Fusion of Stereo Aerial Images and Official Surveying Data for Mapping Curbstones Using AI

Semantic segmentation and object extraction from aerial images have made tremendous progress along with the evolution of deep learning neural network architectures. However, collecting high quality training data is still the bottleneck for many applications, in terms of costs and limited visibility of small objects. Conducting the training of artificial intelligence (AI) with accessible and adapted official data reduces the effort and allows to integrate independent in-situ knowledge as reference. Focusing on a prominent road element as example, this work presents a new approach for detecting curbstones from airborne stereo images with the assistance of official surveying data. To adapt the curbstone maps to the oblique view images, reference information is removed in occluded regions. The refined reference masks are fused with airborne imagery and integrated to the training of a Swin transformer segmentation model. In the end, the curbstone segments are transformed into vectors using an advanced vectorization approach. The proposed approach is tested over the city area of Brunswick, Germany.