空照影像密匹配成果偵錯之瓶頸與解決辦法

Bottlenecks and Solutions of Blunder Detection on Dense Matching Results of Aerial Images

李硯婷(Yen-Ting Lee) ； 蔡展榮(Jaan-Rong Tsay)

《航測及遙測學刊》 22卷3期 (2017/09) Pp. 137-155

https://doi.org/10.6574/JPRS.2017.22(3).1

摘要

在攝影測量領域中，影像匹配技術已發展至密匹配(dense matching)的新紀元，此時密匹配成果之偵錯與品質評估面臨一些瓶頸，包括(1)無原始匹配點像坐標、(2)匹配點數量龐大、(3)相鄰匹配點之距離太近而產生相關參數的高相關，導致解算不穩定之現象。本文提出並使用四種簡易實用的密匹配成果偵錯與品質評估法，包括目視檢查法、相對方位法、像片三角法以及獨立測量法。使用的密匹配演算法為SMM、SfM、DAISY及SGM，偵錯成果顯示50.72%和47.00%的密匹配錯誤分別出現於高程1階不連續面(山形屋脊線)與0階不連續面，僅0.05%出現於均調區，此經驗可供密匹配研究與應用之參考。相對方位計算成果得到SMM、SGM錯誤率分別為2.82 %、2.36 %。經像片三角法之評估，SMM匹配精度為0.23 pixel；錯誤率為3.97 %。相對方位與像片三角法之偵錯速度分別為14,984 匹配點/秒、292 匹配點/秒。獨立測量法使用12個地面檢核點檢查成果顯示，SGM密點雲與佈標點高程之差值絕對值，最大值為0.935 GSD、最小值0.006 GSD、平均值0.315 GSD、RMSD等於0.238 GSD，此處的1 GSD=0.168 m。

關鍵字

密匹配 ； 偵錯 ； 品質評估

並列摘要

In photogrammetry, the development phase of image matching is moving into a new era, namely the dense matching. At this phase, there are bottlenecks for blunder detection and quality evaluation of dense matching results, including (1) no output on photo coordinates of matching points, (2) a huge number of matching points, and (3) unstable bundle block adjustment caused by too close matching points. This paper proposes four easy and applicable methods to overcome those bottlenecks. They are visual check, relative orientation (RO) using a huge number of tie points, bundle block adjustment, and comparing with check points. Four matching algorithms are tested, including SIFT-based Multi-image Matching (SMM), Structure from Motion (SfM), DAISY and Semi-Global Matching (SGM). The results show that 50.72% and 47.00% of wrong matching points appear at roof ridge lines and break lines. Only 0.05% of those wrong matching points are located in homogeneous color area. This might be for reference for development and application of image dense matching. Test results of RO method demonstrate that SMM and SGM have the wrong matching rate of 2.82% and 2.36%, respectively. The evaluation results of bundle block adjustment method show that the matching accuracy and wrong matching rate of SMM are 0.23pixel and 3.97%, respectively. Both RO and bundle block adjustment methods have the computation speed of 14,984 points/second and 292 points/second, respectively. By comparing 12 check points, the object points measured by SGM have the absolute elevation differences with maximum 0.935GSD, minimum 0.006GSD, average 0.315GSD and RMSD 0.238GSD, where 1GSD=0.168m.

並列關鍵字

Dense Matching ； Blunder Detection ； Quality Evaluation