Visual Odometry by Solving Relative Orientation Parameters of Consecutive Image Pairs

  • 林 照捷

Student thesis: Master's Thesis


In the past few decades Visual Odometry (VO) in computer vision has been widely developed and discussed in positioning and navigation technique Through analyzing consecutive image pairs the robotic can localize itself autonomously and continuously in an unknown environment without any human input The most famous example is robotic space mission on Mars in 2003 Actually photogrammetry can also achieve localization of the platform and it is known as recovering exterior orientation parameters (EOPs) of camera Because of the different methodology exploited in these two fields VO focuses on real time navigation and photogrammetry aims to the off-line mapping Consequently combinable methods containing VO in computer vision and photogrammetry are used to reconstruct the camera path in an unknown environment in this study This study can be divided into two parts: Relative Orientation Parameters (ROPs) of an image pair and Coherent Relative Orientation Parameters (CROPs) of consecutive images pairs Firstly solving ROPs of an image pair is a critical and fundamental work in VO technique so an appropriate matching strategy and an automatic ROPs method are applied in this study to solve reliable ROPs of an image pair Though automatic method is convenient the projective ambiguity containing scale and four-fold ambiguity should be solved in this study Secondly CROPs can be simply described as the overall EOPs of consecutive images Since the localization of consecutive images is incremental the errors are accumulated over time In this study a preliminary study of network adjustment of ROPs is proposed to suppress the accumulated error in the navigation For validation of feasibility of proposed monocular VO and network adjustment of ROPs two experiments including indoor and outdoor tests are conducted in this study In outdoor test 3 consecutive images pairs are captured in outdoor environment of Department of Geomatics In indoor test 9 consecutive images pairs are captured up the stairs in Department of Geomatics In analysis of monocular VO the results show that the estimated camera path and reconstructed 3D object points are reasonable and the features in reconstruction can be recognized clearly In analysis of network adjustment comparing estimated camera paths with actual camera path the accumulated errors reduce significantly in outdoor test when employing network adjustment Furthermore the deviation of object points derived from different image pairs are decreased in both two tests Consequently the network adjustment in this study is effective to reduce the accumulated errors of the camera path This study shows the feasibility of monocular VO with combinable methods containing photogrammetry and VO in computer vision The proposed network adjustments of ROPs for accumulated errors in navigation is validated In addition real scale of translation in monocular VO can also be solved in network adjustments of ROPs in this study
Date of Award2018 Aug 7
Original languageEnglish
SupervisorYi-Hsing Tseng (Supervisor)

Cite this