都市區有人機傾斜攝影製作真正射影像精度評估與探討

隨著電腦視覺匹配技術之發展，航拍影像的正射糾正逐漸以密匹配產製的高密度點雲消除地物高度和影像傾斜所造成的像點位移，進而製作真正射影像，使地物像點以垂直投影呈現在正確的位置上；且隨著傾斜攝影技術的發展，航拍時可以減少更多都市區地物之遮蔽並讓製作的真正射影像更正確顯示地物（尤其是建物）邊界形狀；再加上深度學習之發展，已有研究使用有人機，即需飛行員駕駛飛機，航拍的都市區航拍影像製作的真正射影像執行相關研究，但尚未有研究探討都市區有人機傾斜攝影製作真正射影像的精度與問題。因此，本研究以臺北市信義區約300公頃範圍，設計垂直航拍地面解析度8公分，以傾斜相機獲取高重疊率的垂直與前、後傾斜影像，再以Agisoft Metashape Pro 1.2版本軟體執行密匹配點雲進而製作真正射影像，於10公頃範圍內選取23個均勻分布之檢核點，比較在立測地形圖上與在真正射影像坐標並計算其位置較差，顯示最大位置誤差0.125m、最小位置0.000m，均方根誤差0.052m，滿足千分之一地形圖之精度。但靠近建物附近地面處之真正射影像破碎模糊；靠近植被附近之真正射影像亦可能錯誤不正確；此外，陰影仍是真正射影像待解決的問題之一。

With the development of computer vision in image matching approaches, the orthorectification of aerial images are gradually using high-density point clouds produced with dense image matching to resolve the image point displacement caused by the height of ground objects and image tilt to further produce true orthoimages. Thus, the object image points can be presented at the correct locations. Additionally, with the development of oblique photogrammetry, more objects in urban areas can be imaged without occlusion and make the object boundaries in the true orthoimages more correct. Moreover, with the development of deep learning, many studies had used urban true orthoimages produced from manned-aircraft aerial images of urban areas, requiring a pilot to fly an aircraft, to perform related studies. However, no research has yet explored the accuracy and issues of producing true orthoimages using manned-aircraft oblique images in urban areas. Therefore, this study selected a test area of about 300 hectares in Xinyi District, Taipei City. An oblique camera was used to obtain vertical images of 8 cm ground sampling distance, forward and backward oblique images with high overlap. Then Agisoft Metashape Pro. version 1.2 software was used to perform dense matching point clouds to further generate true orthimages. Twenty-three evenly distributed check points were selected within 10 hectares to measure the coordinates both in topographic map and true orthoimage to calculate the position difference. The results showed maximum position error is 0.125m, the minimum position error is 0.000m, and the root mean square error is 0.052m, which meets the accuracy of 1/1000 topographic map. However, the orthoimage of the ground close to the building is broken and blurred; the true orthoimage near the vegetation may also be incorrect in addition, shadows are still one of the problems to be solved in the true orthoimage.