The invention discloses a geometric feature-based visual reconstruction method. According to the method, reconstruction is not performed for each pixel during 3D scene reconstruction; instead, firstly objects (geometric components) with geometric features in each scene are extracted, and the 3D scene reconstruction of a single scene picture is performed based on geometric blocks; secondly a plurality of scene pictures are matched by using the geometric features such as the sizes, the pixel information and the like of the geometric blocks as matching features; and finally the 3D scene reconstruction is finished. According to the scheme, the scene reconstruction is finished only by utilizing the geometric features instead of pixels, so that the calculation and storage overheads are greatly reduced and very high matching precision and scene reconstruction precision can be achieved.