| Keywords [eng] |
Spatial-Temporal Change Detection, Point Cloud Data, 3D Scene Alignment, 3D Visualization, Machine learning, Erdvinių ir laiko pokyčių aptikimas, taškų debesies duomenys, 3D scenos lygiavimas, 3D vizualizacija, mašininis mokymasis. |
| Abstract [eng] |
The research focuses on spatial-temporal change detection in point cloud datasets, a growing area of importance as 3D sensing technologies such as LiDAR, photogrammetry, and drone-based mapping systems become more prevalent. These technologies produce vast quantities of high-resolution spatial data, enabling detailed 3D representations of real-world environments. Detecting and interpreting how these environments evolve over time is difficult and is complex challenge. Main aim of this paper was to develop a comprehensive, automated framework for identifying, quantifying, and visualizing changes across time in dynamic settings like urban infrastructure, construction zones, and natural landscapes. The proposed framework addresses not only geometric comparison of 3D scenes, but also semantic understanding by identifying what changed rather than only where the change occurred. The work integrates 3D alignment techniques, deep learning architectures, and simulation-based data generation to improve scalability, interpretability, and automation in 3D change detection tasks. Existing methods have shown the effectiveness of voxel-based, octree, and pointwise distance techniques in computing structural change. However, they often fall short in dealing with noisy, unstructured data or providing contextual insight. To address these limitations, both synthetic and real-world datasets were used throughout the experiments. Synthetic datasets generated in controlled simulation environments enabled evaluation under known ground-truth changes, while real-world datasets allowed robustness testing under realistic conditions. Several methods were evaluated, including geometric baseline approaches, Feature-based MLP, PointNet, and an attention-based transformer model. Experimental results demonstrated that learned models outperform traditional geometric approaches, particularly in more challenging semantic change detection tasks. The proposed attention-based transformer model achieved the most balanced performance across classes and showed improved robustness to semantic-preserving geometric transformations. The developed framework provides an interpretable and scalable approach for analyzing spatial-temporal changes in point cloud data and demonstrates the potential of transformer-based architectures for semantic 3D scene understanding. |
