BMBF Project AuZuKa
Project duration: 01. February 2016 - 31. December 2021
This project is funded by the Federal Ministry of Education and Research under grant number 13N13891
Unwrapped pipe surface. Top: enhanced images created with 3d robot tracking; bottom: classical unwrap.
The inspection of sewer pipes is mandatory to guarantee their functionality. At present, mobile robots with cameras are used to tackle this task. The manual process of damage detection and classification is error prone because of the repeating and tiresome work. Therefor the goal of this project is the development of a system, which is capable of assisting the employee with the automatic detection and classification of damages.
First experiments in this project were based on legacy data from sewer pipe inspections. The pictures were taken with fisheye cameras. To account for artifacts caused by the camera movement we developed an algorithm to estimate the camera position. Furthermore we developed algorithms to eliminate uneven lighting in the images and to generate smooth transitions between subsequent images. Based on the noticeably improved images a Deep-NN was trained for the detection of defects.
Nevertheless, some damages are not visible in 2D images, so depth information is required. A novel camera head will be developed and used for the acquisition of the 3D pipe structure. Based on the generated point clouds a temporal and spatial consistent 3D pipe model can be calculated. Originating from the 3D model Deep Learning techniques will be used to detect and classify defects.
Publications
J. Künzel, T. Werner, R. Möller, P. Eisert, J. Waschnewski, R. Hilpert
Automatic Analysis of Sewer Pipes Based on Unrolled Monocular Fisheye Images, Proc. IEEE Winter Conf. on Applications of Computer Vision, Lake Tahoe, USA, Mar. 2018.
