An Novel Terrain Detection Method For Automated Processing of Laser Scanning Point Clouds

Terrain detection is critical for the automated processing of laser scanning point clouds used in forest studies. More significantly, to produce forest structure (i.e., tree height), it is indispensable to normalize the point cloud values by locating the terrain level and computing the height above the ground. Existing terrain detection methods use a priori information (i.e., laser scanners configurations), intensive user actions (i.e., parameters tuning), or are designed for specific datasets and environments. Moreover, the increasing size of point clouds creates additional challenges regarding computational efficiency.

Our goal is to design an automated, efficient, and robust algorithm for terrain detection. To do so, we divide the entire input points into small parallel sub-regions, namely slices, in the vertical direction (z-axis). Then, we find the lowest points in each slice, followed by interpolating ground curves from these lowest points via b-spline. Later, in each slice, the height over the detected curve is calculated for each point. Finally, all points are updated with the height values. Our method can automatically normalize points and filter non-ground points without external information. Moreover, the approach can be easily parallelized to speed up on large point clouds.

Current Terrestrial Laser Scanning (TLS) point cloud experiments present promising visualization results. We will test more datasets with reference ground datasets like EuroSDR, and include more types of point clouds, like Airborne Laser Scanning (ALS) point clouds. Quantitative and qualitative comparisons with state-of-art methods (e.g., lasground in LAStools) are expected in the experiments.