精化多視角影像密匹配及點雲產製

Refinement of Multi-view Dense Image Matching and Point Cloud Generation

劉宣萱(Hsuan-Hsuan Liu)

指導教授 ： 趙鍵哲

國立臺灣大學/工學院/土木工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202003751

摘要

多視角影像密匹配係透過不同視角之多張影像，搭配密匹配技術高重疊率的設計需求，以提供高品質、大規模尺度及具多餘觀測描述的三維場景點雲。然而，大量多視角影像於計算處理上較為複雜而繁瑣，且針對每一像對分別計算其初始視差值，所需耗損的時間成本亦相對增加；再者，具多組像對重疊條件之多視角影像，倘未善加調製交會幾何，所產製之場景點雲即便具描述幾何之多餘觀測特性，然而點位的不精確性及較大的離散度無法有效助益於後續空間資訊產製之任務。對此，本研究提出一系列優化作業模式解決前述發展困境，首先將影像所提供之交會幾何納入考量，於半全域式匹配法進行前，約制透視中心所提供之幾何條件來量化像對間相似性，並透過影像獲取之配置從中篩選得以滿足預期匹配成果及點雲精度之立體像對組配；之後採以視差傳遞策略降低反覆給定單一像對初始視差值的缺失，並藉由縮小搜尋範圍降低計算耗損，達致兼顧場景重建幾何質量及覆蓋率的同時，亦能加速整體計算效能。此外，本研究透過誤差傳播計算前交三維點位之誤差並結合特徵訊息，產製點雲點位坐標、坐標品質資訊以及點位幾何屬性。據此，除了能刪除錯誤點位或削減其後續應用影響性之外，並能從“減點”的角度依使用需求，選用符合場景表達及重要幾何特徵描述條件之點雲點位，達到點雲精化功效。本研究所擬方法業經兩組實際資料驗證其功效及可行性，所產製之點雲能有效描述場景幾何；另外，於兩測試區域之像對計算總量部分：減少約為44%及14%之效益，而針對其時效提升部分則約為82%。

關鍵字

多視角 ； 影像密匹配 ； 點雲 ； 精化

並列摘要

High-quality, large scale and reliable 3D scene reconstruction can be achieved by dense image matching technique with multiple viewing angles and highly overlapping image acquisition design. However, processing large number of multi-view images is complicated and tedious. Also, when matching multiple stereo pairs, it would take long in getting disparity values if each pair is to be processed independently. In addition, redundantly described scene models with low reliability trouble the exploitation of geospatial information. This paper proposes a two-stage matching strategy to tackle the aforementioned shortcomings. First of all, the key stereo pairs in support of reliable disparity estimation are selected based on the geometric similarity measure viewed from perspective centers. And then, it is followed by propagating the disparity with good estimates attained from semi-global matching on these key pairs to the overlapped pairs that provide better intersection geometry for quality point cloud generation upon matching. In the meantime, the coverage of the scene by the matched pairs has been kept on tracking to satisfy the completeness of 3D reconstruction. Upon the selection of only those pairs with better intersection geometry, much fewer points have been generated, the “point cloud reduction” but with sufficient positioning quality and geometric attributes for deriving geo-spatial information has been realized. The proposed approach has been tested by practical data sets and it is proven that both the efficient image manipulation with the computational time reduction up to 84% and quality point cloud generation well depicting the scene geometry highlight the merit of this study.

並列關鍵字

Dense image matching ； Multi-view ； Point cloud ； Refinement

參考文獻

Ahmadabadian, A. H., Robson, S., Boehm, J., and Shortis, M., 2013. Image selection in photogrammetric multi-view stereo methods for metric and complete 3D reconstruction, Videometrics, Range Imaging, and Applications XII; and Automated Visual Inspection, Vol. 8791, pp.879107.

Chuang, T. Y., Ting, H. W., and Jaw, J. J., 2018. Dense stereo matching with edge- constrained penalty tuning, IEEE Geoscience and Remote Sensing Letters, 15(5):664-668.

Dall'Asta, E., and Roncella, R., 2014. A comparison of semiglobal and local dense matching algorithms for surface reconstruction, ISPRS - International Archives of The Photogrammetry, Remote Sensing and Spatial Information Sciences, XL-5, pp. 187-194.

Dominik, W., 2017. Exploiting the redundancy of multiple overlapping aerial images for dense image matching based digital surface model generation, Remote Sensing, 9(5):490.

Galliani, S., Lasinger, K., and Schindler, K., 2015. Massively parallel multiview stereopsis by surface normal diffusion, In Proceedings of the IEEE International Conference on Computer Vision, pp. 873-881.