Real-Time Drone Object Detection - VisDrone

End-to-end computer vision system for aerial object detection using the VisDrone dataset. The model detects and tracks vehicles, pedestrians and cyclists in real-time drone footage, deployed via REST API and integrated with live hardware through network protocol reverse engineering.

📄 Read the full technical blog post for detailed methodology, experiments, and analysis.

---

Results

Model Performance (VisDrone Validation Set)

Model	mAP50	mAP50-95	Precision	Recall	CPU Latency	FPS
YOLOv8n	33.2%	19.2%	44.7%	33.3%	71ms	14
YOLO26n	37.5%	19.9%	46.5%	35.8%	45ms	22

Key improvements: +4.3% mAP50, +36% faster inference, +2.5% higher recall

Per-Class Performance (YOLO26n)

Class	mAP50	Precision	Recall	Challenge
Vehicle	68.1%	66.0%	62.6%	✅ Large objects, well-detected
Pedestrian	33.3%	43.5%	34.8%	⚠️ Small, dense crowds
Bike	11.0%	29.9%	9.9%	❌ Tiny objects (12×18 px)

Hardware Integration (Syma Z3 Pro Drone)

Real-time inference on live drone video stream via reverse-engineered TCP protocol:

Metric	Value
Display FPS	27.5
Inference Latency	62.4 ms
Frame Drop Rate	0.0%
End-to-End Latency	~200ms
WiFi Range	~30m

---

Project Structure

drone-object-detection/
│
├── data/
│   ├── raw/                          # VisDrone dataset (download separately)
│   │   ├── visdrone.yaml             # YOLO config file
│   │   └── test_video.mp4            # Sample test video
│   └── processed/                    # Converted YOLO format (train/val)
│
├── models/
│   ├── baseline_yolov8n/             # Initial YOLOv8n training
│   └── yolo26n_visdrone/             # YOLO26n (latest, best performance)
│       └── weights/
│           ├── best.pt               # Best checkpoint
│           └── best.onnx             # ONNX export
│
├── notebooks/
│   ├── data_exploration.ipynb        # EDA, object size analysis, class distribution
│   ├── training_model_colab.ipynb    # GPU training on Google Colab
│   └── evaluating_model.ipynb        # Failure analysis, robustness testing
│
├── src/
│   ├── config.py                     # Paths and constants
│   ├── convert_to_yolo.py            # VisDrone → YOLO format converter
│   ├── model_test_image.py           # Single image inference
│   ├── model_test_video.py           # Video inference with metrics
│   ├── live_detection.py             # Real-time drone stream inference (TCP/UDP)
│   ├── optimize_model.py             # ONNX export and benchmarking
│   └── analyze_logs.py               # API performance analysis
│
├── deployment/
│   └── api.py                        # FastAPI REST service
│
├── outputs/
│   ├── plots/                        # Visualizations (failure analysis, robustness)
│   ├── logs/                         # API request logs (JSON)
│   ├── videos/                       # Recorded sessions with detections
│   └── *.csv                         # Robustness test results
│
├── requirements.txt
└── README.md

---

Quickstart

Prerequisites

git clone https://github.com/elouanXP/drone-object-detection
cd drone-object-detection
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

Dataset

Download the VisDrone-DET2019 dataset and extract:

• VisDrone2019-DET-train → data/raw/
• VisDrone2019-DET-val → data/raw/

Convert to YOLO format:

python src/convert_to_yolo.py

Training (Google Colab Recommended)

Open notebooks/training_model_colab.ipynb in Google Colab with T4 GPU:

• Training time: ~1.5 hours (35 epochs)
• Download trained weights: yolo26n_trained.zip

Inference

Single Image:

python src/model_test_image.py --image data/processed/val/images/0000001.jpg

Video:

python src/model_test_video.py --video data/raw/test_video.mp4

Live Drone Stream (Syma Z3 Pro):

python src/live_detection.py

Requires connection to drone WiFi (192.168.30.1). Press q to quit.

REST API

Start the FastAPI server:

python deployment/api.py

Access interactive docs: http://localhost:8000/docs

Test with curl:

curl -X POST "http://localhost:8000/predict" \
  -F "file=@data/processed/val/images/0000001.jpg" \
  -F "conf_threshold=0.3"

---

Technical Stack

Category	Tools
Deep Learning	PyTorch, Ultralytics YOLO26
Computer Vision	OpenCV, PIL
API	FastAPI
Optimization	ONNX Runtime
Tracking	ByteTrack (Kalman filter)
Video Decode	ffmpeg
Network Analysis	nmap, Wireshark
Experiment Tracking	MLflow (optional)
Visualization	matplotlib, seaborn

---

Author

elouanXP | Portfolio