The invention discloses a three-dimensional reconstruction method of a large-scene object. The method comprises the following steps: 1, acquiring a plurality of sequence pictures of the scene object from different visual angles and positions by use of a camera; 2, dividing the sequence pictures into a plurality of units Ni; 3, performing feature detection and matching on three sequence pictures of each unit by use of an SIFT algorithm; 4, initializing a first sequence picture of a unit N1 (i=1), and calculating a basic matrix F and an intrinsic matrix E of the other two sequence pictures; 5, solving camera parameters Ri and Ti and three-dimensional point cloud coordinates Mj of each sequence picture in other units; 6, obtaining seed surface sheets of the scene object; 7, expanding the seed surface sheets; and 8, filtering the seed surface sheets. According to the invention, by use of three-view constraints, mismatch points are effectively removed, an accurate basic matrix is obtained, the reconstruction precision is high, the units do not need coordinate conversion, the calculation complexity is reduced, and through performing dense reconstruction on the basis of a motion recovery structure, obtained point clouds of a three-dimensional model are enabled to be denser.