Abstract
Inspecting deep-tunnel sewerage systems is a hazardous task traditionally performed manually. While Unmanned Aerial Vehicles (UAVs) offer a safer alternative, standard image stitching algorithms often fail in these environments due to repetitive textures and non-ideal camera motion. This paper introduces a "feature-less" framework that leverages known scene geometry and recorded camera trajectories to bypass traditional feature-matching. By projecting high-resolution images onto a 3D cylindrical model, the system creates seamless panoramic visualizations for infrastructure inspection.
Introduction
Traditional image stitching relies on identifying matching features (e.g., SIFT or ORB) between overlapping frames. However, in tunnels, repetitive patterns like bricks or uniform concrete surfaces often lead to matching failures. Furthermore, UAVs navigating these spaces exhibit complex motion—including forward translation and jitter—that violates the assumptions of standard panoramic stitching, which typically assumes pure rotation or a perfectly flat scene.
This research addresses these challenges by proposing a geometry-driven approach. Instead of searching for visual "landmarks," the framework uses the UAV's pose data and the known cylindrical geometry of the tunnel to mathematically determine the placement of every pixel.
Methodology
The core of the proposed framework is a deterministic geometric solution for image projection that prioritizes scene priors over visual evidence.
Cylindrical Projection and Inverse Warping
The process begins by transforming 2D pixel coordinates from the captured images into 3D rays using the camera's intrinsic parameters, such as focal length and optical center. These rays are then intersected with a virtual cylinder of a known radius ($r$).
To ensure high visual quality and avoid aliasing or "holes" in the final output, the authors employ inverse warping. Rather than projecting pixels forward from the source image, the algorithm iterates through every pixel in the target cylindrical map and "looks back" to find the corresponding intensity value in the source images based on the UAV’s recorded pose (Rotation $R$ and Translation $t$). This method explicitly accounts for cases where the UAV is flying off-center or wobbling.
Unity Integration
The resulting stitched textures are wrapped back onto 3D cylindrical models within the Unity game engine. This allows inspectors to navigate a virtual reconstruction of the tunnel, providing a more intuitive interface for identifying defects than flat panoramic strips.
Experimental Results
The framework was validated using a synthetic dataset generated in Blender and Unity, featuring a 3-meter radius tunnel with a brick texture to provide clear visual feedback on alignment.
- Robustness to Motion: The system successfully handled simulated UAV jitter and off-center flight paths. By using ground-truth trajectory data, the algorithm corrected for random movements that would typically break feature-based stitching.
- Reconstruction Quality: In noiseless scenarios, the algorithm produced perfect, seamless panoramic strips, effectively "straightening" the curved tunnel walls for inspection.
- Visualization: The Unity-based interface demonstrated the feasibility of a virtual "fly-through" for identifying simulated defects on the tunnel walls, proving the utility of high-resolution narrow-field-of-view cameras over lower-resolution 360° alternatives.
Discussion and Future Work
The primary advantage of this method is its reliance on geometry over visual features, making it robust in dark, repetitive environments. However, the system's performance is currently tied to the accuracy of the UAV's pose estimation (IMU/SLAM) and the assumption of a constant, known tunnel radius.
Future research will focus on integrating depth sensors (such as LiDAR) to automatically estimate tunnel dimensions and handle more complex geometries, such as curved pipe sections or varying diameters. The authors also plan to validate the framework using real-world datasets from operational sewerage tunnels.
Summary updated