MACHINE LEARNING METHODS FOR 3D RECONSTRUCTION OF URBAN ENVIRONMENTS

Abstract

This article examines machine learning methods used to analyze visual data (photos and videos) from surveillance cameras for the purpose of constructing 3D models of urban infrastructure. The study highlights the relevance of automating spatial analysis in response to the growing volume of visual information. The research applied modern machine learning algorithms, including logistic regression, decision trees, random forest, clustering methods, and deep neural networks (U-Net, ResNet, ViT). A quantitative comparison of all models was performed using MAE, IoU, and F1-score metrics, allowing for an objective evaluation of their applicability in 3D urban modeling. The scientific novelty lies in the integrated application of machine learning and computer vision techniques to create digital twins of urban environments. The practical significance is demonstrated by the potential use of the developed solutions in territorial planning and urban infrastructure monitoring systems