从多视角卫星图像生成准确DSM的方法研究

Research on the Method of Generating Accurate DSM from Multi perspective Satellite Images

神经辐射场(NeRF)已成为一种新兴的三维重建方法,然而由于缺乏几何约束,其在多视角卫星摄影测量中的应用很难获得准确的数字表面模型(DSM)。为解决这个问题,文中提出了通过几何约束神经辐射场(GC-NeRF)从多视角卫星图像生成准确的DSM,其关键是针对卫星相机视点分布紧密的特征设计一个几何损失项,可以通过促使表面更薄来约束场景的几何结构,极大地提高了生成DSM的准确性。在WorldView-3卫星图像上的实验表明,GC-NeRF可以利用多视角卫星图像生成更精确的DSM。

Neural radiation field has become an emerging 3D reconstruction method,but due to the lack of geometric constraints,its application in multi view satellite photogrammetry is difficult to obtain accurate digital surface models(DSM).To address this issue,the paper proposes generating accurate DSMs from multi view satellite images using Geometric Constrained Neural Radiation Fields(GC NeRF).The key is to design a geometric loss term for the tightly distributed viewpoint features of satellite cameras,which can constrain the geometric structure of the scene by making the surface thinner,greatly improving the accuracy of generating DSMs.Experiments on WorldView-3 satellite images have shown that more accurate DSMs can be generated using multi view satellite images.