Point Cloud and Digital Surface Model Generation From High Resolution Multiple View Stereo Satellite Imagery

Abstract

ABSTRACT:Nowadays, multiple-view stereo satellite imagery has become a valuable data source for digital surface model generation and 3D reconstruction. In 2016, a well-organized multiple view stereo publicly benchmark for commercial satellite imagery has been released by the John Hopkins University Applied Physics Laboratory, USA. This benchmark motivates us to explore the method that can generate accurate digital surface models from a large number of high resolution satellite images. In this paper, we propose a pipeline for processing the benchmark data to digital surface models. As a pre-procedure, we filter all the possible image pairs according to the incidence angle and capture date. With the selected image pairs, the relative bias-compensated model is applied for relative orientation. After the epipolar image pairs' generation, dense image matching and triangulation, the 3D point clouds and DSMs are acquired. The DSMs are aligned to a quasi-ground plane by the relative bias-compensated model. We apply the median filter to generate the fused point cloud and DSM. By comparing with the reference LiDAR DSM, the accuracy, the completeness and the robustness are evaluated. The results show, that the point cloud reconstructs the surface with small structures and the fused DSM generated by our pipeline is accurate and robust.

To Bundle Adjust or Not: A Comparison of Relative Geolocation Correction Strategies for Satellite Multi-View Stereo

“…The Computer Vision community has proposed different methods to correct the pointing error of RPC camera models. Bundle adjustment based solutions are a generally accepted practice that consist in detecting inter-image tiepoints and applying a compensating function to the original RPCs so that the back-projections of the tie-points are coincident in the 3D world [14,5,22,13,18].…”

“…The maximum correlation translations are employed to register all DSMs to the frame of reference of the first input stereo pair, which is expected to be the best according to the selection criterion used (see Section 3.1). After the alignment, the point-wise median is used to perform the DSMs fusion, as in [13]. Remark that the fusion is done using the DSMs previous to interpolation.…”

“…Previous work already highlighted the importance of the criterion used to select input stereo pairs for MVS pipelines dealing with satellite imagery [11,13]. Poor choices lead to pairs of views sharing less visual content and output DSMs with larger errors and incomplete areas, making the correction of the pointing error harder both for image and geometry based methods.…”

Dense Multi-View Image Matching for DSM Generation From Satellite Images

“…Points visible in only an image pair are chosen in one group named non-reliable point cloud (single), and other points are named reliable point clouds. For DSM generation, (Kuschk, 2013, Qin, 2017, Gong et al, 2018 generate DSM for each stereo pair. In the end, all these DSMs are fused by simple median filter (Kuschk, 2013, Gong et al, 2018 or adaptive depth fusion method that considers the spatial consistency (Qin, 2017).…”

“…Recently, researchers have used the multi-image techniques to generate a digital model from close range images (Ahmadabadian et al, 2013, Innmann et al, 2019, aerial (Ameri et al, 2002, Haala et al, 2012 and satellite images (Zhang et al, 2006, Krishna et al, 2008, d'Angelo et al, 2012, Giribabu et al, 2013, Qin, 2017, Gong et al, 2018. The accuracy of the final point clouds directly related to the accuracy of image matching algorithms.…”