💻 computer science

Autonomous Inspection of Power Line Insulators with UAV on an Unmapped Transmission Tower

This paper presents an autonomous online inspection algorithm that utilizes camera-LiDAR sensor fusion and deep learning to detect and localize power line insulators on unmapped transmission towers, demonstrating significant time savings in simulation and high-precision localization accuracy in real-world flights.

Original authors: Václav Riss, Vít Krátký, Robert Pěnička, Martin Saska

Published 2026-03-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Václav Riss, Vít Krátký, Robert Pěnička, Martin Saska

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a high-voltage power line tower standing tall in a field. It's covered in ceramic "insulators" (the white, ribbed discs that hold the wires) which are the unsung heroes keeping electricity from shorting out. Over time, these insulators get dirty, cracked, or broken. If they fail, the whole grid could go down, causing blackouts and expensive repairs.

Traditionally, checking these insulators is a tough job. You either have to climb a ladder (dangerous and slow) or fly a drone manually (requires a skilled pilot and lots of attention).

This paper introduces a "Smart Drone" that can inspect a tower all by itself, even if it has never seen that specific tower before. Think of it as a drone that doesn't need a map or a blueprint; it just flies up, looks around, figures out where the insulators are, and takes perfect photos of them in one go.

Here is how it works, broken down with some everyday analogies:

1. The "Eyes and Ears" (Sensor Fusion)

The drone is equipped with two main tools:

A Camera (The Eyes): It takes pictures to find the insulators. It uses a special AI brain (called YOLOv11n) trained to recognize what an insulator looks like, even if the background is messy with trees or clouds.
A LiDAR (The Ears/Touch): This is a laser scanner that shoots out thousands of invisible laser beams to measure distance. It creates a 3D "cloud" of points representing the tower.

The Magic Trick: The drone combines these two. It uses the camera to say, "Hey, I see an insulator right there!" and then immediately uses the laser data to say, "Okay, exactly how far away is it, and what is its 3D shape?" It's like looking at a shadow (the image) and then reaching out to touch the object to know its true size and location.

2. The "Detective Work" (Localization)

Once the drone spots an insulator, it needs to know exactly where it is to take a good photo. The raw laser data is messy—it sees the insulator, but also the metal tower, the wires, and maybe a bird.

To clean this up, the researchers used three different "detective" algorithms:

DBSCAN: Imagine a group of friends holding hands. This algorithm groups the laser points that are close together (the insulator) and ignores the ones far away (the tower structure).
RANSAC & PCA: These are like drawing a straight line through a messy scatter of dots to find the true center and direction of the insulator, ignoring the "noise" or outliers.

The best combination found was DBSCAN + RANSAC. It's like using a sieve to catch the good points and then a ruler to measure them perfectly. This allows the drone to pinpoint the insulator's location with an error of only about 6 inches (16 cm)—close enough to take a crystal-clear photo.

3. The "One-Trip Strategy" (Online Inspection)

Old methods were like a two-step dance:

Step 1: Fly around the tower once just to map it and find where the insulators are.
Step 2: Fly around a second time, following a pre-planned path, to take the actual photos.

This new method is a one-step sprint. The drone flies around the tower, and the moment it spots an insulator, it instantly calculates the best spot to hover and snap a photo. It does this on the fly, without stopping to make a map first.

The Result?
In computer simulations, this "one-trip" method saved 24% of the time compared to the old two-trip method. It's like ordering a pizza: instead of calling to order, waiting for the driver to find the house, and then waiting for the food, the driver arrives with the pizza already in hand.

4. Real-World Proof

The team didn't just test this in a computer game; they flew a real drone (with a powerful computer on board) around a real high-voltage tower.

The Challenge: Real life is messy. The camera lens got foggy, and the light wasn't perfect.
The Success: Even with these issues, the drone successfully found the insulators, calculated their location, and captured high-quality inspection images. The error rate remained incredibly low, proving the system is robust enough for real-world use.

The Bottom Line

This paper presents a "self-driving inspector" for power lines. It removes the need for pre-made maps, reduces the time the drone spends in the air, and does it all with a level of precision that ensures the photos are sharp enough to spot a tiny crack in an insulator. It's a significant step toward making our power grid safer, cheaper to maintain, and less reliant on human pilots.

1. Problem Statement

High-voltage power line insulators require regular inspection to prevent mechanical or electrical failures caused by contamination or physical damage. Current inspection methods face significant challenges:

Manual/Remote Control: Traditional inspections are labor-intensive, requiring ground access or manually piloted UAVs.
Dependency on Prior Maps: Existing autonomous systems typically require a pre-existing 3D map of the transmission tower structure and the exact locations of insulators. This necessitates a two-flight strategy: first, a scanning flight to create a map, and second, an inspection flight to visit pre-computed waypoints.
Inefficiency: The two-flight approach increases mission time and operational costs.
Localization Limitations: Previous methods using binocular cameras or pure LiDAR often suffer from high localization variance or inefficiency in capturing detailed images in a single pass.

The core problem addressed is how to enable a UAV to autonomously inspect an unmapped transmission tower in a single flight, without prior knowledge of the tower's exact geometry or the number/position of insulators.

2. Methodology

The authors propose an online inspection algorithm that integrates camera and LiDAR sensor fusion to detect, localize, and inspect insulators in real-time.

A. Sensor Fusion Pipeline

The system fuses data from a monocular RGB camera and a LiDAR sensor (Livox Mid-360):

Insulator Detection (Vision): A lightweight Convolutional Neural Network (YOLOv11n) detects insulators in the RGB image stream. The model was optimized using TensorRT for real-time inference.
Data Projection: LiDAR point cloud data is projected onto the 2D image plane using known extrinsic transformations between the LiDAR, UAV body, and camera frames.
Point Cloud Filtering: Points in the 3D space that project inside the 2D bounding boxes generated by the YOLO detector are isolated. This filters out background noise and tower structures, retaining only points likely belonging to the insulator.
Accumulation: To densify the sparse LiDAR data, point clouds from three consecutive detection events are cumulated before localization.

B. Insulator Localization Algorithms

Four distinct methods were proposed and compared to estimate the 3D center and orientation of the insulator from the filtered point cloud:

DBSCAN: Density-based spatial clustering.
RANSAC: Random Sample Consensus for line fitting.
DBSCAN + RANSAC: Clustering followed by RANSAC line fitting to refine the orientation and center.
DBSCAN + PCA: Clustering followed by Principal Component Analysis (PCA) to iteratively filter outliers and refine the center.

The algorithms assume the insulator has a cylindrical shape. They select the cluster nearest to the UAV, estimate the longitudinal axis, and compute the center point.

C. Online Inspection State Machine

The UAV operates on a state machine with two primary modes:

Exploration: The UAV flies a pre-planned path around the tower (based on approximate GPS coordinates and tower dimensions) to search for insulators.
Inspection: Upon detecting and localizing an insulator, the system calculates optimal inspection waypoints to capture high-resolution images. The UAV pauses exploration, flies to these waypoints, captures images, and then returns to the exploration path.

Safety: The system uses tower dimensions (height/width) to define safety volumes, preventing collisions with power lines or the tower structure.
Memory: Detected insulator positions are stored in a world frame buffer to prevent re-inspecting the same object and to handle insulators on the far side of the tower (which are skipped for the current pass).

3. Key Contributions

Single-Flight Strategy: The first approach to inspect unmapped transmission towers in a single flight without prior structural maps, significantly reducing mission time.
Camera-LiDAR Fusion: A robust pipeline combining deep learning detection with geometric point cloud filtering to achieve precise 3D localization in real-time.
Comparative Localization Analysis: A systematic evaluation of DBSCAN, RANSAC, and PCA combinations, identifying DBSCAN + RANSAC as the most robust method for handling outliers and varying point densities.
Real-World Validation: Successful deployment on a custom UAV platform (Jetson Orin Nano, Pixhawk 6C, Livox LiDAR) in a high-voltage environment.

4. Results

Simulation Results

Detection Accuracy: The YOLOv11n model achieved a mean Average Precision (mAP50–95) of 0.81.
Localization Error: The DBSCAN + RANSAC method achieved the best performance with a mean localization error of 0.36 ± 0.21 m. It outperformed DBSCAN + PCA by approximately 30% in overall error.
Time Efficiency: Compared to the traditional two-flight strategy (scan then inspect), the single-flight online approach saved up to 24% of total inspection time for scenarios with 8 inspection waypoints. Even with 24 waypoints, it remained more efficient (only 4.6% additional time compared to the two-flight optimal path).

Real-World Experiments

Hardware: Deployed on a UAV with a Livox Mid-360 LiDAR and Basler camera.
Detection Performance: The mAP50–95 dropped to 0.17 in real-world conditions due to challenging factors like fogged lenses and non-ideal lighting, highlighting a gap between simulation and reality.
Localization Accuracy: Despite detection challenges, the DBSCAN + RANSAC method achieved a mean localization error of:
- XY Plane: $0.16 \pm 0.08$ m
- Altitude (Z): $0.16 \pm 0.11$ m
Variance Improvement: The proposed method achieved more than an order of magnitude lower variance in horizontal localization error compared to previous binocular camera-based approaches (which had errors around 0.92 m).
Outcome: The system successfully captured detailed inspection images of insulators on a live high-voltage tower without prior mapping.

5. Significance

This work represents a significant step forward in autonomous infrastructure inspection. By eliminating the need for a pre-scan mapping flight, the proposed system drastically reduces the time and cost associated with power line maintenance. The demonstrated ability to localize insulators with sub-20cm accuracy using a single-flight strategy makes the technology viable for widespread deployment in utility companies, enhancing grid reliability while minimizing human risk and operational downtime. The paper also highlights the critical importance of robust sensor fusion to compensate for the limitations of vision-only or LiDAR-only approaches in complex, unstructured environments.