Drone Detection and Tracking Using YOLOv11x
Short Description
This repository presents a YOLOv11x-based deep learning model designed for drone detection and tracking. Optimized for small, fast-moving aerial targets, the model integrates detection with multi-object tracking (e.g., ByteTrack or DeepSORT) to maintain object identities over time. Built in PyTorch and trained using the Ultralytics API, the pipeline also includes optional heatmap visualization to monitor target movement. The full codebase, training details, and example inference scripts are available on
GitHub.
Model Details
YOLOv11x is the extra-large configuration of the YOLO family, known for its high accuracy on small object detection tasks. This model has ~56.8M parameters and achieves real-time inference speeds on modern GPUs. It was trained in Google Colab Pro+ using an NVIDIA L4 GPU, with a focus on detecting a single class: drone. The architecture and training settings have been chosen to balance precision, recall, and inference latency for aerial surveillance applications.
Intended Use
This model is primarily intended for research, prototyping, and real-time monitoring scenarios involving aerial objects. It is particularly suitable for security, wildlife monitoring, or airspace awareness systems where drone detection is critical. However, it should not be directly deployed in safety-critical or regulatory enforcement systems without extensive additional testing, calibration, and legal compliance checks.
Training Data
The dataset consists of just over 1,000 training images and ~350 validation images, all annotated in YOLO format for a single class (drone). Data augmentations included mosaic composition, random scaling, and HSV color jitter. Images were resized and letterboxed to 640×640 during training. While the dataset is sufficient for proof-of-concept and research, domain adaptation is strongly recommended for deployment in different environments.
Training Procedure & Hyperparameters
The model was fine-tuned from pretrained YOLOv11x weights using the following setup:
- Optimizer: AdamW
- Initial Learning Rate: 0.001
- Batch Size: 16
- Epochs: 32
- Image Size: 640×640
- Pretrained weights:
yolov11x.pt
Training was conducted entirely on Colab Pro+ with GPU acceleration, and the best.pt checkpoint was selected based on validation mAP performance.
Evaluation
On the validation set, the model achieved a precision of 0.922, recall of 0.831, mAP@50 of 0.905, and mAP@50–95 of 0.546. Inference latency is approximately 8.9 ms/image, enabling real-time operation. These results demonstrate strong precision for the drone detection task, though generalization to unseen conditions depends on dataset similarity.
Inference, Tracking & Visualization
Inference is straightforward using the provided scripts. The model can process both images and videos, outputting bounding boxes with confidence scores. When combined with a tracking algorithm such as ByteTrack or DeepSORT, it maintains object identities across frames. The repository also includes optional heatmap visualization tools to summarize object movement patterns over time.
Example usage:
1model = YOLO("best.pt")
2result = model.track(source="video.mp4", conf=0.3, iou=0.5, show=True)
Limitations & Risks
- Limited dataset size may reduce generalization to new domains.
- Single-class detection; cannot differentiate between drone types or other aerial objects.
- Possible false positives in cluttered or noisy backgrounds.
- May miss detections under severe occlusion, extreme weather, or poor lighting.
- Misuse potential for privacy-invasive surveillance applications.
- Requires compliance with local laws and ethical guidelines.
Recommendations
For improved robustness, users should fine-tune the model on in-domain data, adjust detection thresholds to match operational needs, and apply domain-specific augmentations. When using tracking, experiment with parameters to optimize ID consistency in different scenarios.
License
This project is licensed under the
MIT License.
