University of Twente Student Theses
Using synthetic data to improve the performance of UAV-based vehicle detection and speed estimation models
Song, Yaozheng (2025) Using synthetic data to improve the performance of UAV-based vehicle detection and speed estimation models.
|
PDF
2MB |
| Abstract: | UAVs can flexibly monitor roads and have a broad range of applications in the field of vehicle speed estimation and traffic monitoring. However, existing UAV datasets are very limited, especially aerial data with vehicle annotations are rare and the coverage of the scene is narrow, which hinders the improvement of the performance of data-driven deep learning models. To address these issues, a semi-automated process is proposed to reconstruct a real outdoor scene, acquire and generate synthetic training data with annotations, and use these data to train deep learning models to improve their vehicle speed estimation performance. Our process is divided into four main steps: scene reconstruction, data acquisition and annotation, model training, and testing and comparing model performance. This method uses synthetic data to overcome the limitations of real data by enriching the training samples with realistic synthetic images and highly accurate annotations. Such synthetic images have been proved that they can significantly improve the robustness and accuracy of the model. In our experiments, models trained with synthetic datasets generalise well to real UAV videos, while models pre-trained with real datasets and fine-tuned with synthetic datasets have further improved performance. The results show that synthetic data can improve the accuracy of detection and speed estimation of deep learning models. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/106610 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page