Description
Concrete structural health monitoring (SHM) is urgently required with more and more severe aging problems happen. Conventional manual inspection approach is labor-intensive and time-consuming. In this paper, we propose a wall-climbing robotic approach for visual inspection of concrete structures using a deep neural network and 3D semantic reconstruction to build a 3D map overlaid with defects. The wall-climbing robot uses a negative pressure module to operate on both vertical and horizontal surfaces. An RGB-D camera is mounted in the robot with an Intel-NUC computer as visual positioning and image processing core to make metric measurement, Our main contribution is the inspection neural network that performs pixel-level segmentation to detect cracks and spallings and mark them on a 3D map for better visualization. We build a semantic dataset which includes 820 labeled images and training on the dataset with 12,000 iterations. We introduce a 3D semantic fusion method to build the 3D map with defects highlighted. The field test and experimental results demonstrate that our wall-climbing robot and inspection system can perform a robust 3D metric inspection.
Location
St. Louis, Missouri
Presentation Date
06 Aug 2019, 2:00 pm - 2:20 pm
Meeting Name
INSPIRE-UTC 2019 Annual Meeting
Document Type
Article - Conference proceedings
Document Version
Final Version
File Type
text
Language(s)
English
Source Publication Title
Proceedings of the 9th International Conference on Structural Health Monitoring of Intelligent Infrastructure (2019: Aug. 4-7, St. Louis, MO)
Included in
Visual SHM for Concrete Infrastructure using a Wall-Climbing Robot
St. Louis, Missouri
Concrete structural health monitoring (SHM) is urgently required with more and more severe aging problems happen. Conventional manual inspection approach is labor-intensive and time-consuming. In this paper, we propose a wall-climbing robotic approach for visual inspection of concrete structures using a deep neural network and 3D semantic reconstruction to build a 3D map overlaid with defects. The wall-climbing robot uses a negative pressure module to operate on both vertical and horizontal surfaces. An RGB-D camera is mounted in the robot with an Intel-NUC computer as visual positioning and image processing core to make metric measurement, Our main contribution is the inspection neural network that performs pixel-level segmentation to detect cracks and spallings and mark them on a 3D map for better visualization. We build a semantic dataset which includes 820 labeled images and training on the dataset with 12,000 iterations. We introduce a 3D semantic fusion method to build the 3D map with defects highlighted. The field test and experimental results demonstrate that our wall-climbing robot and inspection system can perform a robust 3D metric inspection.
