A Deep-Learning-Based Method for Extracting an Arbitrary Number of Individual Power Lines from UAV-Mounted Laser Scanning Point Clouds

Sha Zhu, Qiang Li, Jianwei Zhao, Chunguang Zhang, Guang Zhao, Lu Li, Zhenghua Chen, Yiping ChenRemote Sensing 2024 · DOI · OpenReview

TL;DR

This paper introduces EM-Net, a novel deep learning architecture for automatically extracting individual power lines from UAV-mounted LiDAR point clouds. It addresses challenges like massive point clouds and imbalanced data by employing a multi-branch network with specialized loss functions for feature learning and classification, enabling robust extraction of an arbitrary number of power lines.

What Problem Does This Paper Address?

Traditional power line inspections often rely on manual methods or manned helicopters, which are risky, labor-intensive, and challenging in difficult terrains. While remote sensing, particularly UAV-mounted LiDAR, offers a safer and more efficient alternative, extracting individual power lines from the resulting 3D point clouds presents significant hurdles. These include the sheer volume of data, the imbalance between power line points and other environmental features, and the difficulty in distinguishing between multiple, often discontinuous, power lines. Existing methods often require manual parameter tuning, limiting their efficiency and generalizability across diverse scenarios.

Methods

The core of this work is the EM-Net, a multi-branch deep learning network designed for robust individual power line extraction. It comprises three main components:

  • Backbone Network: The paper leverages RandLA-Net as the backbone. RandLA-Net is chosen for its efficiency in handling large-scale point clouds by using techniques like random sampling and local feature aggregation (LFA) to progressively downsample and extract features.
  • Power Line and Power Tower Extraction Branch: This branch classifies each point in the cloud into one of three categories: power line, power tower, or "other." To combat the significant class imbalance (where power line/tower points are a minority), it employs a weighted cross-entropy loss function. This loss assigns different weights to categories, giving more importance to the minority classes, thus improving their classification accuracy.
  • Individual Power Line Feature Learning Branch: This is the most novel aspect. It aims to learn discriminative embeddings for points belonging to different power lines. The goal is to ensure that points from the same power line are clustered closely in the embedding space, while points from different power lines are pushed far apart. This is achieved using a discriminative loss function (LDisc), which comprises variance, distance, and regularization terms. This learned embedding is then fed into an unsupervised Mean-Shift clustering algorithm to separate individual power lines without needing to pre-specify their number.

Datasets

The authors evaluated their method on two distinct datasets collected by a RIEGL HawkScan X3 UAV. These datasets were acquired from different locations, providing a good testbed for the method's generalizability. The point clouds had a density of approximately 200 points per square meter. For evaluation, the data was manually annotated into three categories (power line, power tower, others) and further refined to label individual power lines. The datasets contained between four and ten power lines, and importantly, included instances of discontinuous power lines.

Key Insights and Ideas

  • The Power of Discriminative Embeddings: The key insight is that learning a feature space where individual power lines are inherently separable, even without explicit supervision for each line, is crucial. This allows for subsequent unsupervised clustering to handle an arbitrary number of lines.
  • Addressing Class Imbalance is Critical: The paper highlights how imbalanced data in point clouds (few power line points vs. many background points) can severely degrade performance. The weighted cross-entropy loss is a practical and effective solution for this common problem in point cloud analysis.
  • Unsupervised Clustering for Arbitrary Numbers: Relying on Mean-Shift clustering after learning discriminative features elegantly solves the problem of not knowing the number of power lines beforehand, a significant advantage over methods requiring a fixed number of clusters.
  • UAV LiDAR as a Viable Data Source: The work reinforces the potential of UAV-mounted LiDAR for detailed infrastructure inspection, offering precise 3D information independent of lighting conditions.

Key Results

The proposed EM-Net demonstrated strong performance:

  • Individual Power Line Extraction: Achieved an average precision of 0.986, recall of 0.975, and F1-score of 0.981 across five-fold cross-validation on the two datasets.
  • Comparative Performance: Outperformed established methods like DBSCAN, RANSAC, and RECONSTRUCT in precision, recall, and F1-score, particularly in challenging scenarios like occluded or interrupted power lines and lines crossing power towers.
  • Weighted Cross-Entropy Impact: The weighted cross-entropy loss showed a significant improvement, particularly in recall for power towers (nearly 7% boost) and power lines (1% boost), leading to an overall F1-score of 95.97% for towers and exceeding 99% for power lines.

Limitations and Future Work

  • Bandwidth Sensitivity: While the method performs well, the paper notes that the performance is somewhat sensitive to the bandwidth parameter used in the Mean-Shift clustering. The optimal bandwidth was found to be around 1.0, which was linked to the dv threshold in the discriminative loss.
  • Computational Cost: Although RandLA-Net is efficient for large-scale point clouds, deep learning methods can still be computationally intensive.
  • Future Work: The authors suggest further exploration into optimizing the bandwidth selection and potentially extending the method to handle more complex scenarios or different types of infrastructure. The paper also implicitly points towards the need for more diverse and challenging benchmark datasets for power line extraction.

Summary updated

Abstract

In recent years, laser scanners integrated with Unmanned Aerial Vehicles (UAVs) have exhibited great potential in conducting power line inspections in harsh environments. The point clouds collected for power line inspections have numerous advantages over remote image data. However, point cloud-based individual power line extraction, which is a crucial technology required for power line inspections, still poses several challenges such as massive 3D points, imbalanced category points, etc. Moreover, in various power line scenarios, previous studies often require manual setup and careful adjustment of different thresholds to separate different power lines, which is inefficient for practical applications. To handle these challenges, in this paper, we propose a multi-branch network to automatically extract an arbitrary number of individual power lines from point clouds collected by UAV-based laser scanners. Specifically, to handle the massive 3D point clouds in complex outdoor scenarios, we propose to leverage deep neural network for efficient and rapid feature extraction in large-scale point clouds. To mitigate imbalanced data quantities across different categories, we propose to design a weighted cross-entropy loss function to measure the varying importance of each category. To achieve the effective extraction of an arbitrary number of power lines, we propose leveraging a loss function to learn the discriminative features that can differentiate the points belonging to different power lines. Once the discriminative features are learned, the Mean Shift method can distinguish the individual power lines by clustering without supervision. The evaluations are executed on two datasets, which are acquired at different locations with UAV-mounted laser scanners. The proposed method has been thoroughly tested and evaluated, and the results and discussions confirm its outstanding ability to extract an arbitrary number of individual power lines in point clouds.
BibTeX
@article{zhu2024deep,
  author = {Sha Zhu and Qiang Li and Jianwei Zhao and Chunguang Zhang and Guang Zhao and Lu Li and Zhenghua Chen and Yiping Chen},
  title = {{A Deep-Learning-Based Method for Extracting an Arbitrary Number of Individual Power Lines from UAV-Mounted Laser Scanning Point Clouds}},
  journal = {Remote Sensing},
  year = {2024},
  doi = {10.3390/rs16020393},
}
Figure