A Photogrammetric Computer Vision Approach for 3D Reconstruction and Volume-Change Measurement of Unsaturated Soils

Existing methods for measuring volume changes of unsaturated soil specimens during triaxial tests have several limitations. Recently a photogrammetry-based method has been proposed to overcome these limitations. Although this method has many advantages over existing methods, such as low-cost, high-accuracy, no requirements for camera positions, and so on, it relies on a commercial software PhotoModeler to detect coded targets which involves tedious manual correction of coded target IDs and thus is not efficient to use. The objective of this study is to make the abovementioned photogrammetry-based method simpler, faster, and more automated to use. To this end, deep learning aided detection method has been proposed for the automatic detection of coded targets. Image processing technique has been proposed to automatically correct coded target IDs without manual operation. Based on the coded target detection results, a photogrammetric computer vision 3D reconstruction approach also has been proposed to reconstruct the 3D models of the cylindrical soils sample. Validation tests have been performed to validate the proposed approach. It is shown that it has accuracy comparable to commercially available software, and the average difference between results obtained from the proposed method and commercially available software is 0.01 mm.