Towards an automated AI-based framework for floor plan compliance checks for residential buildings

Technical Summary: Towards an Automated AI-based Framework for Floor Plan Compliance Checks for Residential Buildings

Problem Statement
The paper addresses the critical challenge of evaluating residential apartment design quality against government policy reforms (e.g., SEPP 65, BADS, SPP 7.3) aimed at improving resident well-being in Australia. Current compliance checking is a manual, labor-intensive process requiring human experts to interpret complex architectural drawings (often raster images or PDFs) against textual regulations. This manual approach suffers from scalability issues, making large-scale assessments (e.g., across thousands of apartments) impractical. Furthermore, existing automated methods are fragmented; they typically focus on single-apartment semantic analysis or building element identification but lack a unified framework capable of handling multi-unit floor plans, extracting structured data, and performing rule-based compliance reasoning simultaneously. The diversity of file formats, drawing conventions, and the presence of shared circulation spaces in high-rise buildings further complicate automated interpretation.

Methodology
The authors propose a conceptual framework for automated compliance checking that integrates computer vision, natural language processing, and graph-based reasoning. The system is structured into three primary engines:

1. Data Extraction Engine: This module converts raw floor plan images (PDF, PNG, JPEG) into a structured, machine-readable building model.

   • Preprocessing: Converts multi-page PDFs to high-resolution images and applies enhancement (noise reduction, sharpening) to improve feature visibility.
   • Advanced Text Recognition (OCR+): Extracts textual metadata (scale, room labels, dimensions) using libraries like PaddleOCR or Tesseract. Crucially, it recovers the drawing scale to convert pixel measurements into real-world dimensions.
   • Multi-task Deep Learning: Utilizes Convolutional Neural Networks (CNNs) for object detection (fixtures, doors, windows), semantic segmentation (walls, rooms), and instance segmentation to distinguish overlapping objects. It also employs computer vision techniques to detect the north arrow for building orientation analysis, which is vital for solar and ventilation checks.
   • Structured Building Model: Transforms the extracted visual and textual data into a JSON-based attributed graph. In this graph, spatial entities (rooms, balconies) are nodes, and topological relationships (adjacency, connectivity via doors) are edges. This representation preserves both geometry and topology, enabling efficient spatial reasoning.

2. Rule Engine: This component translates natural language building codes into executable, machine-readable rules.

   • It leverages Large Language Models (LLMs) to parse regulatory texts (e.g., from the High Life Study, BADS, SEPP 65) and decompose them into structured JSON objects.
   • Each rule is defined with specific attributes: rule identifier, applicability (e.g., apartment type), condition, threshold, comparison operator, and scoring logic (e.g., binary pass/fail or proportional scoring).
   • This approach aims to overcome the bottleneck of manual rule encoding and allows the system to adapt to evolving policies by updating the rule set rather than redesigning the system.

3. Compliance Check Engine: This engine evaluates the structured building model against the encoded rules.

   • It compares extracted geometric and topological data (e.g., room area, window access, circulation paths) against the thresholds defined in the Rule Engine.
   • The engine incorporates LLMs to interpret measurements and spatial relationships contextually, handling rules that depend on geometric understanding rather than just numeric thresholds.
   • It generates structured JSON reports detailing compliance status, specific violations, and remediation guidance, ensuring the decision-making process is transparent and traceable.

Key Contributions

1. Identification of Gaps: The paper systematically identifies the limitations of existing automated approaches, specifically their inability to handle multi-apartment floor plans with shared spaces and their lack of integration between visual parsing and regulatory reasoning.
2. Conceptual Framework: It proposes a novel, unified framework that bridges the gap between unstructured floor plan images and automated compliance assessment. The framework uniquely combines deep learning for geometric extraction with LLM-driven rule generation and graph-based topological analysis.
3. Scalable Architecture: The proposed modular design allows for the adaptation to different jurisdictions and evolving building codes by separating the data extraction, rule encoding, and evaluation processes.

Results and Experimental Findings
The paper presents initial experiments to assess the feasibility of the framework components rather than a fully deployed system tested on a massive dataset.

• Data Extraction: The framework successfully demonstrates the conversion of floor plan images into structured JSON models, capturing room boundaries, furniture, and orientation.
• Rule Encoding: Experiments show that LLMs can effectively translate textual design standards (e.g., minimum internal area requirements) into structured JSON rules with appropriate thresholds and scoring logic.
• Compliance Checking: A sample case study (APT_101) demonstrates the engine's ability to verify a 2-bedroom apartment's internal area (80.0 m²) against a regulatory threshold (70.0 m²), producing a clear "PASS" result with traceable rationale.
• VLM Limitations: The authors tested a Visual Language Model (VLM) on single-unit and multi-unit floor plans. While the VLM performed reasonably well on annotated single-unit plans, it struggled significantly with multi-unit plans lacking text labels, exhibiting hallucinations in room counts, dimensions, and apartment types. This highlights that current VLMs are insufficient as standalone tools for precise geometric compliance in complex, unannotated scenarios.

Significance and Claims
The paper claims that its proposed framework offers a scalable, consistent, and transparent approach to automated compliance checking, which is essential for supporting large-scale studies like the "High Life Study" and enforcing design standards across thousands of apartments. By automating the extraction and evaluation of design metrics, the system aims to reduce manual workload, improve reproducibility, and enhance the ability of policymakers to monitor design quality over time.

The authors maintain a modest stance, acknowledging that the framework is currently conceptual and that significant challenges remain, particularly regarding the diversity of floor plan quality, the detection of north arrows in varied styles, and the segmentation of individual units in multi-apartment layouts. They emphasize that a "Human-in-the-Loop" verification step is necessary for low-confidence outputs and that future work must focus on developing dedicated datasets for multi-unit plans and refining the integration of specialized AI components. The paper does not claim to have solved the problem entirely but rather provides a viable architectural pathway toward a fully automated solution.