Computer vision on autonomous 4G cameras (OFITECO)
Full-stack service to operate autonomous 4G cameras (solar, no fixed IP) and analyse events with computer vision over distributed critical infrastructure.
The problem
Monitoring installations in remote locations with no fixed network, no public IP and no power supply. The answer: solar-powered 4G cameras (4K, PTZ, battery) that send JPEG photos every 10 minutes instead of continuous video — to preserve battery and mobile data — plus a backend that operates and analyses them remotely.
What I built
A full-stack web service to operate a network of autonomous solar 4G cameras (PTZ, 4K, battery) from the browser and run optional computer vision on the captures, integrated with the xdata digital-twin platform.
- Control of cameras with no fixed IP via MQTT + a P2P bridge over 4G: on-demand capture, PTZ (355° pan / 140° tilt), presets with auto-photo, real-time battery and remote siren / reboot / wake-up.
- MQTT snapshots instead of RTSP: ~90% less bandwidth over 4G, with the camera sleeping between captures to preserve battery.
- Per-camera auto-capture (from 5 min to 24 h), managed server-side (one asyncio task per camera) and persistent across restarts.
- Dual storage: JPEG photo + metadata CSV (battery, status, location) per capture, with gallery and lightbox.
- Computer vision (optional module, CPU-only): a YOLOv8n (ONNX) + zero-shot CLIP + change detection (SSIM) pipeline that detects intrusions (person / vehicle / animal), classifies weather (fog / snow / night) and flags changes against the previous photo, with ray-casting exclusion zones to cut false positives.
Why it matters
- No civil works, no cabling (autonomous solar camera + 4G SIM): adding a new camera means registering it in the web panel.
- AI on CPU (Intel i5, no GPU): zero inference cost, ~700 ms/photo — plenty for 10-minute capture intervals.
- Photos and CSVs live only on the client’s host (bind mounts), never in a third-party cloud.
Technical metrics (verifiable)
- 30 REST endpoints in 7 groups (Swagger / OpenAPI) · 3 Docker services (MQTT broker · P2P bridge · FastAPI app) with a multi-stage Dockerfile.
- 63 tests (pytest) · 8-check quality gate before every push (tests, build, syntax, secrets, tech debt, WCAG AA, security, docs).
- Vision pipeline at ~700 ms/photo on CPU with no GPU: YOLOv8n 58 ms + CLIP ~600 ms + change detection ~14 ms.
Stack
Frontend: Svelte 5 · Vite 7 · TypeScript. Backend: Python 3.12 · FastAPI · aiomqtt · Pydantic. Vision: YOLOv8n (ONNX) · CLIP ViT-B/32 · SSIM. Comms: MQTT (Mosquitto) · P2P bridge over 4G. Deployment: Docker Compose.
Professional project at OFITECO (VINCI Group) · private repository.
Other work with a related technical focus:
xdata-bim — Infrastructure Digital Twin (OFITECO)
Web-based digital-twin platform (BIM + GIS + IoT) for monitoring critical infrastructure in real time. Built solo, from architecture to production.
Pulso — Structural Health Monitoring (OFITECO)
Web platform that auscultates bridges and buildings via vibrations: automatic operational modal analysis (OMA), tracking and damage alerts. In production.
PlantsCare IoT
End-to-end agricultural IoT, from hardware to cloud: custom ESP32 nodes, MQTT/CoAP, InfluxDB, EfficientNet-B0 image classification and AWS deployment.