查詢結果分析
來源資料
頁籤選單縮合
| 題 名 | 應用深度特徵於多視角車載影像匹配之研究=Deep Feature Matching for Multi-View Images in Mobile Mapping Systems |
|---|---|
| 作 者 | 張智安; 陳沛丞; 陳亭霓; 楊軒; 林展慶; 李冠毅; 洪愷頡; 呂振永; | 書刊名 | 航測及遙測學刊 |
| 卷 期 | 30:3 2025.09[民114.09] |
| 頁 次 | 頁139-151 |
| 專 輯 | 航測及遙測工程應用特刊 |
| 分類號 | 312.831 |
| 關鍵詞 | 深度學習; 深度特徵; 影像匹配; 多視角影像; 交通標誌; Deep learning; Deep features; Image matching; Multi-view images; Traffic signboard; |
| 語 文 | 中文(Chinese) |
| DOI | 10.6574/JPRS.202509_30(3).0001 |
| 中文摘要 | 隨著移動測繪技術的進步,多視角車載影像逐漸成為道路觀測的重要資料來源,但傳統方法難以克服畸變與視角差異。為了提升影像匹配的精度和泛化能力,本研究探討基於深度學習的深度特徵匹配技術(Deep Feature Matching, DFM),利用VGG19預訓練模型與CNN卷積神經網路萃取特徵,並結合兩階段匹配策略與RANSAC演算法過濾錯誤點以確保可靠性,達成多視角車載影像的精確匹配和物空間三維定位。實驗採用三相機車載影像,以交通標誌作為定位目標。結果顯示,DFM在尺度差異、畸變與遮蔽等情境下,匹配成功率與精度均優於SIFT,特別在畸變與遮蔽下表現更佳。交通標誌定位成功率約70%,平均誤差小於0.5 m,證實DFM具備多視角三維定位的應用潛力。 |
| 英文摘要 | With the advancement of mobile mapping technology, multi-view imagery has become an important source for road observation. However, traditional matching methods struggle to overcome the challenges posed by distortions and variations in viewing angles between images. To enhance the precision and reliability of image matching, this study investigates a deep feature matching technique based on deep learning. By utilizing convolutional neural networks (CNN) to extract deep features, accurate matching and 3D spatial positioning of multi-view images can be achieved. The study employs Deep Feature Matching (DFM) technology, which is based on the pre-trained VGG19 model. Through a two-stage matching strategy and the RANSAC (Random Sample Consensus) algorithm, erroneous matching points are filtered out to ensure the reliability of the matching results. The experimental data consists of multi-view images, with traffic signs serving as the target objects for image matching and positioning. The research results reveal that, compared with the traditional SIFT method, DFM demonstrates a higher success rate and improved positioning accuracy in various image scenarios, including scale differences, shape distortions, and occlusion conditions. Notably, DFM achieves significantly more matching points than SIFT in distortion and occlusion scenarios. Furthermore, the analysis of traffic sign positioning indicates that the success rate of traffic sign positioning reaches 70%, with an average error of less than 0.5 m for successfully located points. This finding highlights the practical application potential of DFM in 3D positioning using multi-view images in complex scenarios and confirms that it achieves higher success rates and accuracy. |
本系統中英文摘要資訊取自各篇刊載內容。