Algorithm Research On Motion Estimation Of Unstructured Images Based On Consumer Camera

Structure from motion,a.k.a.SFM,is a critical problem in computer vision and photogrammetry field.It is a process of estimating instantaneous camera pose in the capture and scene structure information simultaneously merely by a set of overlapping images or videos of the same object,taken from different viewpoints,which accuracy and efficiency will directly have an influence on following 3d reconstruction.Although the whole procedure is basically established regarding to this problem around 2000 and multi-view geometry reconstruction has been studied for many years,a considerable number of researchers still have focused on such difficulty up to now since the process is relatively complicate,need to set many parameters,also is hard to find the globally optimal solution and exists bottleneck in reconstruction speed for large-scale images.So far,there really has been no a general and mature reconstruction software based on multi-sources images while several brilliant foreign commercial software is still in dominant position.Consequently it is essential to conduct sparse reconstruction algorithm research for arbitrary images on this basis.Aerial triangulation in the traditional photogrammetry which is equivalent to structure from motion process in computer vision needs strict camera calibration parameters and strips information.The general process is that doing relative orientation for a single model first and then connecting those single models.Besides,orientation is always along the direction of X in the photogrammetric coordinate.However,in computer vision field,the process of taking photos is casual.We usually don't require any priori information,and we can acquire approximate camera parameters directly from image EXIF tags in the most cases then do a forward optimization for camera parameters in the bundle adjustment process.This method can reduce the limitation to image devices.In the context of having difficulty obtaining camera parameters for most images such as smart phone photo,digital camera photo,internet photo,even UAV image and aerial image,it is advisable to adopt the solution come from computer vision to solve this problem.Based on this point,the main contents of this paper include:1.A systematic summary of theories foundation and algorithm principle of image reconstruction.Introducing three typical reconstruction algorithms in computer vision field and evaluating respective advantages and disadvantages and its application condition.Concluding the general solution and flow for image reconstruction and having a comprehensive summary about functions and features for several state-of-theart commercial software and open-source software.Based on these work,we generalize two crucial problems in the image reconstruction: one is how to identify the data association facing massive image data,the other is that which reconstruction strategy we should adopt since different strategies could generate different results.2.A detailed analysis of two and three view reconstruction algorithm.This process corresponds to relative orientation and absolute orientation in photogrammetry.We delve into eight-point algorithm and five-point algorithm respectively and its characteristics.We adopt different algorithms in calibrated and non-calibrated camera cases and combine it with RANSAC algorithm to acquire robust solution.Besides,we introduce a novel parameterization of the perspective three point problem for a direct computation of absolute camera position and orientation.The result shows that the algorithm is efficient in the course of image reconstruction.We study a multi-model geometric verification method in the stage of match constraint to choose the best suitable model,and thus lay a solid foundation for following reconstruction.3.A comparison of three types of feature extraction algorithm's(SIFT,SURF,ORB)influence on image reconstruction.The traditional solution in image reconstruction still mostly employ SIFT algorithm because of its stability and robustness,but the efficiency of SIFT is not high and perfect in the aspect of massive data and real-time.This paper apply the other two algorithms SURF and ORB to image reconstruction and compare the performance of these three algorithms and its influence on reconstruction.In the end,we offer a strategy to improve reconstruction speed in the condition of keep stability.4.This paper implement and improve traditional incremental image reconstruction algorithm.We adopt SIFT and SURF algorithm respectively to extract features and match it.When obtaining exhaustive image matching information which is used to generate track information as a whole input.Then we find initial image pair by offering user manually and automatic searching.After identifying initial pair,we try to discover the most number of 2D-3D corresponding point between remaining images and previous images to identify the next reconstruction image.Next,executing one bundle adjustment just only to motion is necessary.We determine if there should do global bundle adjustment depending on specific parameters we set.It turns out that this solution is practicable.Finally,it reaches the desired effect.Besides,we proposes a method of picking the best next view based on the distribution of feature points in the target image.Compare to the traditional method of choosing the most 2D-3D correspondences points,this strategy can improve accuracy and robust of reconstruction to some extent.