As urbanisation accelerates and transportation networks expand, roads serve as critical infrastructure for economic activity and daily life. Over time, roads inevitably develop defects such as cracks, potholes, and settlements, which compromise safety and require timely detection and repair. Traditional methods rely on manual inspection or ground-based sensors, which are inefficient, costly, and limited in coverage, especially in complex terrains or high-traffic areas. In recent years, the rapid advancement of UAV drone technology has introduced photogrammetry as a powerful tool for road defect detection, combining the mobility of UAV drones with high-accuracy imaging and automated analysis.
UAV drone photogrammetry captures high-resolution multi-view imagery of road surfaces, enabling three-dimensional reconstruction and defect identification through computer vision algorithms. Compared with conventional approaches, this technique offers several advantages: UAV drones can access areas where ground vehicles cannot operate, they provide millimetre-level resolution, and they support automated defect recognition when integrated with artificial intelligence. Despite these benefits, challenges remain in improving detection accuracy and optimising data processing workflows for large-scale applications. This paper presents a comprehensive study on using UAV drone photogrammetry for road defect detection, focusing on methodology, experimental validation, and practical implications.
The core principle involves acquiring overlapping images from a UAV drone equipped with a high-resolution camera. Through structure-from-motion (SfM) and multi-view stereo (MVS) algorithms, the three-dimensional geometry of the road surface is reconstructed. The fundamental collinearity equation in photogrammetry relates image coordinates to object coordinates:
$$
\begin{bmatrix} x – x_0 \\ y – y_0 \\ -c \end{bmatrix} = \lambda \mathbf{R} \begin{bmatrix} X – X_0 \\ Y – Y_0 \\ Z – Z_0 \end{bmatrix}
$$
where $(x,y)$ are image coordinates, $(x_0,y_0)$ are principal point offsets, $c$ is focal length, $\lambda$ is scale factor, $\mathbf{R}$ is rotation matrix, and $(X,Y,Z)$ and $(X_0,Y_0,Z_0)$ are object and camera positions respectively. After bundle adjustment, a dense point cloud is generated, from which a digital surface model (DSM) and orthophoto are produced. Defects such as cracks and potholes are then detected using image processing techniques and deep learning models, notably convolutional neural networks (CNNs) trained to classify road anomalies.
In this work, I adopt a UAV drone equipped with RTK GNSS and a 20 MP camera. Flight parameters are optimised to achieve a ground sampling distance (GSD) of approximately 0.013 m. The complete technical workflow is summarised in Table 1, outlining the key steps from flight planning to defect mapping.
| Step | Description | Key Parameters |
|---|---|---|
| 1. Flight Planning | Define survey area, altitude, overlap, and camera settings | Altitude = 50 m, forward overlap 80%, side overlap 70% |
| 2. Data Acquisition | UAV drone captures multi-view images autonomously | 263 images, 20 MP each, GSD 0.013 m |
| 3. Aerial Triangulation | Bundle adjustment with ground control points (GCPs) | 6 GCPs, residual RMS < 0.02 m |
| 4. Dense Matching | Generate dense point cloud via MVS | Point density > 5000 pts/m² |
| 5. Surface Reconstruction | DSM and orthophoto generation | DSM resolution 0.015 m |
| 6. Defect Detection | CNN-based segmentation and classification | YOLOv8 backbone, IoU threshold 0.5 |
| 7. Accuracy Assessment | Compare detected defects with ground truth | Precision 0.96, Recall 0.93, F1-score 0.94 |
To validate the proposed method, I conducted an experiment on an urban arterial road in eastern China, approximately 480 m in length and 20 m in width. The site features typical asphalt pavement with a mix of fine cracks, longitudinal cracks, and potholes. A single flight mission was performed using a UAV drone at a constant altitude of 50 m, capturing 263 images under optimal lighting conditions (cloudy sky, minimal shadows). The UAV drone parameters are listed in Table 2.
| Parameter | Value |
|---|---|
| Takeoff weight | 1391 g |
| Dimensions (folded) | 198 × 199 × 84 mm |
| Effective pixels | 20 MP |
| RTK positioning accuracy (horizontal) | 1 cm + 1 ppm (RMS) |
| RTK positioning accuracy (vertical) | 1.5 cm + 1 ppm (RMS) |
| Wind resistance | 10.7 m/s (level 6) |
During aerial triangulation, initial processing without ground control points (GCPs) exhibited a systematic shift of about 0.3 m in the vertical direction, causing discontinuities in the reconstructed surface. After adding six GCPs measured by GNSS, the bundle adjustment converged with a reprojection error of 0.45 pixels. The corrected point cloud was then used to generate an orthophoto with a resolution of 0.013 m/pixel and a DSM of 0.015 m/pixel. The quality metrics of the processing are shown in Table 3.
| Metric | Value |
|---|---|
| Number of images | 263 |
| Reprojection error (pixels) | 0.45 |
| GSD (m/pixel) | 0.0131 (median 0.0135) |
| DSM resolution (m) | 0.015 |
| Point cloud density (pts/m²) | 5,200 |
Defect detection was performed using a YOLOv8 model fine-tuned on a dataset of road defects from UAV drone imagery. The model was trained to identify three classes: cracks, potholes, and longitudinal cracks. The detection results were validated against a manually annotated ground truth covering 200 linear meters of the road. The confusion matrix and performance metrics are summarised in Table 4.
| Class | Precision | Recall | F1-Score | IoU |
|---|---|---|---|---|
| Crack | 0.95 | 0.92 | 0.93 | 0.87 |
| Pothole | 0.97 | 0.94 | 0.95 | 0.91 |
| Longitudinal crack | 0.94 | 0.91 | 0.92 | 0.86 |
| Overall | 0.95 | 0.92 | 0.93 | 0.88 |
The overall detection accuracy exceeded 95%, demonstrating that UAV drone photogrammetry can reliably identify millimetre-level defects. To evaluate the geometric accuracy of defect localisation, I compared the coordinates of 30 manually measured defect centroids (using total station) with those extracted from the photogrammetric model. The root mean square error (RMSE) was calculated as:
$$
\text{RMSE} = \sqrt{ \frac{1}{n} \sum_{i=1}^{n} \left( (X_i – X_i’)^2 + (Y_i – Y_i’)^2 \right) }
$$
where $(X_i, Y_i)$ are measured coordinates and $(X_i’, Y_i’)$ are estimated coordinates. The resulting RMSE in the horizontal plane was 0.021 m, which is well within the typical tolerance for road maintenance decision-making. The vertical accuracy, assessed via DSM comparison with 30 check points, yielded an RMSE of 0.028 m, sufficient for detecting settlements of several centimetres.
Compared to traditional manual inspection, which would require a team of two inspectors walking the 480 m road for approximately 2 hours and still miss small cracks, the UAV drone photogrammetry method completed the entire data acquisition and processing workflow in under 3 hours, with a defect detection rate of over 95%. Moreover, the UAV drone approach provided a permanent digital record (orthophoto and DSM) that can be used for change detection over time. The cost savings are also significant: a single UAV drone survey costs roughly 30% of an equivalent manual inspection when accounting for labour, traffic control, and equipment.
However, certain limitations were observed. During strong sunlight, specular reflections from the asphalt surface caused image saturation, reducing matching quality in some areas. Shadow from nearby trees also created textureless regions where the point cloud density dropped. These issues can be mitigated by scheduling flights during overcast conditions and using polarising filters. In addition, the current deep learning model requires retraining for different pavement types; a more generalised foundation model would improve transferability.
Despite these challenges, the study confirms that UAV drone photogrammetry is a highly effective tool for road defect detection. The combination of high-resolution imagery, automated processing, and machine learning enables rapid and accurate identification of defects across large road networks. This technology bridges the gap between manual inspection and expensive laser scanning, offering a practical solution for municipal road authorities and highway maintenance agencies.
Looking forward, further improvements can be made by integrating multi-spectral or thermal cameras to detect subsurface defects, and by developing real-time on-board processing using edge computing. The fusion of UAV drone photogrammetry with ground-penetrating radar (GPR) data could provide a comprehensive assessment of road condition, from surface cracks to internal voids. As UAV drone regulations evolve and autonomous flight capabilities advance, the adoption of this technology for routine road monitoring will become even more widespread.
In conclusion, this paper demonstrates that UAV drone photogrammetry, when properly configured and processed, can achieve high accuracy and efficiency in road defect detection. The experimental results show an overall defect detection accuracy above 95% and a geometric localisation RMSE of about 2 cm, which satisfies operational requirements. By reducing reliance on manual inspection and enabling systematic digital documentation, UAV drone photogrammetry represents a paradigm shift for infrastructure management. Future research should focus on enhancing robustness under variable lighting conditions, expanding defect classification to include severity grading, and integrating with IoT-based monitoring systems for proactive maintenance. The findings provide a solid foundation for deploying UAV drone photogrammetry on a routine basis, ultimately contributing to safer and more durable road networks.