應用深度學習於衛星影像道路標線萃取之研究

Abstract

道路是地理資訊系統中經常使用的資料項目之一，過去已有需多研究著重於從航遙測影像中萃取出道路的空間位置，然而許多與道路相關的屬性資料，如道路寬度、車道數量、行車方向等，大多仍需由人工進行判釋，再添加至道路圖層的屬性欄位中，難以快速建置數量龐大的道路圖資。為了因應快速變遷的空間資訊，保持道路圖資的完整性及時效性為一相當重要的課題。因此，本研究將針對高解析度衛星影像，建立一個自動萃取道路標線資料的流程，以快速建置完整的道路資訊。 研究流程包含影像前置處理程序及針對不同道路標線類別資料之標記方法，首先以人工標記方式，從遙測影像中建立不同道路標線類別之訓練資料集，訓練類別包含車道線、行車分向線、機車停等區、枕木紋行人穿越道、以及五種方向之指向線等。再以YOLOv4深度學習演算法訓練各類別的物件偵測模型參數，接著對各類別偵測結果進行準確度分析，最後特別針對線型道路標線進行線段偵測及萃取流程以獲得向量格式之成果。研究成果顯示，以深度學習偵測線型及非線型道路標線時，F1-score分別可達96%及92%，呈現不錯的分類精確率；而在道路標線向量化的部分，透過邊緣偵測及細化等影像處理程序可成功萃取出車道線及行車分向線，F1-score可達92%，精度評估部分平面均方根誤差可達20公分以內。

Road information is frequently used data in Geographic Information System. In the past, much research has focused on extracting the spatial location of roads from remote sencing images. However, many attribute data related to roads, such as road width, number of lanes, and traffic directions, etc. still require manual interpretation and addition to the attribute fields of road layers. This makes it difficult to rapidly build large-scale road map data. In order to cope with the rapidly changing spatial information and maintain the completeness and timeliness of road map data, this study aims to develop an automated process for extracting road markings from high-resolution satellite images to quickly build comprehensive road information. The research process involved image preprocessing procedures and marking methods for different road marking categories. First, a training dataset for different road marking categories was established using manual annotation from remote sensing images. The training categories include lane lines, traffic divergence lines, scooter waiting areas, pedestrian crossing, and direction lines in five directions. Then, the YOLOv4 deep learning algorithm was used to train object detection models for both non-linear and linear road markings, along with the analysis of the accuracy of the detection results for each category. Finally, a specific process for line detection and extraction was developed for linear road markings to obtain results in vector format. The research results showed good classification accuracy, with F1-scores of 96% and 92% for the detection of linear and non-linear road markings, respectively. Furthermore, through image processing procedures such as edge detection and thinning, the road lane markings and directional markings were successfully extracted and vectorized. The F1-score achieved a rate of 92% and planar root mean square error are within 20 centimeters.