Hi, I am currently doing my Master's Thesis on semantic building modelling and trying to extract buildings by classification from a photogrammetric point cloud. However, the building classification gets mixed up with the vegetation classification. Could you give me advice on what the reason for that could be? A good classification is imperative for my work, since the resulting buildings are clustered and their respective point clouds are used to generate footprints and later the entire 3D models.
My best workflow after a few weeks of trying still does not deliver satisfactory results. It has the following structure (input cloud is in ECEF WGS84 coordinates):
- Transformation to UTM
las2las -i h100_l80_q80_g_o_dense_georef_clipped.las
-o h100_l80_q80_g_o_dense_georef_clipped_utm2.las -ecef -target_utm 32N - Ground estimation + height calculation
lasground_new64 -i h100_l80_q80_g_o_dense_georef_clipped_utm.las -city
-o h100_l80_q80_g_o_dense_ground.las -step 15 -ultra_fine -ecef -compute_height - Classification
lasclassify64 -i h100_l80_q80_g_o_dense_ground.las
-o h100_l80_q80_g_o_dense_classified.las -ground_offset 5 -keep_overhang -wide_gutters
This leads to the following result:
The result is relatively ok, but many trees, especially in the lower left part of the cloud, are wrongly classified. Furthermore, only the building rooftops are detected correctly (even when not using the -ground_offset 5 during classification).
During most of my tests, the buildings were classified partly as vegetation. Thereby, I have played with most of the parameters in lasground and lasclassify, such as step size, input reference system, offsets etc. Here is the result that I usually get (also from the upper cloud):
Both point clouds are photogrammetric, whereas the first dataset was gathered by me. Could the input data be the reason for the surprising result? The reconstruction was conducted in COLMAP, the georeferencing in Cloud Compare. I have also tried the same workflow with the wrapper transformers inside the Feature Manipulation Engine (FME) and get the same results.
I would appreciate any help or advice. Thank you in advance!
Best regards
Hristiyan Dimitrov