應用深度線特徵於建築環境内之點雲套合

Abstract

點雲相比於影像，具備三維的特徵且更能反映現場真實狀況，能幫助許多工程上的應用，例如進度管理、品質管理等等。在點雲的應用上，點雲套合是一件重要的前置任務。現今，我們需要將點雲手動整合到統一的座標系統，這是一項重複性且耗費人力的工作，而自動化的點雲套合受困于難以在低重叠、跨源頭的點雲取得對應特徵，進而影響其套合成效。本研究基於在建築環境内多是結構線的緣由，提出一個以線特徵進行點雲自動套合的框架，此框架分由4個模組組成，包括線的分割 (Line Extraction)，線特徵的萃取與配對 (Line Feature Extraction and Matching)，離群值檢驗 (Outlier Pruning)和線的套合模組 (Line-based Registration)。此外，我們設計了一個深度神經網路（LineNet），以進行線的特徵萃取和配對。我們的線特徵套合框架在點雲套合的資料集（3DMatch）上達到26.46%的套合成功率，超過目前最好的線特徵套合模型。最後我們也對使用線特徵進行套合的重要因素進行討論，其中包括針對各個模組的品質因子的討論。

Point clouds have a broad range of applications in construction monitoring and management, such as progress monitoring and quality inspection. Nonetheless, point cloud registration is a crucial task before most point cloud applications. Currently, the alignment of point clouds into the same global coordinate system largely requires manual work, which is repetitive and labor-intensive, and the automated point cloud registration suffers from finding correct correspondences between cross-source or low overlapping point clouds. The salient and abundant line geometry in man-made structures contains the potential to automate the registration process. This research aims to perform point cloud registration by using line geometry as a matching feature. We propose a line-based point cloud registration framework and design the LineNet, a line feature encoding network to perform the feature extraction and matching. We experimented with the proposed framework on the representative 3DMatch dataset, and the overall registration recall is 26.46%, which have great improvement against the latest line-based method. Some discussions on the important factors to perform a good line-based point cloud registration have been made.