Position And Orientation Estimation Of Large Scale Image Based On Distributed SFM

Structure From Motion(SFM)is an algorithm to estimate image orientation and3 D coordinates of sparse point cloud from 2D images.When the iamge scale is large,the parameters are very large and there are many redundant observations,so SFM is constrained by efficiency,memory,accuracy.For position and orientation estimation of large scale image,this paper studies the distributed SFM algorithm to decompose largescale problem into many small-scale problems.For the redundancy of observations,we propose track selection algorithm.Firstly,the distributed SFM algorithm clusters images,divide image into several clusters.After block partition,each cluster is oriented based on Increamental SFM,which can be calculated in parallel.Finally,all clusters are combined into a global scene by using the consistency constraints of common parameters between clusters.Bundle Adjustment is the core of SFM and it’s also a very time-consuming part,in which the redundant observations and initial values with large errors add the algorithm’s time consumption.In order to improve the efficiency of Bundle Adjustment,we propose track selection method,the preserved tracks need to be distributed evenly and accurately to ensure the accuracy and integrity.We divide the space of 3D points into 3D grids,eliminate redundant tracks with low accuracy in each grid,then judge whether the reserved tracks cover all the cameras and “camera-camera”edge,if it doesn’t meet the requirements,we select rejected tracks to add until the above requirements are met.The efficiency of Bundle Adjustment can be improved by track selection.Finally,the paper introduces parallel Bundle Adjustment.Divide the large scene after initial orientation into several sub scenes,The internal parameters of the sub scene are optimized separately,and then makeing the parameters of all sub scenes meet the global consistency based on the consistency constraint of 3D common points.