Clustering-Based Plane Segmentation Neural Network for Urban Scene Modeling.

Area of Science:

• Computer Vision
• 3D Scene Understanding
• Machine Learning

Background:

• Urban scene modeling is crucial for applications like 3D mapping and AR/VR.
• Accurate segmentation of man-made structures (roads, buildings) from 3D point clouds is challenging.

Purpose of the Study:

• To present a novel clustering-based plane segmentation neural network for urban 3D point clouds.
• To improve the accuracy of modeling man-made structures in urban environments.

Main Methods:

• Introduced hybrid K-means plane segmentation (HKPS), a neural network for segmenting unorganized 3D point clouds.
• Employed hybrid K-means clustering, considering both Euclidean and cosine distances for point grouping.
• The network estimates the optimal number of planes without requiring labeled training data.

Main Results:

• The HKPS method demonstrated superior performance in plane segmentation compared to conventional approaches.
• Evaluated using the Virtual KITTI dataset, confirming its effectiveness.
• Achieved accurate segmentation of planes, crucial for urban scene modeling.

Conclusions:

• The proposed HKPS method offers an effective, unsupervised approach for urban 3D point cloud segmentation.
• This technique advances 3D map generation, city digitization, and AR/VR applications.
• Publicly available code facilitates further research and development.