| 摘要: | 在某些情況和特定目的下，我們需要估算大型物件的體積，例如建築物、雕像、飛機等等。本文提出一套具備高度彈性且成本效益高的方法，使用來自低成本無人機拍攝的空拍影像、智慧型手機拍攝影像以及Google Earth 三維街景資料等多種來源的影像，搭配 RealityCapture 建立的高解析度點雲模型，實現對大型物件的體積估算。
本方法整合多階段的點雲處理與影像分析流程。首先，透過空間裁切及降噪技術清除背景與非目標物資訊，並使用 RANSAC 擬合地面平面並進行法向量校正，使點雲地面法向量與世界座標 Z 軸對齊，提升切片分析的準確性與一致性。接著，利用 DBSCAN 聚類演算法以及 KDTree 半徑檢查，去除地面與浮動雜訊點。為進一步提升低密度區域的完整性，本研究導入 Poisson 重建與輪廓補點技術，補足點雲稀疏區域，使整體點雲更為完整、連貫。在體積估算過程中，我們將點雲依 Z 軸高度切分為數個切片，並將每個切片投影至 XY 平面產生二值影像。為解決點雲密度不均所導致的輪廓破碎問題，本文引入自適應膨脹演算法，針對每個像素區域的局部密度動態調整膨脹核大小，並結合侵蝕處理修補輪廓與去除雜點。處理後的影像使用 OpenCV 擷取輪廓，計算出每層的切片面積，再乘以切片厚度並逐層累加，最後估算完整體積。此外，本文亦提供體積估算過程中的視覺化展示，包含每層切片的三維點雲與對應輪廓影像，有助於觀察估算細節與分析異常切片。
實驗部分，我們選擇多個真實世界大型建物作為驗證對象，包括中正紀念堂、陶朱隱園以及內湖區住宅，分別涵蓋不同複雜度與形體特徵的結構。將實際體積與估算體積進行誤差比對後，驗證本方法在多種資料來源與不同點雲品質條件下，皆展現出穩定且高準確度的體積估算能力。整體誤差控制在可接受範圍內，顯示本方法具有良好的適應性，也適用於資源有限、無須昂貴設備的應用場景。
;In certain situations and for specific purposes, it is necessary to estimate the volume of large objects, such as buildings, sculptures, aircraft, and so forth. This study proposes a highly flexible and cost-effective method that utilizes images from various low-cost sources—including aerial photographs taken by unmanned aerial vehicles (UAVs), images captured by smartphones, and 3D Street View data from Google Earth—combined with RealityCapture to construct high-resolution point cloud models for estimating the volume of large objects.
The proposed approach integrates a multi-stage workflow encompassing point cloud processing and image analysis. First, background and non-target information are removed through spatial cropping and noise reduction techniques, including plane fitting using RANSAC and normal vector correction to align the point cloud ground normal with the Z-axis of the world coordinate system, thus enhancing the accuracy and consistency of slice analysis. Subsequently, DBSCAN clustering and KDTree radius checks are employed to eliminate ground and floating noise points. To further improve the integrity of low-density regions, Poisson reconstruction and point completion techniques are introduced to fill sparse areas of the point cloud, making the overall data completer and more continuous. During the volume estimation process, the point cloud is sliced along the Z-axis into several layers, and each slice is projected onto the XY plane to generate binary images. To address issues of fragmented contours caused by uneven point cloud density, this paper introduces an adaptive dilation algorithm that dynamically adjusts the size of the dilation kernel based on the local density around each pixel region and combines it with erosion operations to repair contours and remove noise. The processed images are then used with OpenCV to extract contours, calculate the area of each slice, and multiply it by the slice thickness to cumulatively estimate the overall volume. Furthermore, the study provides visualization during the volume estimation process, including 3D point clouds and corresponding contour images for each slice, facilitating observation of estimation details and analysis of abnormal slices.
In the experimental section, several large real-world buildings are selected for validation, including Chiang Kai-shek Memorial Hall, Agora Garden, and residential buildings in the Neihu District, covering structures of varying complexity and geometrical characteristics. By comparing the estimated volumes with actual measurements, this method demonstrates stable and high-accuracy volume estimation capabilities across various data sources and point cloud quality conditions. The overall estimation errors are maintained within acceptable limits, indicating the method’s adaptability and suitability for applications in resource-constrained environments without the need for expensive equipment. |
