Semantische Segmentierung optischer Sensordaten für Anwendungen in der Binnenschifffahrt

Abstract

Inland waterway transport (IWT) is an extremely important backbone for heavy good transportation with severe economical influence and the potential for the reduction oftraffic-related greenhouse gas emission. As IWT is expected to increase, updated chart data is required. Traditional survey methods are intense in cost and time. This work presents a processing scope for self-updating inland waterway charts. The required data can be gathered through optical sensors, that are fitted on IWT vessels. In semantic segmentation, every pixel in a RGB image is assigned to a defined class. This machine-learning problem is used to distinguish between various objects in a (IWT related) scene and thus to survey the infrastructure. For this task, the new BerlinIWT dataset is proposed. Existing datasets in this field may contain more examples, but do not provide an adequate number of classes. Training a neural network on the datasets MaSTr1325 and BerlinIWT leads to remarkable results. Spatial mapping information is completed with LiDAR (light detection and ranging) data. The acquired 3D point clouds provide precise distance information with a reasonable maximum range. The sensor compensates the flaws of (stereo) cameras, that are suitable for scene understanding, but inappropriate for distance measurements. The most suitable technique for the combination of LiDAR and camera data is discussed. For the ongoing scope towards simultaneous localisation and mapping (SLAM), two different methods for optical flow estimation are compared. Finally, further processing steps are pointed out and the application is discussed with respect to a traffic-telematics related use-case.