Building Structure Recognition Using BIM-Rendered Re-Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for recognizing building structures in multi-story buildings face challenges such as varying data accuracy due to skill levels, high effort in data registration and annotation, difficulty in handling large amounts of point cloud data, and poor recognition of structures with high light reflectivity, especially metal pipes, without adequate automation and site-tailored models.
Innovation Solution
A system utilizing machine learning models generated through a process involving BIM data, correct images, and virtual observation images, with re-learning and image processing to enhance recognition accuracy, particularly for structures with high light reflectivity, by using a machine learning model generation device and recognition device that includes correct image generation, virtual observation image processing, and re-learning to improve model precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If LiDER or remote sensing technology is used to measure construction status, then measurement automation is achieved, but data accuracy varies depending on skill level and measurement conditions
Solution Approach 1:
The patent uses BIM (Building Information Modeling) data as a digital copy or representation of the actual building structure. Instead of directly measuring the physical building with LiDER, the system creates a virtual 3D model from BIM data and performs measurements and recognition on this digital copy, thereby eliminating the skill-level dependency of manual measurement while maintaining high accuracy through the structured nature of BIM data.
2Measurement precision
If multiple portions of construction site are measured to ensure data accuracy, then measurement coverage is improved, but time and effort for data registration increases
Solution Approach 1:
The patent performs preliminary processing by pre-processing BIM data into a format suitable for machine learning recognition. The BIM data is converted into 3D point cloud data and processed to create training datasets with annotations before actual recognition tasks. This preliminary preparation eliminates the need for time-consuming manual measurement and registration during construction site visits, as the system can directly process the pre-prepared digital models.
3Measurement precision
If manual identification of structures and measurement of positions and sizes is performed, then recognition accuracy is improved, but work effort and time increase
Solution Approach 1:
The patent replaces manual mechanical measurement and identification processes with an automated machine learning-based recognition system. The system uses trained neural networks to automatically identify structures, measure positions and sizes, and generate recognition results from 3D point cloud data. This substitution of manual labor with automated AI processing maintains high recognition accuracy while dramatically improving work efficiency and eliminating the time-consuming manual annotation process.
4Quantity of substance
If a large number of photographic images are collected for machine learning training, then model learning data sufficiency is improved, but data collection complexity and cost increase
Solution Approach 1:
Instead of collecting actual photographic images from construction sites, the patent uses BIM data as a digital copy to generate synthetic training data. The BIM models are processed to create 3D point cloud data and rendered images that simulate real construction site conditions. This approach provides unlimited training data quantity without the complexity of physically visiting and photographing actual construction sites, as the digital models can be generated and manipulated computationally.
5Quantity of substance
If structures with high light reflectivity such as metal pipes are photographed, then complete structural documentation is achieved, but recognition accuracy decreases due to blown-out highlights
Solution Approach 1:
The patent uses 3D point cloud data generated from BIM models as a digital copy of structures, eliminating the lighting problems inherent in photographing reflective surfaces. The point cloud representation captures geometric information without being affected by light reflectivity, allowing accurate recognition of metal pipes and other highly reflective structures without the blown-out highlights that plague traditional photogrammetry.
6Measurement precision
If site-tailored machine learning models are regenerated for each construction site, then recognition accuracy for specific structures is improved, but time and costs for model regeneration increase
Solution Approach 1:
The patent segments the machine learning model into a general-purpose base model trained on diverse construction site data and site-specific fine-tuning components. The base model handles common structures and conditions, while only the necessary parameters are fine-tuned for each specific construction site using local BIM data. This segmentation allows rapid adaptation to new sites without regenerating the entire model, maintaining high accuracy for site-specific structures while minimizing regeneration time and costs.
Data Source
AI summary
Provided is a building inside structure recognition system for recognizing a structure in a building by using a machine learning model. A building inside structure recognition system according to the present invention comprises: a machine learning model generation device that generates a first machine-learned model by executing machine learning in which a correct image generated from building information modeling (BIM) data is set as correct data and a virtual observation image generated by rendering the BIM data is set as observation data, and a second machine-learned model by inputting at least an image for re-learning into the first machine-learned model to execute re-learning; and a building inside structure recognition device that recognizes a structure in a building by using the second machine-learned model generated by the machine learning model generation device.