Book chapter
Data Augmentation Pipeline for Enhanced UAV Surveillance
Pattern Recognition, pp 366-380
03 Dec 2024
Abstract
The growing use of Unmanned Aerial Vehicles (UAVs) presents considerable safety and security challenges, requiring improved capabilities for drone detection. On the other hand, collecting real-world data is costly, time consuming, and in some cases subject to rules and regulations. This paper addresses this issue by introducing a novel two-phase data augmentation approach aimed at improving the accuracy and efficiency of long-range drone detection systems. The initial phase involves the generating of synthetic data, using Unreal Engine, to increase the diversity and richness of the dataset. Subsequently, we employ Cycle-consistent Generative Adversarial Network (CycleGAN) to translate these synthetic images into more realistic representations, thus bridging the gap between synthetic and real datasets. This methodology not only seeks to refine the precision of drone detection algorithms but also presents a cost-effective solution for creating extensive and varied drone detection datasets. To evaluate the efficacy of our proposed data augmentation pipeline, we trained the You Only Look Once (YOLO) detection model under three distinct scenarios: utilizing purely real data, augmenting real data with synthetic data, and augmenting real data with CycleGAN-translated synthetic data. The performance of these models was assessed using four separate, unseen real datasets: DetFly, Drone Detection, UAV Detect, and New Batch. Our findings indicate a marked improvement in drone detection capabilities when the training dataset includes both synthetic and translated images, with the most significant enhancement observed in scenarios where real data is augmented with translated data.
Metrics
10 Record Views
Details
- Title
- Data Augmentation Pipeline for Enhanced UAV Surveillance
- Creators
- Solmaz ArezoomandanJohn KlohokerDavid K. Han
- Contributors
- Apostolos Antonacopoulos (Editor)Subhasis Chaudhuri (Editor)Rama Chellappa (Editor)Cheng-Lin Liu (Editor)Saumik Bhattacharya (Editor)Umapada Pal (Editor)
- Publication Details
- Pattern Recognition, pp 366-380
- Series
- Lecture Notes in Computer Science
- Publisher
- Springer Nature Switzerland; Cham
- Number of pages
- 15
- Resource Type
- Book chapter
- Language
- English
- Academic Unit
- Electrical and Computer Engineering
- Scopus ID
- 2-s2.0-85211916698
- Other Identifier
- 991021985404604721