透過您的圖書館登入
IP:18.206.12.31
  • 學位論文

運用街景影像及物件語意關係建構深度學習模型進行建物騎樓偵測與製圖

Arcade Detection and Mapping Based on a Deep Learning Model Constructed Using Street View Images and Object Semantic Relations

指導教授 : 賴進貴
共同指導教授 : 郭巧玲(Chiao-Ling Kuo)

摘要


騎樓係由一系列門拱結構而構成的特殊建築樣式,廣泛分佈於東南亞地區,並儼然已發展成一種重要的建築風格。在文化上,騎樓已成為了許多東南亞國家和地區的典型建築文化景觀;在功能上,騎樓是城市中步行空間的要素之一,可在熱帶和亞熱帶氣候區為行人提供遠好於露天人行道的通行體驗,因此騎樓分佈是都市規劃和空間資訊等相關領域發展與應用之重要資訊。然而在目前各類資源中,所能提供的騎樓地理資料仍然很少,這導致空間規劃與分析無法完備;而對於騎樓位置的調查或資料建置,仍以人工踏勘調查方式為主,需要耗費大量建置成本。有鑑於此,發展高度自動化之騎樓資料產製,以利實際規劃與分析,並可供一般行人作為出行參考,有其必要性。 街景影像具有覆蓋範圍廣和獲取便捷等優勢,近年來,以深度學習為代表的人工智慧技術蓬勃發展,尤其於影像辨識有相當卓越之表現。因此本研究基於深度學習方法、結合街景影像資料,發展出了一套高度自動化的騎樓物件偵測方法。其以YOLOv5物件偵測算法為核心,集成物件語意關係學習模塊,整合Google街景影像、開放街圖和1/5000地形圖等資料,以實現對騎樓的高效偵測與製圖。本研究以台北市中心的7個行政區為實驗區,實驗結果顯示,在自建的騎樓影像資料集上,偵測結果擁有超90%之平均精度(AP)。此外,本研究亦實現以測繪原理演算偵測騎樓,完成實驗區內的騎樓製圖,製圖結果的整體誤差約在0.6M,可以滿足步行導航等任務之需要。 本研究驗證了物件偵測演算法對Google街景中騎樓建築這種抽象程度較高之物件的偵測能力;亦進一步明確了Google街景用於測量計算任務的精確度。本研究成果除有助於創造更友善的都市步行環境外,亦可作為整合人工智慧與空間資訊方法研究之參考。

並列摘要


An arcade is a special architectural style consisting of a variety of arch structures, and is widely distributed in Southeast Asia, where it has become an important architectural style. In terms of culture, the arcade has become a typical cultural landscape in many Southeast Asian countries and regions; in terms of function, the arcade is one of the significant ways of offering walking space in cities. Arcades can not only provide a much better experience than normal uncovered sidewalks for pedestrians in tropical or subtropical climates but also play an important role in the development and applications of urban planning. However, arcade data are still scarce, which hinders the spatial planning of a city. At present, arcade surveys and data mapping are mainly manual, and are thus costly. To overcome the obstacles mentioned above, it is necessary to develop a highly automatic means of data production for arcade geodata to boost urban planning and geo-spatial analysis, and for pedestrians to use as a reference for travel. There are many advantages of street view images, including extensive coverage and simple access. Recently, artificial intelligence (AI) technology, deep learning in particular, has been widely explored and applied, especially in image recognition. Therefore, this study proposes a highly automated object detection method based on the deep learning method using street view images to explore arcades and achieve arcade mapping. The YOLOv5 object detection algorithm is used as the core, and an object semantic relationship learning module is used to provide others with a confidence score. This method is also integrated with Google Street View images, Open Street Map (OSM) and 1/5000 topographic maps to achieve efficient detection and mapping of arcades. The study area covers seven most densely populated districts in Taipei City, Taiwan. The study results show that the average precision (AP) of our arcades dataset is more than 90%. Furthermore the detection results were mapped by using a mapping algorithm. The overall average error of the current technique is about 0.6 m, which meets the requirements of pedestrian navigation and urban development research. This study verifies the ability of the object detection algorithm to detect objects with a high degree of abstraction, such as arcades in Google Street View; it also further specifies the accuracy of Google Street View for measurement computation tasks. The results of this study not only can be used to improve the urban pedestrian environment, but also serve as a reference for integrating artificial intelligence and spatial information technology.

參考文獻


中央社. (2022). 花蓮騎樓攤商頻遭開罰議員籲顧生計縣府允研議
內政部. (1947). 建築管理規則.
王怡青. (2017). 從步行者觀點探討使用者對「臺南市騎樓暢通計畫」政策認知及滿意度­以北門路為例(碩士學位論文). 國立中山大學. 高雄市.
史寶剛, 尹海偉, 孔繁花, 劉佳. (2021). 基於 ENVI­met 模型的南京市新街口地區垂直綠化降溫效應評價. 現代城市研究, (12), 125–132.
吳欽憲. (2009). 騎樓的景觀與社會空間識覺——以臺北市為例 (碩士學位論文). 國立臺灣師範大學.

延伸閱讀