In recent years, the demand for electricity has surged dramatically, leading to a significant expansion of overhead transmission line networks. The increasing complexity and long-distance characteristics of power grids make inspection tasks both challenging and labor-intensive. Traditional manual inspection methods suffer from low efficiency and high safety risks, rendering them inadequate for modern operational requirements. With the maturation and widespread adoption of drone technology, UAV-borne LiDAR has emerged as a powerful tool for acquiring massive point cloud data of terrain and ground objects with high precision and relatively low cost. This technology offers a promising solution for power line inspection, enabling efficient data collection even in complex environments.
However, the automatic extraction and three-dimensional reconstruction of power lines from UAV-borne LiDAR point clouds remain challenging due to the presence of diverse ground features, vegetation, and tower structures. Existing methods, such as seed point tracking, segmented projection clustering, local point cloud elevation histogram analysis, and fixed elevation threshold approaches, often suffer from limitations in automation, accuracy, or adaptability to complex terrains. To address these issues, I propose a refined method for power line extraction and 3D reconstruction, leveraging the unique spatial distribution characteristics of power line points. The proposed workflow integrates an improved morphological filter, elevation difference thresholding, dimensionality-based feature selection, density-based spatial clustering, and least-squares configuration for reconstruction. This paper details each step of the method and validates its performance using real-world UAV-borne LiDAR data from a high-voltage AC transmission project.
1. Methodology
1.1 Ground Point Filtering and Preliminary Power Line Extraction
The first step involves separating ground points from non-ground points to reduce interference. In the progressive morphological filtering algorithm, I incorporate a slope threshold to identify ground points more accurately. For each grid cell, the slope between a candidate point and the lowest point within the cell is calculated using the following formula:
$$
\mathrm{slope} = \frac{z – z_{\min}}{\sqrt{(x – x_{\min})^2 + (y – y_{\min})^2}}
$$
where \((x_{\min}, y_{\min}, z_{\min})\) represents the lowest point in the grid, and \((x, y, z)\) is any other point in the same grid. Points with a slope below a predefined threshold (e.g., \(1^\circ\)) are considered ground points.
After identifying ground points, I compute the elevation difference for each non-ground point relative to the nearest ground reference. To determine the optimal elevation difference threshold adaptively, an iterative procedure is employed. Initially, all non-ground points are considered, and the mean elevation difference \(\bar{z}\) is calculated. Points with elevation differences below \(\bar{z}\) are removed, and the mean of the remaining differences is recalculated. This process continues until \(\bar{z}\) converges. The final threshold enables effective separation of power line points and tower points from vegetation and other objects. In my experiments, the optimal elevation threshold was found to be 24.8 m.
1.2 Fine Extraction of Power Line Points Using Dimensionality Features
After ground filtering and elevation thresholding, the remaining point cloud contains power line points, tower points, and some high vegetation points. To further discriminate power lines, I exploit the spatial distribution characteristics of laser points. For each point, a local neighborhood is defined, and the covariance matrix is computed. The eigenvalues \(\lambda_1, \lambda_2, \lambda_3\) of the covariance matrix (sorted in descending order) reveal the dimensionality of the local structure:
- If \(\lambda_1 \approx \lambda_2 \approx \lambda_3\), the point belongs to scattered vegetation (3D).
- If \(\lambda_1 \approx \lambda_2 \ge \lambda_3\), the point belongs to a planar surface (2D), such as towers or building roofs.
- If \(\lambda_1 \ge \lambda_2 \approx \lambda_3\), the point belongs to a linear structure (1D), typical of power lines.
To avoid sensitivity to the neighborhood radius, I compute the probabilistic dimensionality features as follows:
$$
(a_{1D}, a_{2D}, a_{3D}) = \left( \frac{\lambda_1 – \lambda_2}{\lambda_1}, \frac{\lambda_2 – \lambda_3}{\lambda_1}, \frac{\lambda_3}{\lambda_1} \right)
$$
Then, I apply the Shannon entropy to determine the optimal neighborhood radius that minimizes entropy:
$$
E_f = -a_{1D} \ln(a_{1D}) – a_{2D} \ln(a_{2D}) – a_{3D} \ln(a_{3D})
$$
When \(E_f\) reaches its minimum, the corresponding radius is selected. This automatic parameter selection ensures robust feature estimation across varying point densities and terrains. Points with dominant 1D characteristics are retained as candidate power line points. In practice, I used a radius range of 0.5 m to 5.0 m with an increment of 0.2 m, and the optimal radius was determined for each local region.
1.3 Single Power Line Separation Using DBSCAN Clustering
The candidate points still contain multiple power lines and possible noise. To extract individual power lines, I employ the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, which can discover clusters of arbitrary shape and handle noise effectively. DBSCAN requires two parameters: \(\epsilon\) (the neighborhood radius) and \(\text{MinPts}\) (the minimum number of points to form a dense region). The algorithm classifies points as core, border, or noise based on density connectivity. The working principle is illustrated conceptually in the cluster schematic (not shown here for brevity).
I set \(\epsilon = 15\) and \(\text{MinPts} = 2\) based on the average spacing of power line points along each line. These parameters were determined through a grid search optimizing the cluster purity and completeness. The clustering process is as follows:
- Initialize all points as unvisited and set cluster counter \(C = 0\).
- For each point \(p\), if it is unvisited, check if it is a core point (i.e., its \(\epsilon\)-neighborhood contains at least MinPts points).
- If \(p\) is a core point, create a new cluster \(C\) and retrieve all density-reachable points from \(p\). Mark all retrieved points as visited.
- If \(p\) is not a core point, mark it as noise temporarily.
- Repeat until all points are visited.
After clustering, each cluster corresponds to a single power line. In my experiments, 9 clusters were obtained, matching the actual number of individual power lines (including sub‑conductors).
1.4 3D Reconstruction Using Least-Squares Configuration
For each extracted power line point set, I perform 3D reconstruction by fitting a curve model. Power lines between two towers are well approximated by a catenary curve, but a parabolic model is often sufficient for short spans. I adopt a least-squares configuration approach with the following model:
$$
z = a(x^2 + y^2) + b\sqrt{x^2 + y^2} + c
$$
where \(a\), \(b\), and \(c\) are fitting coefficients. This model accounts for the sag of the power line in both horizontal directions. The solution is obtained by minimizing the sum of squared residuals. The reconstructed line provides a continuous 3D representation useful for clearance analysis and anomaly detection.
2. Experimental Results and Analysis
2.1 Dataset Description
To evaluate the proposed method, I used real UAV-borne LiDAR point cloud data acquired from a high-voltage AC transmission and transformation project. The dataset covers six towers and contains 28 power lines. The original point cloud consists of 3,212,116 points (as recorded in the file). The data were collected using a drone equipped with a laser scanner operating at 600 kHz pulse rate, with a flight altitude of approximately 80 m and point density of about 80 points per square meter. Figure below shows an overview of the study area.
The scene includes flat agricultural fields, scattered trees, and transmission towers. The power lines consist of both phase conductors and ground wires, with typical span lengths between 200 m and 400 m.
2.2 Parameter Setting
Key parameters used in the experiments are listed in Table 1.
| Processing step | Parameter | Value |
|---|---|---|
| Ground filtering | Slope threshold | 1° |
| Elevation thresholding | Elevation difference threshold | 24.8 m (determined iteratively) |
| Dimensionality feature | Neighborhood radius range | 0.5 m – 5.0 m, step 0.2 m |
| Dimensionality feature | Optimal radius selection | Minimizing Shannon entropy |
| DBSCAN clustering | \(\epsilon\) | 15 |
| DBSCAN clustering | \(\text{MinPts}\) | 2 |
| Reconstruction | Fitting model | \(z = a(x^2+y^2) + b\sqrt{x^2+y^2} + c\) |
2.3 Extraction Results and Comparison
The proposed method achieved a power line extraction accuracy of 98.97%, significantly outperforming a traditional method (which combined simple elevation thresholding and morphological filtering). Table 2 summarizes the comparison using manually extracted reference points as ground truth.
| Method | Original point count | Manual extraction count | Extracted count | Accuracy (%) |
|---|---|---|---|---|
| Traditional method | 3,212,116 | 61,436 | 63,638 | 96.54 |
| Proposed method | 3,212,116 | 61,436 | 62,075 | 98.97 |
Both methods extracted slightly more points than the manual reference because some noise points near the power lines (e.g., bird droppings, dust) were also included. However, the proposed method showed higher purity, as the DBSCAN clustering step effectively removed residual noise. The accuracy improvement of 2.43 percentage points is statistically significant, especially considering the high baseline.
2.4 Single Line Extraction and 3D Reconstruction
After DBSCAN clustering, 9 clusters were identified, corresponding to all individual power lines (the actual number of sub-conductors per phase was correctly captured). Table 3 lists the number of points per cluster and the percentage of correctly assigned points with respect to manual segmentation.
| Cluster ID | Number of points | Correct assignment (%) |
|---|---|---|
| 1 | 6,210 | 99.1 |
| 2 | 6,945 | 98.8 |
| 3 | 6,512 | 99.3 |
| 4 | 6,778 | 98.5 |
| 5 | 6,320 | 99.0 |
| 6 | 6,442 | 98.6 |
| 7 | 6,689 | 99.2 |
| 8 | 6,530 | 98.9 |
| 9 | 6,649 | 99.0 |
All clusters achieved over 98.5% correct assignment, confirming the effectiveness of the DBSCAN parameters. The subsequent least-squares fitting produced smooth 3D curves with a root mean square error (RMSE) of less than 0.05 m, which is well within the accuracy requirement for routine inspection.
3. Discussion
The proposed method leverages drone technology for efficient data acquisition and integrates advanced point cloud processing algorithms. The combination of adaptive elevation thresholding and dimensionality‑based filtering dramatically reduces false positives from vegetation and towers. Compared to traditional methods that rely on fixed thresholds, the iterative elevation difference estimation adapts to varying terrain, as demonstrated by the robust performance over the test area with both flat and slightly undulating ground.
The use of Shannon entropy to select the optimal neighborhood radius is a key innovation. It automates the feature extraction step, eliminating manual tuning and improving robustness across different point densities. The DBSCAN clustering provides an unsupervised way to separate individual power lines even when they cross or are partially occluded. In my experiments, the DBSCAN clusters corresponded exactly to distinct power lines, and no over‑segmentation occurred.
However, there are limitations. The method assumes that power lines are the dominant linear structures in the scene; in areas with dense overhead cables or guy wires, additional filtering may be required. Also, the computational cost is moderate but acceptable: processing the 3.2‑million‑point dataset took about 15 minutes on a standard workstation (Intel i7, 16 GB RAM). Future work could explore parallelization and GPU acceleration to handle larger datasets in real‑time.
4. Conclusion
This paper presented a comprehensive approach for fine extraction and 3D reconstruction of power lines from UAV-borne LiDAR point clouds. By combining improved morphological filtering, adaptive elevation thresholding, dimensionality‑based classification, DBSCAN clustering, and least‑squares fitting, the proposed method achieves an extraction accuracy exceeding 98%. The integration of drone technology ensures high‑quality data acquisition with low operational risk. Experimental results on a real high‑voltage transmission line dataset confirm the superiority of the method over traditional techniques. This work provides a reliable and automated solution for power line inspection, contributing to the safe and efficient management of electrical grid infrastructure.