基于单目图像的列车事故场景三维重建

3D reconstruction of train accident scene based on monocular image

为了辅助铁路事故应急救援方案的制定,提出了一种基于单目图像的列车事故场景快速三维重建方法。考虑不同应用场景的2种相机投影模型,采用SIFT算法提取图像特征并与事故列车的CAD模型相匹配,通过引入车厢之间的几何约束,将三维重建转换为求解带约束的非线性最小二乘问题,最终还原事故主体的位置与姿态。为了定量与定性验证该方法的计算效果,分别对模拟列车事故场景与真实列车事故场景进行车厢的三维重建。在模拟列车事故场景中采用了较精确的有限相机投影模型进行离线标定,在真实列车事故场景中采用了较稳定的针孔模型进行自标定。分析结果表明:通过对模拟场景的定量分析,重建两节车厢中用于测量的8个节点的最大相对误差为4.54%,平均相对误差为1.85%;通过对真实场景的定性分析,结合地形信息校正,同样能够实现车厢位置与姿态的三维还原;最终借助三维可视化引擎,可在视觉上还原整个事故环境全貌。该方法还可用于应急救援电子沙盘的开发以进行铁路事故分析和安全教育。

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 vehicles 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 the method,the mimicked train accident scene and the 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.Analysisresult shows that through quantitative analysis of mimicked scene,the maximal and average relative error of 8nodes for measurement in reconstructing two vehicles are 4.54% and 1.85%respectively.Through qualitative analysis of real scene,the 3D reduction of position and pose for vehicles can also be realized by combining the topographic information correction.The whole accident environmental panorama reduces visually with the help of 3D visualization engine.