變通式相機參數及像點精化估計

Abstract

近年來非量測型相機為攝影測量實務應用中常用的設備之一，為顧及攝影測量品質，通常以相機率定程序確認像點精化項目及精化量值，並修正物像對應關係。然而相機在獲取影像時，尤其是變焦式相機，可能因拍攝場景率定困難，或當時無自率光束施行條件，無法當場進行率定；另對於無段式變焦鏡，若選擇非預設刻度的像主距，其數值無法事後回復。這些狀況使得相機率定工作難以進行或者獲致不可靠之結果。 本研究綜合上述實務需求及資料處理彈性，使相機在諸多像主距下建構率定結果資料庫，依據觀測資料分析相機參數及像點精化分別與像主距的多項式函數關係，配合考量表頭檔像主距誤差、觀測量品質以及決定各項參數的最適擬合階數，推求符合任何資料型態的相機參數擬合模型及像點精化改正模型，建立當率定工作無法施行、免於施行條件下或像主距無法回復之像點精化有效替代方案，以產生等效或接近等效的像點精化成果。 基於實務考量並透過實驗成果顯示，在任何資料型態下，加入表頭檔像主距誤差及觀測量精度，能增加資料的可信度並提升擬合精度；另外，相同資料型態若採用固定單一的多項式函數，可能使部分擬合成果的偏誤較大，因此逐一針對不同項目及最適擬合階數的自動判定，將可有效減少擬合成果的偏誤。除此之外，本研究亦建構多項式總體精化擬合方法，經實驗顯示其像點精化之成果品質優於經由相機個別參數擬合化算者，展現模式簡化、程序便捷及提供足夠品質之實務應用潛力。

Non-metric digital cameras have recently gained their increasing popularity in photogrammetric applications. To achieve quality performance of photogrammetric tasks, the estimation of camera parameters, among others, is of great concern. Camera calibration is designed to determine camera parameters, including interior orientation parameters and distortions, for effectively refining the image point so that object-to-image correspondence under collinearity property can be well justified. There are, however, some situations where camera calibration, especially for zoom-dependent cameras, is hard or impossible to operate, or the focal length employed in the site can not be fully preserved elsewhere, making camera calibration inapplicable or unreliable. Therefore, alternative ways of supplying camera parameters must be considered. This research employed correction models, instead of actual calibration, to determine the camera parameters by referring to the recorded calibrated data sets of the very same camera on different principal distances. Two types of model have been formed. One features in estimating each kind of camera parameters in a separate fashion, while the other integrates all parameters and forms an effective polynomial function to estimate the overall amount of refinement. It is revealed from the experimental results that both models offer satisfactory estimations for image point refinement, especially when the quality of the calibrated data sets, the uncertainty of the focal distance of the target shown in metadata, and the best fitting order are taken into consideration through least-squares adjustment. Furthermore, the second model where the overall refinement is achieved by a single polynomial function gains better refinement than the first one, suggesting a convenient and sufficient alternative for image point refinement under no actual camera calibration.