點雲自動建構向量模型之策略

Automatic Point Cloud Structuralization for Vector Models

宋政哲(Cheng-Che Sung) ； 莊子毅(Tzu-Yi Chuang)

《航測及遙測學刊》 26卷1期 (2021/03) Pp. 31-43

https://doi.org/10.6574/JPRS.202103_26(1).0003

摘要

本研究基於深度學習技術提出物件角點萃取模型及角點向量化模型，並針對建物點雲中的幾何結構，如板、梁、柱、牆作為向量化之標的物，建立點雲轉換向量模型的自動化機制。經實際測試，本研究在物件分類品質的平均精確度可達到50%以上，物件角點位置之誤差最高不超過25 cm，其中梁柱類別的角點誤差更小於10 cm，而在向量模型的角點連結正確率可達70%以上。顯示本研究方法能有效地將構件點雲自動地轉換為向量模型，同時賦予物件之語義屬性，模型成果可視為後續加值應用或成果精煉的基礎模型，達到提升自動化作業之效益。

關鍵字

點雲結構化 ； 三維深度學習 ； 建築資訊模型 ； 角點偵測 ； 角點向量化

並列摘要

This study proposes a novel learning-based point cloud structurization method especially for building components such as plates, beams, columns, and walls. The proposed model net learns from the point clouds generated by existing building information modeling (BIM) components to predict the geometric model of a newly given point cloud consequently. It is worthy to note that the BIM-to-Point cloud approach overcomes the difficulty of 3D training data acquisition that is usually confronted in most deep learning applications. The average precision (AP) of each category achieved a precision of above 50% in the classification task, in which the wall category yielded a precision of 77%. On the other hand, the predicted object corner positions reported an accuracy better than 25cm, in which the beam category revealed an accuracy of up to 10cm. Moreover, the quality of the geometric model in vector-linking prediction reached a precision of 70% suggesting that the proposed learning-based method can indeed reconstruct the geometric model of a given point cloud automatically.

並列關鍵字

Point Cloud Structuring ； 3D Deep Learning ； Building Information Modeling ； Corner Detection ； Corner Vectorization