rely on datasets with eight or nine classification classes, with the segmentation performance on highly imbalanced datasets with only two classes yet unexplored. We develop a pre-processing and data loading pipeline and use the SoftGroup sparse convolutional neural network (CNN) library to segment a custom urban point cloud dataset, and we explore the effects of different training parameter values on the segmentation accuracy. From our experiments, we are able to achieve a maximum average mIOU percentage of 72.8% on our test dataset.
