Extracting Structural Elements From 3D Point Clouds in Indoor Environments via Machine Learning Techniques

The utilization of three-dimensional point clouds is an advanced approach for detecting the geometry of objects within a building environment. Nonetheless, a vast amount of data still needs to be manually processed. Intelligent automation frameworks could be deployed to overcome such issues. Hence, this study proposes a machine learning-based framework for successfully classifying structural components in indoor environments. The proposed framework consists of four stages: pre-processing, feature extraction, feature selection, and interpretability of classification results using an explainable machine learning method. According to the proposed framework, the chi-squared test stands out for optimum local neighborhood radius determination and feature selection. The CatBoost model has the highest accuracy of 82.96%, whereas the Random Forest model's accuracy is 82.09%. However, the training time for the Random Forest is 27 times shorter than the CatBoost. Hence, both models could be preferred to other machine learning models for practical applications due to the good balance between accuracy and calculation efficiency. Additionally, the model with the highest accuracy, CatBoost, is evaluated using the Shapley Additive exPlanations to understand the impacts of features on predictions, and according to the results, Z coordinate and verticality had a relatively high impact on the model, while others had low impacts. The proposed framework uses machine learning to classify indoor point clouds, balancing processing time and accuracy for computational efficiency in practical applications. Hence, the framework could be utilized to automate the digitalization efforts of indoor environments effectively.