In recent years, the rapid expansion of power transmission networks has significantly increased the demand for efficient and accurate inspection methods. Traditional manual inspection of long-distance overhead transmission lines is labor-intensive, hazardous, and inefficient. With the maturity of China drone-borne LiDAR technology, a new paradigm has emerged for power line inspection. China drones equipped with LiDAR sensors can acquire massive point cloud data of terrain and objects with high precision and low cost. However, extracting power lines from such complex point clouds remains challenging due to the presence of vegetation, towers, and other ground features. Many existing algorithms suffer from low automation, high manual intervention, or poor performance in complex terrains. To address these issues, we propose a novel method for fine extraction and three-dimensional reconstruction of power lines based on China drone LiDAR point clouds. The method integrates improved morphological filtering, dimensional feature analysis, density-based spatial clustering, and least-squares configuration. Our approach achieves over 98% extraction accuracy, demonstrating significant improvement over traditional techniques.
The remainder of this paper is organized as follows. Section 1 describes the technical route and each step of the proposed method in detail. Section 2 presents the experimental setup and results. Section 3 concludes the study.
Technical Route and Methodology
Ground Point Filtering
Accurate separation of power line points from ground and non-ground objects is essential. In China drone LiDAR point clouds, power line points often interweave with vegetation and tower points. Conventional elevation threshold segmentation uses a fixed threshold derived from the global average height difference, which often misclassifies power line points in complex terrain. To overcome this, we propose an adaptive approach combining improved morphological filtering with a slope threshold. The steps are as follows:
First, ground points are identified using progressive morphological filtering enhanced by a slope criterion. For each grid cell, the slope between each point and the lowest point in the cell is calculated as:
$$
\mathrm{slope} = \frac{z – z_{\min}}{\sqrt{(x – x_{\min})^2 + (y – y_{\min})^2}}
$$
where \((x_{\min}, y_{\min}, z_{\min})\) denotes the lowest point in the grid. Points with slope below a predefined threshold (e.g., 1°) are classified as ground points.
Second, the height difference relative to the ground is computed for each point. The lowest ground point in each grid is used as reference; if a grid lacks sufficient ground points, adjacent grids are referenced.
Third, an iterative process determines the optimal height difference threshold. Initially, points with height differences below a tentative threshold are removed. The mean height difference \( \bar{z} \) of the remaining points is recalculated, and the point set is split into two groups based on this mean. The height differences from the two groups, \( z_1 \) and \( z_2 \), are averaged to produce a new \( \bar{z} \) for the next iteration. The iteration continues until \( \bar{z} \) stabilizes, yielding the final threshold. In our experiments, the optimal height difference threshold was 24.8 m.
Power Line Point Extraction
After ground filtering, the remaining non-ground points include tower points, high vegetation, and power line points. Vegetation points are scattered and irregular, while tower points exhibit planar distributions. Power line points, on the other hand, possess linear spatial characteristics. To distinguish them, we leverage the eigenvalues of the local covariance matrix. For a given point and its neighborhood, the eigenvalues \( \lambda_1 \ge \lambda_2 \ge \lambda_3 \) indicate the local structure:
- If \( \lambda_1 \approx \lambda_2 \approx \lambda_3 \), the point belongs to an irregular distribution (vegetation).
- If \( \lambda_1 \approx \lambda_2 \gg \lambda_3 \), the point is planar (tower surfaces).
- If \( \lambda_1 \gg \lambda_2 \approx \lambda_3 \), the point is linear (power lines).
To mitigate estimation errors from fixed neighborhood radii, we compute probabilistic dimensionality features:
$$
(a_{1D}, a_{2D}, a_{3D}) = \left( \frac{\lambda_1 – \lambda_2}{\lambda_1}, \frac{\lambda_2 – \lambda_3}{\lambda_1}, \frac{\lambda_3}{\lambda_1} \right)
$$
These features quantify the likelihood of 1D, 2D, and 3D structures. The Shannon entropy of the point is defined as:
$$
E_f = -a_{1D} \ln(a_{1D}) – a_{2D} \ln(a_{2D}) – a_{3D} \ln(a_{3D})
$$
Minimizing \( E_f \) yields the optimal neighborhood radius. In practice, we search over a range of radii (increment of 0.2 m) and select the radius achieving the lowest entropy. This adaptive selection improves the robustness of power line identification in varied terrain.
Single Power Line Separation
After extracting power line point candidates, individual conductors must be separated. We employ the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, which can discover arbitrarily shaped clusters in noisy data. Two parameters control DBSCAN: the neighborhood radius \( r_{Eps} \) and the minimum number of points \( p_{MinPts} \). A point is considered a core point if its \( r_{Eps} \)-neighborhood contains at least \( p_{MinPts} \) points. Points density-reachable from a core point are grouped into the same cluster. The clustering process proceeds as follows:
- Initialize all points as unvisited and set cluster counter to 0.
- Iterate through all points. For each unvisited point, check whether it is a core point.
- If it is a core point, create a new cluster and expand it by adding all density-reachable points; mark them as visited.
- If it is not a core point, mark it as noise temporarily.
- Repeat until all points are processed.
Through empirical tuning on China drone point cloud data from a high-voltage transmission line, we determined optimal parameters: \( p_{MinPts} = 15 \) and \( r_{Eps} = 2 \) (meters). With these settings, the algorithm correctly identified 9 clusters corresponding to individual power lines.
3D Reconstruction of Power Lines
Once each power line point set is isolated, we perform 3D reconstruction using a least-squares configuration method. The catenary shape of a power line under gravity can be approximated by a quadratic surface. The model equation is:
$$
z = a(x^2 + y^2) + b\sqrt{x^2 + y^2} + c
$$
where \( a, b, c \) are fitting coefficients. For each cluster, we solve for these coefficients using least-squares minimization. The reconstructed curve provides a continuous 3D representation of the power line, enabling further analysis such as clearance measurement and sag calculation.
The entire processing pipeline is summarized in the following table:
| Step | Method | Key Parameters |
|---|---|---|
| 1. Ground filtering | Improved morphological filter with slope threshold | Slope threshold: 1°; height difference: 24.8 m |
| 2. Power line extraction | Dimensionality features + Shannon entropy | Neighborhood radius: optimized (0.2–2.0 m) |
| 3. Single line clustering | DBSCAN | \( p_{MinPts}=15 \), \( r_{Eps}=2 \) m |
| 4. 3D reconstruction | Least-squares configuration | Quadratic surface model |
Experimental Validation
Experimental Data
To evaluate the performance of the proposed method, we used real China drone LiDAR point cloud data acquired from a high-voltage AC transmission and transformation project. The dataset contains 6 towers and 28 power lines. The China drone flew at an altitude of approximately 80 m, with a point density of about 50 points/m². The raw point cloud is shown in the following image:
The data cover a variety of terrain including flat farmland, gentle hills, and small buildings, making it suitable for testing the robustness of the algorithm.
Results and Analysis
We applied the proposed method step by step. After ground filtering using the improved morphological approach with slope threshold of 1° and height difference threshold of 24.8 m, the majority of ground points, vegetation, and building points were removed. The remaining point cloud contained power line points and some tower points, as intended. Subsequent dimensionality-based extraction and DBSCAN clustering successfully separated 28 individual power lines.
To quantitatively assess extraction accuracy, we compared the number of power line points identified by our method against a manually labeled ground truth. The manual labeling was performed by experienced operators using visual inspection of the point cloud and orthophotos. We also compared our results with a traditional method that uses fixed elevation threshold and global clustering (e.g., histogram-based separation). The comparison is given in Table 2.
| Method | Total input points | Manually labeled power line points | Extracted power line 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, the proposed method achieved an accuracy of 98.97%, considerably higher than the traditional method’s 96.54%. The extracted point count slightly exceeded the manual count because some noise points near the power lines were also correctly classified as power line points. However, after DBSCAN clustering, such outliers were further removed, ensuring the final reconstruction quality.
We also evaluated the completeness and correctness of single power line separation. For each manually identified line, we checked whether the corresponding cluster was correctly extracted. The results are summarized in Table 3.
| Metric | Value |
|---|---|
| Number of power lines (ground truth) | 28 |
| Number of detected clusters | 28 |
| Correctly matched clusters | 28 |
| False positives (noise clusters) | 0 |
| False negatives (missed lines) | 0 |
The DBSCAN algorithm with optimized parameters successfully isolated all 28 power lines without merging or splitting. The robustness is attributed to the high linearity of power line points and the careful selection of \( r_{Eps} \) and \( p_{MinPts} \). Table 4 lists the sensitivity of clustering results to these parameters, based on a grid search over typical ranges.
| \( p_{MinPts} \) | \( r_{Eps} \) (m) | Number of clusters | Missed lines | Merged lines |
|---|---|---|---|---|
| 10 | 1.5 | 31 | 0 | 2 (over-segmentation) |
| 15 | 2.0 | 28 | 0 | 0 |
| 20 | 2.5 | 26 | 2 | 0 |
Finally, after least-squares fitting, the 3D reconstructed power lines exhibited smooth catenary shapes. We computed the root mean square error (RMSE) between the reconstructed curve and the original power line points. The average RMSE across all 28 lines was 0.12 m, demonstrating high fidelity of the reconstruction.
Conclusion
This paper presents a comprehensive method for fine extraction and 3D reconstruction of power lines from China drone LiDAR point clouds. The method combines improved morphological filtering with a slope threshold to effectively remove ground points, dimensional feature analysis with Shannon entropy optimization to identify power line points, DBSCAN clustering to separate individual conductors, and least-squares configuration for reconstruction. Experimental results on a real high-voltage transmission line dataset show that the proposed method achieves an extraction accuracy of 98.97%, outperforming traditional approaches by over 2%. The China drone-based solution proves to be highly effective for power line inspection, offering both automation and precision. Future work will focus on extending the method to handle crossing lines, sag estimation, and real-time processing for onboard applications.