The proliferation of UAV technology has enabled efficient aerial data acquisition for applications like disaster response, topographic surveying, and environmental monitoring. Low-altitude fixed-wing camera UAVs capture high-resolution images, but individual frames cannot cover entire survey areas. Stitching these into seamless panoramas is essential for comprehensive spatial analysis. This study addresses computational bottlenecks in large-scale stitching by integrating POS data with optimized geometric correction.
1. Preprocessing of Low-Altitude Fixed-Wing UAV Aerial Images
Raw images from non-metric cameras exhibit lens distortion, noise, and uneven illumination. Distortion correction uses calibrated parameters:
$$ \begin{cases} \Delta x = x(K_1 r^2 + K_2 r^4) + 2P_2 xy + P_1(r^2 + 2x^2) \\ \Delta y = y(K_1 r^2 + K_2 r^4) + 2P_1 xy + P_2(r^2 + 2y^2) \end{cases} $$
where \( r^2 = x^2 + y^2 \), and \( K_1 \), \( K_2 \), \( P_1 \), \( P_2 \) are distortion coefficients. Histogram equalization enhances low-contrast images, while a 5×5 median filter reduces noise. Filter performance is evaluated using statistical metrics:
| Evaluation Metric | Band | Mean Filter (5×5) | Median Filter (5×5) |
|---|---|---|---|
| Gray Mean | R | 110.126 | 109.277 |
| G | 112.276 | 111.409 | |
| B | 117.569 | 116.845 | |
| Average Gradient | R | 3.736 | 4.080 |
| G | 3.528 | 3.878 | |
| B | 3.409 | 3.784 |
2. Geometric Correction Using POS Data
POS data (latitude, longitude, roll φ, pitch ω, yaw κ) enable geometric correction through collinearity equations. Image-to-ground coordinate conversion is expressed as:
$$ \begin{bmatrix} u \\ v \\ w \end{bmatrix} = \mathbf{R} \begin{bmatrix} x \\ y \\ -f \end{bmatrix}, \quad \mathbf{R} = \begin{bmatrix} a_1 & a_2 & a_3 \\ b_1 & b_2 & b_3 \\ c_1 & c_2 & c_3 \end{bmatrix} $$
where rotation matrix elements are derived from attitude angles. Ground coordinates are calculated using:
$$ X = (Z – Z_s) \frac{a_1 x + a_2 y – a_3 f}{c_1 x + c_2 y – c_3 f} + X_s, \quad Y = (Z – Z_s) \frac{b_1 x + b_2 y – b_3 f}{c_1 x + c_2 y – c_3 f} + Y_s $$
Principal point coordinates are optimized using Total Least Squares (TLS) to minimize fitting errors. For coordinates \((x_i, y_i)\), TLS minimizes:
$$ \phi = \sum_{i=1}^{m} (a_0 + a_1 \hat{x}_i – y_i)^2 + \sum_{i=1}^{m} (\hat{x}_i – x_i)^2 $$
TLS reduces directional deviations by 5–10% compared to Least Squares in simulated error tests.
3. Stitching via ArcGIS Engine Mosaic Datasets
Corrected images are geotagged using TWF files containing top-left coordinates, rotation angles, and ground sampling distance (GSD). Mosaic datasets in ArcGIS Engine manage spatial indexing to overcome memory limitations:
$$ \text{GSD} = \frac{\text{Flight Altitude} \times \text{Sensor Width}}{\text{Focal Length} \times \text{Image Width}} $$
Key steps include:
- Batch projection of TIFF images
- Mosaic dataset creation
- Dynamic stitching using spatial queries
4. System Implementation and Evaluation
A C#-based system was developed in Visual Studio 2012 with ArcGIS Engine. The modular architecture includes:
| Module | Functions |
|---|---|
| View Management | Data loading, zoom, measurement |
| Map Annotation | Point/line/polygon drawing |
| Image Processing | Stretching, correction, stitching |
Tests used 71 images (0.08m GSD). Stitching accuracy was validated against ground control points:
| Sample Point | ΔX (m) | ΔY (m) | RMS (m) |
|---|---|---|---|
| 1 | -2.22 | 0.21 | 2.22 |
| 2 | -1.35 | -1.93 | 2.35 |
| 3 | 4.94 | -3.84 | 6.25 |
Errors exceeded 4m in high-relief areas due to unmodeled terrain effects. The camera drone stitching system processed 1,500+ images (40 km²) on 4GB RAM hardware.
5. Conclusions
This research presents a POS-based workflow for low-altitude fixed-wing camera UAV image stitching. Key innovations include:
- TLS-optimized POS data fitting improves geolocation accuracy
- Mosaic datasets enable large-scale processing on standard hardware
- The stitching system achieves 2–3m planimetric accuracy in flat terrain
Limitations include terrain-induced seams without DEM data. Future work will integrate feature-based matching with POS data to enhance robustness for complex topography. The camera UAV approach significantly reduces computational requirements while maintaining geospatial accuracy.