Individual tree extraction from urban LMMS data
Organizations involved: IGN (France)
The workflows address large point clouds from Laser Mobile Mapping Systems (LMMS). LMMS is a fairly new and still developing data acquisition technology capable of generating enormous point clouds in a short time. This showcase addresses how to efficiently and reliably identify and extract trees from big LMMS data. Methods for extracting objects from LMMS data are still immature and addressed by many research groups. It is considered challenging to successfully identify individual trees from part of a point cloud classified as sampling trees. Pre IQmulus algorithms for extraction of metric and quantitative information from point clouds were typically demonstrated on small point clouds (< 100.000.000 points) sampling a few adjacent facades at most.
The IQmulus algorithms have been tested on a 10km-long mobile mapping dataset, as acquired by IGN’s Stereopolis vehicle within two hours of operation in downtown Toulouse resulting in a data size of 121 Gbyte (1 012 160 278 points). This data set was split in 517 tiles of up to 3M data points. Running the workflow for the full Toulouse data set resulted in the automated extraction of more than 4000 trees, see Figure 16. The processing on the Fraunhofer cloud using 12 nodes took around five hours (7 minute per tile+ overhead for scheduling), compared to 52 hours on a desktop computer (6 minutes per tile). The accuracy of the three detection was around 80%.
On this figure the yellow pins (4000+) are the automatically identified trees by the workflow.