3D reconstruction of train accident scene based on monocular image.

Abstract

To help with making an emergency rescue plan for train accidents, a rapid 3D reconstruction method of train accident scene based on a monocular image was proposed. Taking two camera projection models for different application scenarios into consideration, the SIFT algorithm was introduced to extract and match image feature with the CAD model of an accident train. Geometric constraints between carriages were provided to transform the 3D reconstruction to solving a nonlinear least square problem with constraints, by which the position and pose of accident subjects were reduced at last. To quantitatively and qualitatively verify the calculation performance of this method, the mimicked train accident scene and real train accident scene were respectively used to carry out 3D reconstruction. The precise finite camera projection model was applied in the mimicked train accident scene to carry out offline calibration, and the stable pin-hole model was adopted in the real train accident scene to carry out auto calibration. Analysis result shows that through quantitative analysis of mimicked scene the maximal and average relative error of 8 nodes for measurement in reconstructing two carriages are 4.54% and 1.85% respectively. Through qualitative analysis of the real scene, the 3D reduction of position and pose for carriages can also be realized with combining the topographic information correction. The whole accident environmental panorama can be reduced visually with the help of 3D visualization engine. This method can also be used in developing emergency rescue electronic sand table for train accident analysis and safety education.