Urban Point Cloud Mining Based on Density Clustering and MapReduce

Abstract

This paper proposes an approach to classify, localize, and extract automatically urban objects such as buildings and the ground surface from a digital surface model created from aerial laser scanning data. To achieve that, the approach involves three steps: (1) dividing the original data into smaller, more manageable pieces using a method based on MapReduce gridding for subspace partitioning, (2) applying the DBSCAN algorithm to identify interesting subspaces depending on point density, and (3) grouping of identified subspaces to form potential objects. Validation of the method was conducted in an architecturally dense and complex portion of Dublin, Ireland. The best results were achieved with a 1-m3-sized clustering cube, for which the number of classified clusters most closely equaled that which was derived manually (correctness=84.91%, completeness=84.39%, and quality=84.65%).