In recent years, the demand for electricity has grown dramatically, leading to a significant expansion of overhead transmission lines. The increasing complexity and long-distance nature of power grids make inspection tasks both difficult and labor-intensive. Traditional manual patrol methods suffer from low efficiency and high risk, failing to meet modern requirements. The advent of UAV drone LiDAR technology has provided a revolutionary solution, enabling rapid acquisition of large-scale, high-precision point cloud data of terrain and objects at relatively low cost. This technology is particularly attractive for power line extraction and reconstruction. However, existing methods still face challenges in automation, accuracy, and robustness under complex terrain conditions. In this study, we propose a novel approach based on the spatial distribution characteristics of power line points using UAV drone LiDAR data. Our method integrates improved morphological filtering, height difference threshold segmentation, dimensionality-based feature analysis, density clustering, and least-squares configuration for 3D reconstruction. Extensive experiments on real-world high-voltage AC transmission line data demonstrate that our method achieves an extraction accuracy exceeding 98%, significantly outperforming traditional approaches.
Methodology
Ground Point Filtering
Accurate separation of ground points from non-ground points is critical for power line extraction. Traditional height difference threshold segmentation often fails in complex terrains because it uses a fixed global threshold. To address this, we introduce an improved morphological filtering algorithm that incorporates slope constraints. The slope at each point is computed as:
$$
\text{slope} = \frac{z – z_{\min}}{\sqrt{(x – x_{\min})^2 + (y – y_{\min})^2}}
$$
where \((x_{\min}, y_{\min}, z_{\min})\) is the lowest point within a grid cell, and \((x, y, z)\) are coordinates of any other point in the same cell. Points with slope below a predefined threshold (set to 1° in our tests) are classified as ground points. Next, we compute the height difference between each non-ground point and the nearest ground reference. An iterative procedure determines the optimal height difference threshold: starting from an initial value, we iteratively remove points below the threshold, recompute the mean height difference, and update the threshold until convergence. In our experiments, the final optimal height difference threshold was found to be 24.8 m.
Power Line Point Cloud Extraction
After ground filtering, the remaining points include tower points, tall vegetation, and other above-ground objects. To isolate power line points, we leverage the spatial distribution anisotropy of laser points. For each point, we compute the eigenvalues \(\lambda_1, \lambda_2, \lambda_3\) of the covariance matrix of its local neighborhood. The dimensionality features are defined as:
$$
(a_{1D}, a_{2D}, a_{3D}) = \left( \frac{\lambda_1 – \lambda_2}{\lambda_1}, \frac{\lambda_2 – \lambda_3}{\lambda_1}, \frac{\lambda_3}{\lambda_1} \right)
$$
The Shannon entropy of the point is:
$$
E_f = -a_{1D} \ln(a_{1D}) – a_{2D} \ln(a_{2D}) – a_{3D} \ln(a_{3D})
$$
We search for the optimal neighborhood radius that minimizes \(E_f\), ensuring that the dimensional characteristics best capture linear (power line) structures. Points with a high linearity (i.e., \(\lambda_1 \gg \lambda_2 \approx \lambda_3\)) are retained as candidate power line points. This step effectively removes scattered vegetation and planar tower surfaces.
Single Power Line Separation via DBSCAN Clustering
The extracted power line candidate points still contain multiple lines interwoven. We apply the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to separate individual power lines. DBSCAN requires two parameters: the radius \(\epsilon\) (rEps) and the minimum number of points \(MinPts\) (pMinPts). Through empirical tuning on our dataset, we set \(MinPts = 15\) and \(\epsilon = 2\) (in the coordinate unit of the point cloud). The algorithm works as follows:
- Initialize all points as unvisited. Set cluster counter \(C = 0\).
- For each unvisited point \(p\), retrieve its \(\epsilon\)-neighborhood. If the number of points in the neighborhood is less than \(MinPts\), mark \(p\) as noise. Otherwise, create a new cluster and expand it by recursively adding density-reachable points.
- Continue until all points are processed.
The clustering result yields distinct groups, each corresponding to a single power line. In our test data, with the chosen parameters, DBSCAN produced 9 clusters, matching the actual number of power lines in the scene.
3D Reconstruction of Power Lines
Once single power line point sets are obtained, we perform 3D reconstruction using the least-squares configuration model. The power line catenary curve is modeled as:
$$
z = a (x^2 + y^2) + b \sqrt{x^2 + y^2} + c
$$
where \(a, b, c\) are fitting coefficients determined by minimizing the sum of squared residuals. This model accurately captures the sag and shape of overhead lines. After fitting, the 3D line can be visualized and used for downstream applications such as clearance analysis and inspection planning.
Experimental Validation
Dataset Description
We conducted experiments on real UAV drone LiDAR point cloud data acquired from a high-voltage AC transmission project. The dataset covers 6 towers and 28 power lines, with a total of 3,212,116 points. The original point cloud includes ground, vegetation, buildings, towers, and power lines. The complex environment provides a rigorous test for our algorithm.
Results and Analysis
Ground Filtering and Height Threshold Segmentation
After applying the improved morphological filtering with slope threshold 1° and iteratively determined height difference threshold 24.8 m, we obtained a set of non-ground points that mostly include power line points and tower points. Vegetation and building points were effectively removed, preventing over-segmentation of power lines. The intermediate result is shown qualitatively in the original paper (Figure 4).
Dimensionality-Based Refinement
Using the optimal neighborhood radius determined by minimizing Shannon entropy, we computed dimensionality features for each candidate point. Points with high linearity (1D probability > 0.8) were retained. This step reduced the number of candidate points by approximately 30% while preserving over 99% of true power line points.
DBSCAN Clustering for Single Line Extraction
Applying DBSCAN (MinPts=15, \(\epsilon\)=2) to the refined point set successfully separated 9 distinct power line clusters, corresponding to the 9 actual lines in the test area (note: some lines may be superimposed in the point cloud due to perspective). The extraction result is presented in the original paper (Figure 5). Visual inspection confirms that each cluster corresponds to a single conductor without mixing.
3D Reconstruction Accuracy
We fitted the catenary model to each cluster. The root mean square error (RMSE) between the fitted curve and the original points was less than 0.15 m for all lines, demonstrating high fidelity. The reconstructed 3D power lines are shown in Figure 6 of the original paper.
Quantitative Evaluation
To objectively evaluate extraction performance, we compared our method against the traditional approach (high-difference threshold without dimensionality filtering) using manually extracted power line points as ground truth. The results are summarized in the table below.
| Method | Total Original Points | Manual Extraction Points | Algorithm Extraction Points | Accuracy (%) |
|---|---|---|---|---|
| Traditional Method | 3,212,116 | 61,436 | 63,638 | 96.54 |
| Proposed Method | 3,212,116 | 61,436 | 62,075 | 98.97 |
As shown, our method achieved an accuracy of 98.97%, significantly higher than the traditional method’s 96.54%. The slightly higher number of extracted points in both methods compared to manual extraction is due to the inclusion of a few noisy points near power lines; however, after DBSCAN clustering, these noise points are effectively eliminated in the final single-line clusters.
Parameter Sensitivity Analysis
We also investigated the sensitivity of our method to key parameters. The slope threshold for ground filtering was varied from 0.5° to 3°. Within the range 0.8°–1.2°, the extraction accuracy remained above 98%. For the height difference threshold, values between 22 m and 26 m yielded similar performance. The DBSCAN parameters were also tested: \(MinPts\) from 10 to 20 and \(\epsilon\) from 1.5 to 2.5. The best combination was found to be (15, 2), consistent with our earlier selection. This robustness confirms the practical applicability of our method.
Conclusion
In this paper, we presented a comprehensive framework for power line extraction and 3D reconstruction from UAV drone LiDAR point cloud data. The key contributions include: (1) an improved morphological filtering method combined with an iterative height difference threshold determination for effective ground point removal; (2) a dimensionality-based refinement step that leverages Shannon entropy to optimally select neighborhood scale and identify linear structures; (3) the use of DBSCAN clustering to accurately separate individual power lines in cluttered scenes; and (4) a least-squares catenary model for faithful 3D reconstruction. Experimental results on real-world high-voltage transmission data demonstrate that our method achieves an extraction accuracy of over 98%, significantly outperforming traditional approaches. The high automation level and robust performance make it well-suited for practical power line inspection and management applications. Future work will focus on extending the method to handle crossed lines and integrating with UAV drone real-time processing.