Automatic Registration for Microsatellite Push-Frame Images

Abstract

In recent years commercial Earth observation satellites have become smaller in size and weight The advantage of miniaturizing satellites is that the cost of launching can be significantly reduced allowing people to operate large numbers of satellites in space at relatively low cost and thereby increasing the time resolution of satellite observations One of the successful examples is the SkySat-1 microsatellite In addition to the above advantages the SkySat-1 differs from conventional satellites in that it uses a push-frame sensor Therefore in the same flight strip images continuously captured in a short time can have a high overlap Satellite imagery is usually applied after radiometric and geometric corrections In terms of geometric correction the conversion between object space and image space is traditionally established by using rigorous sensor models or Rational Function Models (RFMs) On the other hand the method of establishing the relationship between object space and image space has also developed in the field of computer vision A well-known method is Structure from Motion (SfM) Although SfM is generally used for close-range photogrammetry and the purpose is more focused on reconstructing the model or coordinates of the 3D scene based on the imaging mode of the microsatellite push-frame sensor this study aims at exploring the feasibility of automatic registration of microsatellite push-frame images based on the SfM-based approach The research approach can be divided into two stages which are the ROP estimation of a single image pair and the global EOP estimation of a set of images In a flight strip the ROPs of any two sufficiently overlapped images can be obtained through the correspondences of image features Obtaining reliable and stable ROPs must rely on accurate adequately numbered and evenly distributed image features Therefore the first part of this study uses image feature matching strategy and with appropriate error matching filtering methods to get accurate feature correspondences The feature correspondences can be used to calculate the essential matrix and further decompose it into ROPs The EOPs of all involved images can then be reconstructed through these ROPs An incremental approach and a one-step adjustment to estimate the EOPs by ROP network adjustment are proposed The incremental approach is divided into two steps In the first step the rotation parameters are calculated In the second step the rotation parameters are regarded as known values and the unknown position parameters and the true scales are solved By contrast the one-step adjustment solves the rotation parameters position parameters and true scales simultaneously In the least-squares adjustment model in addition to the ROP observations the baseline scales of the images estimated by using the microsatellite on-board data are also used as observations This is necessary to resolve the degeneracy in translation estimation In order to prove the feasibility of automatically registering microsatellite push-frame images by SfM-based methods this study simulates the microsatellite push-frame images by high altitude aerial photography and takes 10 consecutive images for ROP and EOP estimation The results are compared with the EOPs solved by aerial triangulation (AT) Experimental results show that both the incremental approach and the one-step adjustment can provide reasonable EOPs of the images while the accuracy of the incremental approach is slightly better than the one-step solution in the calculation of translation parameters

Date of Award	2019
Original language	English
Supervisor	Yi-Hsing Tseng (Supervisor)