Residential buildings—where individuals spend the majority of their lives—require more effective tools for understanding, maintaining, and enhancing indoor environments. In contrast to commercial or public facilities, homes exhibit highly varied layouts, are densely furnished, and reflect occupants’ needs for privacy and comfort. Although 3D laser scanning and point cloud technologies have become increasingly prevalent in the Architecture, Engineering, and Construction (AEC), most processing methods are primarily designed for idealized, unoccupied, or synthetic spaces. Consequently, these methods often struggle when applied to the complex conditions of actual residential settings. In response to this methodological gap, this thesis is centered on advancing point cloud–based building assessment tailored specifically to occupied residential environments. The goal was to develop a pipeline capable of addressing clutter, occlusion, and spatial irregularity, rather than avoiding them. This work was guided by a central question: How can point cloud methods become more practical, robust, and scalable for real-world homes? To address this, the research is organized around three primary objectives, each forming a core chapter of the thesis. In Chapter 2, a benchmark dataset was developed specifically for residential environments. Four fully occupied homes and one under-construction unit with partial interior elements were scanned using both terrestrial and mobile laser scanning technologies to capture the spatial conditions as they exist in daily life. Rather than removing furniture or minimizing environmental noise, elements such as closed curtains, kitchen cabinetry, and multifunctional rooms were retained to reflect real-world complexity. The benchmark comprises five evaluation tasks—area measurement, floorplan generation, adjacency graph creation, navigation planning, and evacuation path modelling—accompanied by standardized evaluation metrics. The objective was to provide a resource enabling researchers to evaluate methods under authentic residential conditions. In Chapter 3, two state-of-the-art floor plan generation methods were reimplemented and tested using the proposed benchmark. Although previously validated on clean or synthetic datasets, both methods encountered significant challenges when applied to residential data. One failed during the geometric reconstruction phase due to the presence of furniture and clutter. The other generated usable outputs but exhibited sensitivity to vertical segmentation, occlusion, and parameter tuning. These limitations were traced to the original validation conditions—typically simplified and controlled environments—and contrasted with the complexities present in the benchmark. The results indicate that while existing methods show potential, they do not generalize effectively to realistic residential settings. In Chapter 4, a new processing pipeline was developed to address the limitations identified in existing methods by directly extracting spatial and topological information from raw point cloud data, without relying on additional images, full 3D reconstruction, or pretrained models. The proposed method integrates vertical slicing, mesh generation, dimensional reduction, space segmentation, and boundary detection to generate area measurements, adjacency graphs, and floor plans. The pipeline was intentionally designed to be lightweight, modular, and adaptable to varying residential building layouts. To validate the proposed method, all five benchmark datasets were used for testing. The results indicated effective performance, with an average precision of 93.5\% in floorplan generation and a root mean square error (RMSE) of 0.49 m² for area measurements. Adjacency graphs were qualitatively assessed through visual comparison with as-built 3D models. The evaluation demonstrated that the method managed layout irregularity, clutter, and wall occlusion more effectively than existing approaches. Each chapter contributes to addressing the main research question; the benchmark, evaluation, and proposed method form a structured, end-to-end framework for real-world residential space analysis, supporting applications such as smart homes, facility management, and emergency planning. This research contributes by advancing spatial reasoning, adjacency modelling, and clutter-aware segmentation in residential indoor environments. It also introduces a new benchmark dataset grounded in real living conditions rather than organized representations. Furthermore, the work provides scalable solutions for applications such as automated building assessment, indoor navigation, and retrofitting analysis without requiring extensive manual annotation or supplementary imaging. Looking forward, the limitations encountered—such as the need for manual parameter tuning, the limited scope of the dataset, and the absence of full automation—highlight valuable directions for future research. Moreover, there is potential to enhance the method through integration with object detection, semantic segmentation, and learning-based adaptation. This thesis reflects a broader vision: to align point cloud research more closely with real-life environments—homes, people, and the spaces we navigate daily. It is intended to lay a foundation for more inclusive, intelligent, and practical applications in residential settings.
Abdullah Elsafty (Thu,) studied this question.