Transport Infrastructure

Fast detection, slow evaluation: Status data are still analyzed manually

Today, 3D data is used as standard when creating basic plans for the conversion and extension of infrastructure. Stationary terrestrial laser scanners (TLS) and ground-based or airborne mobile mapping systems equipped with laser scanners and cameras are used to collect the data. They enable measurements in moving traffic. Traffic jams or expensive closures can thus be reduced or completely avoided. Very large amounts of data (several gigabyte per road kilometer) are generated during the acquisition process, which until now have mainly been evaluated manually, i.e. by sighting. The various objects within the 3D and image data are manually outlined and assigned to a class. This type of evaluation is very time-consuming – unlike data acquisition itself, which is carried out very quickly thanks to modern mobile mapping systems, even when infrastructure elements with large dimensions are involved.

3D-AI: Fraunhofer IPM uses AI for efficient data evaluation

The Deep Learning Framework 3D-AI developed by Fraunhofer IPM provides a much more efficient and reliable interpretation of traffic infrastructure data than it was possible with previous methods. The framework projects the recorded scanner data reliably and accurately into the images of the color cameras. Each RGB image or multispectral image of the scene is assigned a corresponding depth channel. With the RGB-D(epth) data processed in this way and a trained artificial neural network, the data evaluation is proving to be very robust against object variations, different viewing angles or varying lighting conditions.

Within the scope of several projects, a library with annotated training data was created for the interpretation of infrastructure objects, which can serve as a basis for the training of an ANN. More than 30 objects from the categories street surfaces, street assets, vegetation and the like have already been trained.