3D Reconstruction of Steel Rail using Multi-view Cameras and Mask R-CNN

This study proposed a method that integrates multi-view image processing, depth estimation, and point cloud generation to accurately reconstruct a 3D model of a rail. The method is tested by positioning two webcams on the left and right sides of the rail and capturing three images of the entire rail. These images are combined into a panorama using image stitching techniques. Depth estimation is then performed using Monocular Depth Estimation, followed by background separation using Mask R-CNN to isolate the rail. The 3D model is reconstructed through 3D Reconstruction, and its quality is evaluated using an accuracy assessment. The resulting 3D model of the rail demonstrated an accuracy of 84.56% when compared to the ground truth, with the rail head 99.87%, the rail web 65.67%, and the rail foot 72.66%.