Semantic Metric 3D Reconstruction for Concrete Inspection

Abstract: In this paper, we exploit the concrete surface flaw inspection through the fusion of visual positioning and semantic segmentation approach. The fused inspection result is represented by a 3D metric map with a spatial area, width, and depth information, which shows the advantage over general inspection in image space without metric info. We also relieve the human labor with an automatic labeling approach. The system is composed of three hybrid parts: visual positioning to enable pose association, crack/spalling inspection using a deep neural network (pixel level), and a 3D random field filter for fusion to achieve a global 3D metric map. To improve the infrastructure inspection, we released a new data set for concrete crack and spalling segmentation which is built on CSSC dataset [27]. To leverage the effectiveness of the large-scale SLAM aided semantic inspection, we performed three field tests and one baseline test. Experimental results show that our proposed approach significantly improves the capability of 3D metric concrete inspection via deploying visual SLAM. Furthermore, we achieve an 82.4% MaxF1 score for crack detection and 88.64% MaxF1 score for spalling detection on the relabeled dataset.