Feature extraction from 3D lidar point clouds using image processing methods

TL;DR: The objective of this research is to identify and assess automated, readily implementable GIS procedures to extract features like buildings, vegetated areas, parking lots and roads from LiDAR data using standard image processing tools, as such tools are relatively mature with many effective classification methods.

Abstract: Airborne LiDAR data have become cost-effective to produce at local and regional scales across the United States and internationally. These data are typically collected and processed into surface data products by contractors for state and local communities. Current algorithms for advanced processing of LiDAR point cloud data are normally implemented in specialized, expensive software that is not available for many users, and these users are therefore unable to experiment with the LiDAR point cloud data directly for extracting desired feature classes. The objective of this research is to identify and assess automated, readily implementable GIS procedures to extract features like buildings, vegetated areas, parking lots and roads from LiDAR data using standard image processing tools, as such tools are relatively mature with many effective classification methods. The final procedure adopted employs four distinct stages. First, interpolation is used to transfer the 3D points to a high-resolution raster. Raster grids of both height and intensity are generated. Second, multiple raster maps - a normalized surface model (nDSM), difference of returns, slope, and the LiDAR intensity map - are conflated to generate a multi-channel image. Third, a feature space of this image is created. Finally, supervised classification on the feature space is implemented. The approach is demonstrated in both a conceptual model and on a complex real-world case study, and its strengths and limitations are addressed.