APPLIED LASER, Volume. 45, Issue 3, 146(2025)
Point Cloud Classification Method for Medium Voltage Power Lines Based on Improved PointNet++
To address the critical challenges of noise interference, low classification accuracy, and limited robustness in medium-voltage power line point cloud classification, this study presents an enhanced deep learning framework based on improved PointNet++ architecture. Initially, multiple techniques are employed to extract multi-dimensional features, including spatial information, geometric features, and local geometric features, creating a 40-dimensional feature vector for each point in the point cloud. Subsequently, PointNet++ is enhanced by incorporating a point attention module (PAM) and a group attention module (GAM), along with layer normalization and residual connection structures, to improve its detail capture capabilities and mitigate the impact of complex environments on classification performance. Finally, the paper utilizes 10 kV medium-voltage power line corridor data, collected from a specific area, to construct a dataset for validating the method. Experimental validation demonstrates superior performance over conventional machine learning methods and baseline PointNet/PointNet++ models, achieving respective improvements of 1.6%, 5.3%, and 4.6% in precision, recall, and F1-score compared to PointNet++ (XYZ+Features). Visualization analyses confirm the critical role of attention mechanisms in enhancing structural feature clarity and environmental differentiation. The proposed method demonstrates greater accuracy in medium-voltage power line point cloud extraction, clearer structural features, and higher differentiation from the surrounding environment.
Get Citation
Copy Citation Text
Ke Jianyan. Point Cloud Classification Method for Medium Voltage Power Lines Based on Improved PointNet++[J]. APPLIED LASER, 2025, 45(3): 146
Category:
Received: Jul. 19, 2024
Accepted: Jun. 17, 2025
Published Online: Jun. 17, 2025
The Author Email: