二維空間圖資套合基於軌跡式直線特徵

Abstract

二維空間圖資套合需要可靠的共軛特徵，用來解算轉換參數，傳統方法倚賴點特徵，但現今圖資種類多元，資料之間的差異性增加共軛點特徵獲取困難。相較之下，直線特徵具有較豐富幾何屬性與易萃取特性，利用直線軌跡重合條件，可作為共軛特徵。 本研究以向量式直線方程式推演基於直線軌跡二維轉換模式，轉換模式包含相似轉換、仿射轉換與射影轉換，皆具有直線保留特性，透過幾何解析和數值模擬方式分析轉換參數最小求解條件，並以最小二乘平差方法，估計轉換參數的最佳解，資料套合品質評估方式除了檢核點，還納入檢核線指標，使成果評估方式更全面。 模擬實驗分析影響套合成果精度的因子，分為直線特徵幾何強度和觀測量品質兩類。影響直線特徵幾何強度的因子包含分布、數量、方向，而觀測量品質受到線段長度和量測品質兩個因子影響。實驗成果顯示直線特徵的分布、方向兩項因子對參數解算成果的影響最顯著。 實際資料實驗以一組近景影像資料進行套合，基於射影轉換模式，使用19個直線特徵，檢核點均方根誤差可達到0.39像元，套合效果能達到與使用點特徵者相近。另外，以不同時期之1/1000數值地形圖為測試資料，基於相似轉換進行套合，使用7個直線特徵，檢核點均方根誤差為0.348公尺，混合7個直線特徵與5個點特徵求解轉換參數，增加套合成果可靠度，檢核點均方根誤差可達0.220公尺。

The registration of two-dimensional geo-spatial data depends on conjugate features and point features are generally considered for this need. Since the enormous increase in the volume of geo-spatial data, the difference between various types of data makes it difficult to extract conjugate point features. In addition to point features, line features, with frequent appearance in the scene and more geometric contribute, are deemed favorable for being primitives under collinear condition. This study exploits the vector form line features and derives trajectory-based transformation models, including similarity, affine and projective, which are all line-preserving transformations. Besides, the minimum solution of solving transformation parameters is analyzed by geometric and numerical approach. The parameter estimation is based on the least squares adjustment. Some line features are also used for evaluating quality of transformation to assess comprehensively. The simulation analyzes the impact of line feature-related factors, divided into geometry and observable quality. The geometry-related factors include distribution, number and orientation of line features. The observable quality is affected by length and measurement quality. The simulated results show that distribution and orientation of line features have significant influence on transform quality. In experiment, the registration of close-range imagery pairs using 19 line features with projective transformation can reach 0.39 pixel errors, which is close to point-based one. Moreover, the registration of two 1/1000 digital topographic maps using 7 line features based on similarity transformation have 0.348 meter errors. If 7 line features with 5 point features are used as conjugate features, the root mean square errors could decrease to 0.220 meters.