Deep-Learning segmentation and 3D reconstruction of road markings using multi-view aerial imagery

Kurzfassung

The 3D information of road infrastructures are gaining importance with the development of autonomous driving. In this context, the exact 2D position of the road markings as well as the height information play an important role in e.g. lane-accurate self-localization of autonomous vehicles. In this paper, the overall task is divided into an automatic segmentation followed by a refined 3D reconstruction. For the segmentation task, we apply a wavelet-enhanced fully convolutional network on multi-view high-resolution aerial imagery. Based on the resulting 2D segments in the original images, we propose a successive workflow for the 3D reconstruction of road markings based on a least-squares line-fitting in multi-view imagery. The 3D reconstruction exploits the line character of road markings with the aim to optimize the best 3D line location by minimizing the distance from its back projection to the detected 2D line in all the covering images. Results show an improved IoU of the automatic road marking segmentation by exploiting the multi-view character of the aerial images and a more accurate 3D reconstruction of the road surface compared to the Semi Global Matching (SGM) algorithm. Further, the approach avoids the matching problem in non-textured image parts and is not limited to lines of finite length. In this paper, the approach is presented and validated on several aerial image data sets covering different scenarios like motorways and urban regions.