以分割法改善應用YOLO之航測標記點檢測系統

近年來由於無人飛行載具盛行，許多行業都開始跟隨了這波熱潮，航空攝影測量是最近幾年流行的測量技術，解決了以往傳統測量的不方便性，並透過高精度的模型等比例還原該地區的面貌，進而做出測量，但是單純的航拍攝影無法掌控模型的精確程度，因此需要在拍攝前去佈置航測標，來提高建置模型的精度與時間，而自從多視點影像技術出現後，許多相關的軟體對於這些標記點能夠高精準度的標記航測標的中心位置，透過這樣的功能解決以往傳統航空攝影測量的不方便性，例如建模前需要花費大量人力時間進行人工刺點，但是多視點影像技術不是毫無缺點，因為主要演算法採用SIFT，此算法的缺點在於辨識上需要許多條件才能達成所謂的高精度辨識。

深度學習是現今的熱潮，許多行業都結合此技術開發相關應用，但目前測量業並未有相關應用，原因在於台灣相關行業目前還是以傳統測量為主，航空攝影測量為輔，再加上需要具備相關專業知識，因此本研究嘗試使用深度學習中的物體檢測技術來辨識航測標，模擬自動化標記點且改善拼接軟體中的缺點，本研究採用YOLO作為物體檢測的模型，擁有快速且完整的權重模型，現今的技術講求的是高幀數的辨識，對於錨點與目標範圍還有待加強，而航測標的技術是透過照片來生成模型，因此可以不需改變YOLO內部結構。本研究以四分割法來提高模型效能，透過自定義非極大值抑制(NMS)將多餘錨點消除，對於講求精度的測量界來說，直觀上置信度高低也就決定精度高低，如果讓模型的置性區間提高，預測位置與交集面積相互重疊率越高，則獲取物體中心點的機率就越高。

In recent years, Unmanned Aerial Vehicle have been prevalent in the world and have been applied in many industries. Aerial photogrammetry using the newest measured approach to solve the inconvenient measurement previously, and return to the original features through one to one scale high precision models. However, because general aerial photogrammetry cannot control precision of models, we need to do targeting to increase accuracy and time in model development before taking the pictures. Since Image Synthesis Technology for Multi-view 3D Display appeared, numerous software can mark targeting central position correctly, and it reduce the time of artificially marked point. Nevertheless, Image Synthesis Technology still have a disadvantage, its algorithm adapts SIFT system. This mean that there must have plenty of recognizable conditions so that it can achieve high accuracy identification.

Nowadays, deep learning is the main trend that variety of industries apply; however, measurement industry has not executed any kinds of event. The reason is that related industries mainly rely on traditional measurement supplemented by Aerial photogrammetry in Taiwan. As a result, this research using object detection with deep learning to identify targeting, also to simulate mark automatically, and improve disadvantage of SILF. YOLO having fast and completed weight model; nevertheless, it bounding box function cannot achieve high frame identification for object detection. People do not need to change YOLO internal structure. Increasing efficacy by four spilt method, and removing unnecessary anchor points by Non-Maximum Suppression (NMS).