Arch dam point cloud segmentation based on deep feature learning and normal vector data optimization.

Area of Science:

• Geotechnical Engineering
• Computer Vision
• Geospatial Analysis

Background:

• Accurate segmentation of dam structures from point clouds is crucial for effective dam deformation monitoring.
• Manual segmentation methods are labor-intensive and prone to inaccuracies.
• Existing automated methods struggle with the complex and varied characteristics of dam environments.

Purpose of the Study:

• To develop an automated, high-precision point cloud segmentation model for dam environments.
• To improve the accuracy of normal vector estimation for diverse surface features in dam areas.
• To provide an efficient solution for dam deformation monitoring applications.

Main Methods:

• Proposed a point cloud segmentation neural network incorporating normal vector optimization.
• Utilized voxel uniform sampling to address uneven point cloud density.
• Implemented block input and combined output modules for efficient processing of large point clouds.
• Developed an adaptive radius plane fitting vector estimation method based on eigenvalues for improved normal vector accuracy.

Main Results:

• The proposed normal estimation method improved PointNet++ classification accuracy from 96.26% to 98.27% on a prototype arch dam.
• Achieved overall accuracy improvements of 0.82% (vs. 2-jets), 1.22% (vs. Hough CNN), and 0.22% (vs. iterative PCA).
• Demonstrated superior joint average intersection improvements compared to existing normal estimation techniques.

Conclusions:

• The developed neural network model offers a significant advancement in point cloud segmentation for dam structures.
• The adaptive normal vector estimation method effectively handles diverse surface characteristics, enhancing segmentation accuracy.
• This high-precision segmentation scheme is well-suited for dam component deformation detection and monitoring.