🛰️ CubeSat AI-Based Real-Time Collision Prediction System

A ground-based software platform that processes satellite camera imagery and telemetry data, detects nearby space debris, and predicts collision risks NEAR real time. The system uses a simulation layer that generates realistic synthetic orbital imagery and debris scenarios since live satellite feeds not available during development.

Architecture Overview — 5-Layer Pipeline

Layer	Name	Responsibility	Technology
1	Simulation Engine	Synthetic orbital imagery and debris movement	Python, OpenCV, NumPy
2	Data Ingestion	Receives frames & telemetry, queues for processing, persists tracks & alerts	FastAPI, Redis, SQLite
3	Vision & Tracking	Detects debris, tracks across time with object IDs	OpenCV, YOLOv8-nano, SORT
4	Prediction & Risk	Predicts trajectories, computes Pc, plans avoidance maneuvers	NumPy, SciPy, filterpy UKF, astropy
5	Dashboard & Alerts	Real-time 3D orbital display, risk scores, ground track, operator alerts	React, Three.js, WebSocket, Recharts

Data Flow: Simulation → Redis Queue → Detector → Tracker → Predictor → Risk Scorer → Dashboard

Models Used

• Detection: YOLOv8-nano (vision/yolo_config.yaml, model: yolov8n.pt) for debris detection.
• Tracking: SORT with a constant-velocity Kalman filter (vision/sort_tracker.py) for multi-object tracking.
• Prediction: Unscented Kalman Filter (UKF) (prediction/ukf_tracker.py) for trajectory estimation.
• Fallback: OpenCV SimpleBlobDetector (vision/object_detector.py) is used when YOLO weights are unavailable.

Repository Structure

cubesat/
├── simulation/              # Layer 1: Simulation Engine
│   ├── engine.py            # Main simulation engine
│   ├── star_field.py        # Static star field generator
│   ├── debris.py            # Debris object models and animation
│   ├── noise.py             # Noise generation (hot pixels, Gaussian, cosmic rays)
│   ├── telemetry.py         # TelemetryPacket generator
│   ├── config.py            # YAML config loader
│   ├── run.py               # Standalone runner
│   └── scenarios/           # Built-in scenarios
│       ├── safe_flyby.yaml
│       ├── close_approach.yaml
│       └── critical_conjunction.yaml
├── ingestion/               # Layer 2: Data Ingestion
│   ├── api.py               # FastAPI app (/health, /ws/live, REST endpoints)
│   ├── redis_client.py      # Redis stream producer/consumer
│   ├── queue_manager.py     # Frame-drop policy (max 30 frames)
│   ├── database.py          # SQLite persistence layer (tracks & alerts)
│   └── worker.py            # Async processing worker (vision → prediction loop)
├── vision/                  # Layer 3: Detection & Tracking
│   ├── preprocessing.py     # Dark subtraction, CLAHE, hot-pixel correction
│   ├── streak_detector.py   # Canny + Hough Line streak detection
│   ├── object_detector.py   # YOLOv8-nano wrapper with blob fallback
│   ├── detector.py          # Two-stage merged detection pipeline
│   ├── sort_tracker.py      # SORT multi-object tracker with Kalman internals
│   ├── pipeline.py          # Full vision pipeline orchestrator
│   └── yolo_config.yaml     # YOLOv8 training configuration
├── prediction/              # Layer 4: Trajectory & Risk
│   ├── coordinate_transform.py  # Pixel to ECI coordinate conversion
│   ├── orbital_dynamics.py      # J2, drag, SRP perturbation models (RK4)
│   ├── ukf_tracker.py           # UKF with 6-state [x,y,z,vx,vy,vz]
│   ├── closest_approach.py      # TCA computation
│   ├── collision_probability.py # Alfriend-Akella Pc calculation
│   ├── risk_assessor.py         # ADVISORY/WARNING/CRITICAL tier classification
│   ├── maneuver_planner.py      # Impulsive burn planner for collision avoidance
│   └── pipeline.py              # Full prediction pipeline orchestrator
├── dashboard/               # Layer 5: React Frontend
│   ├── src/
│   │   ├── App.jsx
│   │   ├── store.js         # Zustand state management
│   │   ├── hooks/useWebSocket.js
│   │   └── components/
│   │       ├── LiveFeed.jsx          # Live camera feed with overlay
│   │       ├── OrbitalView.jsx       # 3D orbital scene (Three.js / R3F)
│   │       ├── GroundTrack.jsx       # 2D sub-satellite point on world map
│   │       ├── TracksTable.jsx       # Active debris tracks table
│   │       ├── ObjectInspector.jsx   # Sliding sidebar for track deep-dive
│   │       ├── RiskAnalytics.jsx     # Pc history chart for selected track
│   │       ├── RiskTimeline.jsx      # Timeline of risk alert events
│   │       ├── AlertFeed.jsx         # Live alert feed
│   │       ├── ManeuverPanel.jsx     # Recommended avoidance maneuver display
│   │       ├── SystemHealth.jsx      # Subsystem health indicators
│   │       ├── SystemMetricsHUD.jsx  # Top-level mission statistics HUD
│   │       └── CriticalModal.jsx     # Full-screen critical alert overlay
│   ├── package.json
│   └── Dockerfile
├── shared/                  # Pydantic data schemas
│   └── schemas.py           # TelemetryPacket, ImageFrame, DetectionEvent, TrackObject, RiskAlert
├── tests/                   # 60 unit & integration tests
├── notebooks/               # Jupyter prototyping notebooks
├── docker-compose.yml
├── Dockerfile
├── Makefile
├── cubesat.db               # SQLite database (auto-created at runtime)
└── requirements.txt

Quick Start

Prerequisites

• Docker & Docker Compose
• Python 3.11+
• Node.js 20+

Run with Docker Compose mac/linux

# Start all services
make start

# View logs
docker compose logs -f

# Stop all services
make stop

Services available at:

• Backend API: http://localhost:8000
• Dashboard: http://localhost:3000
• API Docs: http://localhost:8000/docs

Local Development

# Install Python dependencies
pip install -r requirements.txt

# Run the FastAPI backend
uvicorn ingestion.api:app --host 0.0.0.0 --port 8000 --reload

# In another terminal, run the simulation
python simulation/run.py --scenario safe_flyby --frames 100

# Run the React dashboard
cd dashboard
npm install
npm run dev

Simulation Scenarios

Three built-in scenarios are available:

Scenario	Description	Expected Outcome
safe_flyby	Debris passes at >20 km distance	No WARNING+ alerts
close_approach	Debris closes from 15 km to 2 km	ADVISORY at 10 km, WARNING at 5 km
critical_conjunction	Three-object conjunction, TCA in 12 min	CRITICAL alert fires

# Run a specific scenario
python simulation/run.py --scenario critical_conjunction --frames 200

# Save frames to disk
python simulation/run.py --scenario close_approach --frames 100 --output-dir /tmp/frames

# Push to Redis (requires running Redis)
python simulation/run.py --scenario safe_flyby --redis

API Endpoints

Method	Path	Description
GET	/health	Health check
POST	/frames	Submit an image frame
POST	/telemetry	Submit a telemetry packet
GET	/tracks	List active debris tracks
GET	/alerts	List recent alerts
WS	/ws/live	Live data stream (100ms updates)

Testing

# Run all tests
make test

# Run with coverage
pytest tests/ -v --tb=short

# Run specific test file
pytest tests/test_prediction.py -v

60 tests cover all modules including:

• Pydantic schema validation
• Simulation frame generation
• Image preprocessing pipeline
• Debris detection & SORT tracking
• UKF trajectory prediction
• Collision probability calculation
• Risk alert tier classification
• Integration scenarios (safe flyby, close approach, critical conjunction)

Linting

make lint
# or
ruff check . --exclude dashboard

Key Design Decisions

1. Simulation-first: The simulation layer is the ONLY component that needs replacement for real satellite data. All downstream modules consume standardized Pydantic schemas.

2. Graceful degradation: Redis, YOLOv8, and other optional dependencies have fallbacks. The system works without them for testing.

3. UKF with perturbations: The Unscented Kalman Filter includes J2 oblateness, atmospheric drag, and solar radiation pressure for realistic LEO orbital mechanics.

4. Alfriend-Akella Pc: Collision probability uses the standard 2D projection method used by US Space Command.

5. SORT tracking: Multi-object tracking requires 4+ consecutive frame confirmations before promoting a track to "active" to suppress false positives.

6. Impulsive maneuver planner: When a CRITICAL alert fires, maneuver_planner.py evaluates candidate in-track and radial burns at a configurable lead time before TCA and recommends the minimum delta-v burn that achieves the target miss distance.

7. SQLite persistence: database.py provides a lightweight persistence layer for tracks and alerts so that history survives backend restarts without requiring a separate database service.

8. 3D orbital visualisation: OrbitalView.jsx uses React Three Fiber and Three.js to render a real-time 3D scene with the satellite globe, debris objects, and predicted trajectory paths.

Dependencies

Python Backend

• fastapi + uvicorn — REST API and WebSocket server
• opencv-python-headless — Image processing
• filterpy — Kalman and Unscented Kalman filters
• numpy + scipy — Scientific computing
• astropy — Coordinate transforms and orbital mechanics
• pydantic + pydantic-settings — Data validation and schemas
• redis — Stream-based message queue
• ultralytics — YOLOv8 (optional, blob detection fallback available)

React Frontend

• react 18 + vite — UI framework and build tool
• three + @react-three/fiber + @react-three/drei — 3D orbital visualisation
• recharts — Time-series charts for Pc history
• zustand — Lightweight state management
• tailwindcss — Utility-first styling

License

MIT