Research And Implementation Of Structure From Motion Based On Unordered Images

Structure from Motion(SfM is a series of methods to reconstruct scene structure(i.e.three-dimensional space point)and calculate camera motion(i.e.camera pose)from image set.Its main methods are incremental reconstruction and global reconstruction.First of all,for these two reconstruction methods,the matching of feature points is the most time-consuming,especially when there is no prior matching information between images(i.e.,unordered set of images).Secondly,compared with the global reconstruction,the advantages of incremental reconstruction are higher accuracy,robust to outliers,and the disadvantages are time-consuming,dependent on the choice of initial image pairs and cannot close loop;compared with the incremental reconstruction,the advantages of global reconstruction are fast reconstruction and accurate closed-loop,the disadvantages are the lower reconstruction accuracy and the sensitivity to outliers.On the basis of the hybrid SfM proposed in 2017,this paper proposes the following improvement methods for the above two problems:1)The SIFT(Scale Invariant Feature Transformation)matching algorithm GeoMatch(Geometric structure and SIFT-based Matching algorithm),which is based on the geometric structure constraints of the scene and the SIFT,is proposed in the feature matching.The experiment shows that GeoMatch is superior to the tree and hash-based traditional matching algorithm in time and accuracy.2)In the rotation estimation stage,we propose a Bayesian inference network(BIN)based algorithm for edge weight in View Graph and a BIN based algorithm for edge weight update algorithm on Community Graph,and propose a stochastic maximum spanning tree(SMST)algorithm to filter out the outliers in rotation estimation.Experiments show that these improvements can be more robust to estimate the rotation.3)Based on the above two improvements,the prototype system SMSfM has been developed.The system has shown good reconstruction quality and time performance in many scenarios.