Xu Huang Department of Civil, Environmental and Geodetic Engineering, The Ohio State University
Rongjun Qin Department of Civil, Environmental and Geodetic Engineering, The Ohio State University
Xiaohu Lu Department of Civil, Environmental and Geodetic Engineering, The Ohio State University

12E

Bundle adjustment for high-resolution satellite images is often operated on Rational Polynomial Coefficients (RPC) for registering them to a unified geodetic framework, which is a critical step in many photogrammetry and computer vision application. However, the growing number of high resolution spaceborne optical sensors has brought a few challenges to the bundle adjustment: 1) most places on earth have been covered by dozens or even hundreds of high-resolution satellite imageries. It is not appropriate to involve all imageries in the bundle adjustment. Firstly, due to bad geometric and radiometric configurations (e.g. maximum of nadir angle > 60° or sun elevation angle ≈ 0°), some imageries may bring high uncertainties in the bundle adjustment. Secondly, involving all imageries will be of high time complexity. Hence, how to select appropriate imageries that can satisfy both the accuracy and time efficiency demands is of a great interest. 2) Processing the large-format satellite images normally come up with a large amount of feature correspondences. Due to the computer memory limit and processing power, it is challenging to handle millions or billions of observations at the same time. Hence, how to efficiently realize the bundle adjustment in large-scale regions is very important.

To handle these challenges, we consider image selection, block division and combination in the bundle adjustment and propose a new method that can efficiently refine RPCs of multi-view large-scale satellite images. We firstly uniformly divide the entire mapping areas into blocks with overlapping regions between adjacent blocks. Secondly, we select high-resolution satellite images for each block. We only consider images that covering the block, from which we compute the completeness of the covering, geometric configuration (e.g. intersection angles), radiometric configuration (e.g. sun elevation angle), imaging time, number of corresponding pairs etc., and formula these factors as scores for each images. We select images from high to low scores until the block has been totally covered twice (to guarantee three-degree overlap in the bundle adjustment). Thirdly, we add a linear system error compensation term in the RPC model and adopt the conjugate gradient methods in the bundle adjustment to iteratively refine the RPCs and the system error compensation terms of each image in the block. After the bundle adjustment of each block, we think there is only Euclidean transformation among these blocks, and register these blocks in the same geodetic framework using absolute orientation models. Finally, we regenerate RPCs for each satellite images by considering these bundle adjustment results.

Experiments with a large number of Worldview 2/3 satellite imageries are performed in this work. To evaluate our image selection strategy, we firstly compare the time efficiencies and adjustment accuracies of our proposed method and the traditional method involving all images. Then, we compared our method with other multi-view large-scale bundle adjustment method. Based on the comparison, we can conclude that the proposed method can efficiently compute accurate, robust adjustment results for multi-view large-scale satellite images, and it is competitive when compared with other state of the art bundle adjustment methods.

11:45 Multi-View Large-Scale Bundle Adjustment Method for High-Resolution Satellite Images, Rongjun Qin