The project covers dataset preparation, model training with different configurations, evaluation, and a basic hardware demo. I trained four YOLOv8 variants to compare data strategies, then built a streaming setup to test detection on live video.
Drone Detection with YOLOv8 – A Learning Project
Computer VisionYOLOv8Object DetectionPythonPyTorchDeep LearningThesis
Project Overview
**Note:** This is a first draft and rough AI translation of the original documentation. The content below may contain inconsistencies or unclear phrasing. I'll update it with a proper manual revision when I get the time. Thanks for your patience! --- This started as a university assignment exploring object detection. The goal was to train a YOLOv8 model to detect drones and see what it takes to go from training to a working demo. I used two public datasets, experimented with augmentation strategies, and built a simple streaming setup with an ESP32-CAM. The project taught me a lot about dataset quality, the gap between test metrics and real-world performance, and the practical challenges of deploying ML models.
Version:v1.0
Time:~80.0 hours
Cost:$0 (if you have a GPU)
Status:completed
Materials
- NVIDIA GPU (CUDA capable) × 1
- ESP32-CAM module × 1
- Laptop for inference × 1
- Anti-UAV Dataset × ~200 videos
- VisioDECT Dataset × ~20k images
- Python 3.8+ × 1
Tools
- PyTorch
- YOLOv8 (Ultralytics)
- OpenCV
- Matplotlib
I used two public datasets to get variety in conditions and drone appearances.
2.1
Anti-UAV Dataset
The primary dataset, containing ~200 training videos at 1080p. I sampled frames to avoid redundancy. Most images show drones at long range against cloudy skies—good for distance detection but limited in lighting variety.
- Resolution: 1920×1080
- Conditions: Mostly cloudy, distant subjects
2.2
VisioDECT Dataset
Added for variety. Contains ~20k images across sunny, cloudy, rainy, and night conditions. Lower resolution (852×480) but drones appear closer and more detailed.
- Conditions: Sun, rain, night
- Advantage: Better visibility of drone features
2.3
Augmentations
I tested geometric augmentations (vertical and horizontal flips) to see their effect on generalization. Results were mixed—more on this in the training section.
All training used PyTorch and Ultralytics YOLOv8. I trained four model variants using a continuous training approach—starting from previous weights rather than scratch each time.
3.1
Setup
- Epochs: 20 per variant
- Batch Size: 16
- Image Size: 640px
- Approach: Continuous training (each model builds on the previous)
3.2
Model 1: Baseline
Anti-UAV dataset only, no augmentations. Established a baseline to compare against.
3.3
Model 2: + Vertical Flip
Added vertical flip augmentation. The idea was that drone silhouettes might be orientation-independent.
3.4
Model 3: + Horizontal Flip
Added horizontal flip augmentation. This turned out to hurt performance—likely because mirroring creates implausible lighting patterns (sun position becomes inconsistent).
3.5
Model 4: Combined Datasets
Combined Anti-UAV and VisioDECT datasets without the horizontal flip. The added variety in conditions and drone appearances improved generalization.
Each model was evaluated on a held-out test set using standard metrics.
4.1
Model 1 (Baseline)
- mAP@50: 0.92
- mAP@50-95: 0.58
- Precision: 0.91
- Recall: 0.88
Reasonable baseline. Some false positives but generally distinguished drones from background.
4.2
Model 2 (+ Vertical Flip)
- mAP@50: 0.93
- mAP@50-95: 0.60
- Precision: 0.92
- Recall: 0.89
Slight improvement. Vertical flip had limited but positive impact.
4.3
Model 3 (+ Horizontal Flip)
- mAP@50: 0.87
- mAP@50-95: 0.52
- Precision: 0.85
- Recall: 0.84
Performance dropped. Confidence scores also decreased noticeably. The augmentation likely introduced unrealistic training examples.
4.4
Model 4 (Combined Data)
- mAP@50: 0.95
- mAP@50-95: 0.64
- Precision: 0.94
- Recall: 0.91
Best overall. Adding diverse real data helped more than augmentation tricks.
4.5
Comparison
Model 4 generalized best, but Model 1 was competitive on the original dataset. More data helps generalization but adds complexity. Not all augmentations are beneficial—horizontal flipping hurt this specific task.
To test beyond offline metrics, I built a simple streaming demo.
Hardware:
- ESP32-CAM module streaming video over WiFi
- Laptop running inference (the ESP32 doesn't have the power to run YOLO)
Software:
- Model weights loaded in PyTorch
- OpenCV for frame capture and visualization
- Simple web interface for the detection feed
Takeaway: The gap between test metrics and real performance is real. The ESP32-CAM's lower resolution and different lighting conditions challenged the model. Useful for understanding practical deployment issues, even if it's not a production setup.
- Data variety matters more than augmentation tricks. Adding VisioDECT helped more than any augmentation I tried.
- Not all augmentations help. Horizontal flipping hurt performance for this task.
- Test metrics don't tell the whole story. Real-world conditions (lighting, resolution, distance) introduced challenges the test set didn't capture.
- Building a demo teaches you a lot. Actually running the model on live video revealed issues I wouldn't have found otherwise.
If I revisit this project:
- Try photometric augmentations (brightness, contrast, noise) instead of geometric ones
- Experiment with newer architectures (YOLOv9, RT-DETR)
- Collect custom data that better matches deployment conditions
- Explore actual edge deployment with model quantization
Results
Model 4 (combined datasets) performed best overall. Dataset diversity mattered more than augmentation strategies. Horizontal flipping hurt performance, likely due to lighting inconsistencies.
- Model 1 mAP@50: 0.92
- Model 1 mAP@50-95: 0.58
- Model 2 mAP@50: 0.93
- Model 2 mAP@50-95: 0.60
- Model 3 mAP@50: 0.87
- Model 3 mAP@50-95: 0.52
- Model 4 mAP@50: 0.95
- Model 4 mAP@50-95: 0.64
- Model 4 Precision: 0.94
- Model 4 Recall: 0.91
Safety Notes
Software project—no physical risks. Watch for disk space with large datasets and GPU temps during training.