Abstract as per original application
(English/Chinese): |
This project aims to apply bi-objective optimization and graph learning to detect architectural openings, such as 3D hollows and maneuvering clearances installed with windows and doors, from point clouds. Openings admit air and light into buildings and, thus, are essential for any livable architecture. Recent advances in LiDAR technology resulted in affordable 3D point clouds on existing buildings. However, LiDAR points are handicapped in surveying openings’ functional hollows and movable components, leading to missing data (e.g., owing to transparency of glass and air), probabilistic data (e.g., multiple open and closed door states during scanning), and ambiguous data (e.g., hard-to-distinguish voids between space, hallways, and openings).
Thus, a dilemma exists between the LiDAR-invisible hollows of openings and LiDAR's limitations. Consequently, the dilemma undermines LiDAR's effectiveness in architectural opening surveying, which is necessary for 3D indoor space networks, visual access to nature, natural lighting 3D maps, local thermal comfort spots, carbon emissions, etc. The dilemma also challenges researchers in other disciplines, e.g., instructing a household robot to open the second drawer or living room window.
The major innovation of this project lies in the bi-objective optimization formulation that addresses two objectives, (i) local semantic graphs of openings and (ii) global design regularity of 3D hollows, to resolve the dilemma. First, dense 3D points of functional hollows are added to LiDAR-visible point clouds to create opening and neighboring supervoxel graphs of interest (ONSI). GPU-powered graph learning labels ONSI’s semantics, e.g., door hollow, maneuvering clearance, window, and indoor space. The global design regularities include symmetry, repetition, and collinearity with other openings. Pareto optima indicate candidate architectural openings. The feasibility of this project was partially verified in a small pilot of ONSI graph learning. The project will extend the study scope, incorporate both objectives, and gauge the overall approach to Hong Kong’s and other smart cities’ datasets.
This project delivers both academic insights and practical advantages. The findings will expand our understanding of how modern optimization and GPU-powered graph learning can be integrated to process LiDAR data. The explainable local ONSI graphs and global regularity patterns may contribute to a breakthrough in theoretical explanations, regarding how architectural domain knowledge can systematically overcome the limitations in LiDAR data, in the case of openings detection. For professionals, the research outcomes will be integrated into a user-friendly LiDAR processing tool for multiple disciplines and industries, e.g., architecture, construction, engineering, automotive, and robotics.
本項目運用雙目標優化與圖學習技術，從點雲數據中檢測建築開口。LiDAR（激光雷達）技術提供新型建築三維點雲數據，然而，LiDAR在檢測開口存在局限（如透明材質數據缺失、可移動部件的多狀態數據不確定性、及不同類孔洞數據歧義）。本項目的首要創新在於提出一類雙目標優化模型，以解決這一難題。該模型同時考慮兩個目標：（i）開口的局部語義圖，及（ii）三維孔洞的全局規律性。首先，將功能性孔洞的不可見三維點集補充到LiDAR可見的點雲中，生成關注的開口與鄰近超體素圖。基於圖學習技術標注體素圖的語義，例如門洞、操作空間、窗戶和室內空間。全局設計規律性包括對稱性、重複性以及與其他開口的共線性。通過帕雷托最優（Pareto optima）確定候選建築開口。本項目的可行性已在一項小規模的體素圖學習試點中得到部分驗證。項目將進一步擴大研究範圍，整合雙目標，並在香港及其他智慧城市的數據集上評估整體方法。研究成果將深化優化與圖學習技術處理LiDAR數據。可解釋的局部體素圖與全局規律模式可能展示建築領域知識系統性地克服LiDAR技術在特定檢測中的局限，適用於建築、施工、工程、汽車及機器人等多個學科與行業。
|
