YOLO-GCOF: A Lightweight Low-Altitude Drone Detection Model
The maturity of the drone industry and the open access to low-altitude airspace have resulted in massive use of low-altitude drones. While the use of drones has brought economic efficiency, the illegal use of drones has raised concerns about increased drone trespass in the no-fly zone and countermea...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10937103/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The maturity of the drone industry and the open access to low-altitude airspace have resulted in massive use of low-altitude drones. While the use of drones has brought economic efficiency, the illegal use of drones has raised concerns about increased drone trespass in the no-fly zone and countermeasure strategies. These concerns underscore the urgent need for low-altitude drone object detection. Traditional radar detection methods often struggle to detect small drones at low altitudes. To address this issue, we propose a lightweight model named YOLO-GCOF, leveraging an optimized version of the YOLOv8 architecture to enhance accuracy and enable efficient deployment on embedded devices. YOLO-GCOF incorporates the GSConv-Integrated Dynamic Group Convolution Shuffle Transformer (GI-DGCST) module as the feature extraction module, which captures fine-grained details and improves the detection of small-scale features. The integration of multi-level feature information is achieved by incorporating the Coordinate-Scale Attention High-level Screening-Feature Fusion Pyramid Networks (CSA-HSFPN) into the neck layer. The HSFPN architecture enables efficient multi-scale feature fusion, and the CSA module optimizes feature selection across scales, improving the model’s ability to detect key features using both scale and coordinate information. The Detect-Olive Group Convolution Head (Detect-OGCH) module, a lightweight parameter-sharing design, is integrated into the head layer to significantly reduce the model’s GFLOPs and parameter count while maintaining detection performance. Lastly, the original Complete Intersection over Union (CIoU) is replaced with the Focaler-SCYLLA IoU (Focaler-SIoU) to improve detection accuracy. The YOLO-GCOF model outperforms the original YOLOv8n, as demonstrated by a 1.1% improvement in mAP@50, alongside reductions in parameter count, computational overhead, and model size by 60%, 49.4%, and 55.1%, respectively. Additionally, the YOLO-GCOF model achieved 91.5% mAP@50 on the DUT-Anti-UAV dataset. These results and successful deployment on Raspberry Pi 4B demonstrate that the YOLO-GCOF is highly effective for lightweight, high-performance low-altitude drone detection. |
|---|---|
| ISSN: | 2169-3536 |