Abstract: The objective of this work is to develop a data processing system that can automatically generate waypoints for navigation of an unmanned aerial vehicle (UAV) to inspect surfaces of structures like buildings and bridges. The input includes data recorded by two 2D laser scanners, orthogonally mounted on the UAV, and an inertial measurement unit (IMU). Algorithms are developed for data registration, surface and obstacle detection, and waypoint generation. Experiments show that the system successfully reconstructs 3D models and generates efficient flight paths for structural inspection.
Introduction
Inspecting large-scale infrastructure such as bridges, dams, and towers is traditionally a dangerous, time-consuming, and expensive process. While Unmanned Aerial Vehicles (UAVs) offer a safer alternative, current workflows often suffer from significant inefficiencies. Pilots typically fly manually or follow pre-set GPS paths that do not account for the specific geometry of the structure or local obstacles. This often results in "data bloat"; for instance, one study collected 12,000 images, but only a few hundred were useful for actual inspection. This research aims to bridge the gap between raw sensor data and intelligent navigation, creating a system that understands its environment to plan optimal inspection paths.
Methodology
The proposed system utilizes a multi-stage pipeline designed for computational efficiency to allow for rapid deployment in construction environments.
Hardware Configuration
The UAV sensor payload consists of two orthogonally mounted 2D laser scanners and an IMU to capture 3D spatial data during yaw maneuvers. Source, p. 2
To avoid the weight and power constraints of heavy 3D LiDAR, the system employs two orthogonally mounted 2D laser scanners and an Inertial Measurement Unit (IMU). As the UAV performs a 360° yaw maneuver, the vertical scanner builds the 3D model while the horizontal scanner estimates translation. By using the IMU to provide rotation data, the system simplifies a complex 6-DOF registration problem into a 3-DOF one, significantly reducing computational load.
Data Processing Pipeline
The end-to-end data processing workflow, from raw sensor input to the generation of navigation waypoints. Source, p. 2
The software architecture transforms raw point clouds into actionable flight plans through several key phases:
- Registration and Pre-processing: Raw data is aligned and filtered using statistical outlier removal. Voxelization is then applied to group points into a 3D grid, equalizing point density and simplifying geometry for faster processing.
- Surface and Obstacle Detection: The system uses RANSAC (Random Sample Consensus) to identify planar surfaces like walls or bridge decks. Points that do not fit these planes are clustered and stored as obstacles in an octree structure.
- Path Planning: Once a target surface is identified, the system calculates camera positions to ensure full coverage with a specific overlap (e.g., 20%). It then employs the A* algorithm on the voxel grid to find the shortest collision-free path between these waypoints.
Experimental Results
The system was evaluated using both controlled geometric datasets and real-world data from a bridge measuring approximately 22m x 10m x 4.5m.
- Efficiency: The total processing time from loading raw data to generating a full waypoint mission was only a few minutes. However, analysis showed that 83% of processing time is consumed by the alignment and registration of point clouds.
- Accuracy: In the bridge experiment, the alignment error between the first and last scans after a full rotation was only 10 cm, which is highly acceptable for large-scale structural inspection.
- Coverage: For a 220 m² bridge surface, the system automatically generated 1,146 stop points and 6,699 waypoints to ensure complete coverage with 20% image overlap.
Discussion and Future Work
A key innovation of this work is the "hardware-software" co-design choice to decouple translation and rotation using the IMU and horizontal scanner. Furthermore, the use of voxelization to equalize density prevents the RANSAC algorithm from being biased toward high-density areas near the sensor, leading to more accurate plane detection.
Currently, the system faces limitations with concave geometries, which sometimes require manual boundary adjustments. Additionally, the path planning assumes a static environment. Future work will explore the deployment of multi-UAV "swarms" to further accelerate the scanning of massive infrastructure projects.
Summary updated