Artificial intelligence for point cloud segmentation and 3D reconstruction
Point clouds of cities depict urban life down to its smallest detail. The point clouds contain a lot of important information such as vehicles, buildings or infrastructure. But how can meaningful information be extracted from these terabytes of data? With the help of artificial intelligence VRVis is able to read out sidewalks and platform edges from point clouds with real-world data.
The data pipeline shows how sidewalks and sidewalk edges are detected from large point clouds using segmentation and 3D reconstruction.
At VRVis, a multi-member research group specializes in geospatial data. Researcher Lisa Kellner is one of the experts who develops new solutions for the visualization and analysis of and interaction with 3D geospatial data.
Laser scan data captures our physical world in large point clouds with data volumes often in the terabyte range. Processing this data is challenging in terms of both volume and extraction of usable information.
For the efficient processing of these gigantic point clouds, we use optimized spatial data structures in combination with special neural network architectures to segment objects and structures in the point clouds and use them for 3D reconstructions. This is a task that is especially important for point clouds of urban areas, but also very challenging. Such point clouds contain very heterogeneous and useful information, such as buildings, vegetation and infrastructure, but also non-stationary objects like pedestrians and vehicles. A successful example for the use of artificial intelligence is the reconstruction of sidewalks and platform edges from real world data using 3D CNNs (Convolutional Neural Networks), which we have already implemented in various projects with corporate partners.