Key Technology Of Position And Orientation Estimation For UAV Images Based On Global Structure From Motion

Confronting the requirements of applications such as reconstruction of battlefield environment and detection of disaster situation,flexible micro UAV(Unmanned Aerial Vehicle)remote sensing system could respond quickly and acquire images data of objective area right the first time.Structure from motion technology in computer vision provides a solution approach for automated processing of UAV data.However,there is still room to improve in timeliness and robustness.Aiming at acquiring 3D(three-dimentional)geometry information of objective area from UAV images effeciently and robustly,multiple view geometry theory in computer vision and error processing and reliability theory in photogrammetry were intergrated,and robust position and orientation estimation for UAV images were researched and explored in this dissertation.The main contributions are as follows:1.The development status of UAV remote sensing systems and data processing technologies has been summarized systematically and comprehensively.Key technologies in multiple view geometry as imaging model building,construction of epipolar geometry relationship between two images,model coordinates estimation,bundle adjustment were stated,and the transforming equation between fundamental matrix and relative orientation elements were deduced.2.In response to the disadvantage of traditional air strips organization pattern for automated and efficient processing of UAV images,an organization method for UAV images based on relationship graph was proposed.The theory of graph was introduced,the relationship undirected graph was designed using both low-precision position and orientation data and image matching method,and by means of estimating overlap condition of images to limit the range for matching,the matching time was shortened substantially as blind ergodic matching avoided.Furthermore,three views matching was used to optimize the graph,and depth-first search was applied to obtain a steady relationship graph for strongly-related images.Then according to parallel rigid theory,the judgement condition of uniquely determining for camera position was given,which could provide theoretical guidance for position estimation.On this basis,3D points cloud was generated incorporating with global SfM(Structure from Motion)with three views constraint.The reliability of reconstructed result was enhanced by introducing three views constraint.3.In response to the sensitivity of rotation averaging based on least squares to outliers,a robust Lie algebra rotation averaging method called IRLARA(Iteratively Robust Lie-Algebraic Rotation Averaging)was proposed.The Lie group and Lie algebra theory was utilized to simplify the product operation of rotation matrixes into linear operation of vectors,thus the computing efficiency was improved.The reliability and accuracy were guaranteed by combining L1 norm optimization and iteratively reweighted least saquare.And the impact of outliers was limited effectively through iteratively weeding out outliers.As a result,the accuracy of rotation parameters was improved obiviously.Furthermore,the global consistency rotation parameters were transformed to WGS84 coordinate system,and a GPS-supported points cloud generating scheme was designed.Adopting GPS data simplified the complexity of SfM greatly,and could increase the stability of scene reconstruction.4.Combining the character of UAV images,the shortcomings of global SfM adopting three views constraint in UAV images processing were analysed,and a global consistency position estimation method using GIRLS(Generalized Iteratively Reweighted Least Squares)was proposed.Based on matched points and epipolar constraint,the global relative translation direction was obtained by L1 norm estimation,and the reliability of observations for global position estimation was enhanced.The accuracy of global position calculating was guaranteed by means of solving quadratic programming within the framework of GIRLS.Integrating images relationship graph building,robust Lie algebra rotation averaging and position estimation method with GIRLS,a global SfM method called IG(IRLARA-GIRLS)was constructed.The treatment efficiency was improved due to avoiding massive computing related to three views,the completeness of reconstructed scene was insured as the connected images relationships and corresponding matched points sets in relationship graph were made the utmost of,and the accuracy and reliability of position and orientation parameters solved was perfected by adopting robust estimation.Experiment results showed that,the relative errors by measuring distances in the generated points cloud model were not larger than 1‰.