DENSE RECONSTRUCTION FOR TUNNELS BASED ON THE INTEGRATION OF DOUBLE-LINE PARALLEL PHOTOGRAPHY AND DEEP LEARNING

Abstract

In scenes of tunnels and underground engineering projects, the 3D modeling based on the traditional horizontal baseline photography method is difficult to make a trade-off between the modeling efficiency and image overlap under narrow space and close photography constraint conditions. Parallel photography provides a better alternative. Furthermore, the undulating tunnel surfaces make occlusion unavoidable, and the pixel scales vary evidently in parallel photography, both of which make it difficult to obtain expected models neither by Semi-Global Matching nor by Patch-based Multi-view Stereo techniques. Comparatively, more accurate 3D reconstruction results can be achieved by using learning techniques, which consider the global semantic information such as specular prior and reflective prior to making the matching more robust. Besides, it is convenient for geologists to acquire photos with smartphones or cameras by a single-line parallel photography method at the tunnel sites. But enough image overlap for reconstruction is still difficult for the above photography way. Therefore, this paper proposes a dense reconstruction method for tunnels using deep learning and double-line parallel photography techniques, and respectively makes comparisons with single-line parallel photography and traditional modeling methods. Experimental results show the feasibility and robustness of the proposed method, especially for tunnel surfaces with complicated textures and occlusions. Moreover, both the completeness and accuracy of the tunnel model with double-line photography are better than with single-line photography.