DroneWatch — On-Device Drone Detection for Law Enforcement

"Dedrone costs $50,000. This runs on a MacBook Air."

Real-time drone detection and tracking built on YOLO26 MLX — running entirely on Apple Silicon with no cloud, no internet, and no $50k hardware contract.

Built for the YOLO26 MLX Build Challenge (Enterprise track) in response to a real pain point: unauthorized drones are an active and growing threat for law enforcement, correctional facilities, and event security. Existing commercial solutions cost $50k–$500k+ and require cloud infrastructure. DroneWatch runs on a $600 Mac mini.

---

Features

• Real-time detection from webcam or video file — YOLO26n fine-tuned on 418 drone/bird/plane images
• Persistent track IDs — drones keep their ID even when briefly occluded
• Permanent drone lock — once confirmed as a drone, the track never downgrades (handles bird/drone overlap)
• Audio alert — fires a ping on first detection with 3-second cooldown to avoid spam
• Annotated video recording — saves evidence MP4 with bounding boxes and track IDs
• 100% on-device — MLX Metal acceleration on M1–M4, no internet required

---

Quick Start

# Clone and install
git clone https://github.com/jesusr04/drone-detection.git
cd drone-detection
python -m venv .venv && source .venv/bin/activate
pip install -e . --no-deps
pip install numpy opencv-python lap scipy

# Run live detection (built-in webcam)
python detect_drones.py

# iPhone as webcam via Continuity Camera
python detect_drones.py --source 1

# Run on a video file
python detect_drones.py --source path/to/video.mp4

# Save annotated output video
python detect_drones.py --source 0 --save

# Use retrained weights with faster tracker
python detect_drones.py --model runs/train/exp/best.safetensors --conf 0.25 --tracker bytetrack

---

CLI Options

Flag	Default	Description
--model	models/yolo26n-drone.safetensors	Path to model weights
--source	0	Camera index or video file path
--conf	0.20	Detection confidence threshold
--tracker	configs/dronewatch.yaml	Tracker config (bytetrack for faster)
--save	off	Save annotated video to results/

Press q to quit.

---

Architecture

Camera / Video File
        ↓
  cv2.VideoCapture
        ↓
  YOLO26n-MLX (fine-tuned on drones)
    model.track(persist=True)
        ↓
  Permanent drone lock logic
    └── 4 consecutive detections → confirmed drone
    └── Confirmed tracks never downgrade (handles bird/drone overlap)
        ↓
  Alert: audio ping + console log
        ↓
  Annotated frame → live display + optional saved video

---

Fine-Tuning

Dataset: 418 labeled images (Bird / Drone / Plane) from Roboflow, supplemented with personal drone footage.

python retrain.py
# Weights saved to runs/train/exp/best.safetensors

Config: configs/drone.yaml

---

Dataset

Split	Images
Train	318
Val	77
Test	23

Classes: Bird (0), Drone (1), Plane (2)

---

Requirements

• Apple Silicon Mac (M1–M4) — MLX is Apple Silicon only
• Python 3.10–3.12
• macOS 12+
• Camera access granted in System Settings → Privacy & Security → Camera

---

Built With

• YOLO26 MLX — pure MLX YOLO26 implementation by webAI
• MLX — Apple's ML framework
• OpenCV — video capture and display
• ByteTrack / BoT-SORT — multi-object tracking with persistent IDs

---

License

AGPL-3.0 — see LICENSE