← Indumia Lab

Vigie IRVE

Video surveillance that understands what it films and raises an alert when a cable is cut or a charger is damaged.

Type
POC · Vision AI (LVLM)
Status
◐ Edge prototype
Year
2026
License
Open source
For
EV charging stations
Code
Coming soon on GitHub

The problem

Cable theft and vandalised chargers, often at night.

An out-of-service charger is expensive to repair and penalises drivers. Conventional CCTV records, but nobody watches the footage live, and motion detection triggers many false alarms.

The approach

Two levels of analysis.

A mini PC connected to the existing IP cameras detects and tracks people on site. Only suspicious scenes are sent to an LVLM, which looks at the context (hood, no vehicle, cable being handled, open cabinet…) and explains its decision.

No cameras to replace.

Pipeline

How a frame is processed.

IP cameras

RTSP H.264 / H.265 streams, multi-camera, hot-plug.

Detection

Fine-tuned YOLO26 + BoT-SORT identity tracking with ReID.

Zones & triggers

Polygons around chargers, new person, timer, de-duplication.

Blurring

Faces blurred on site before anything is sent (YuNet).

LVLM analysis

Gemini or Mistral: description, verdict and confidence score in JSON.

Alert

Snapshot flagged “ALERT” in the monitoring interface.

0,89mAP50 of the fine-tuned YOLO26s model
0,90precision on 8,821 validation images
10security classes: balaclava, tool, damaged cable, vandalism…
20FPScapture on CPU, validated on a real camera

Why it's different

Fewer false alarms.

Context

Context-aware analysis

The LVLM tells normal charging apart from an attack. In the scenes tested, normal use of a charger triggered no alert.

Specialised

A vision model dedicated to EV charging

YOLO26s trained on merged datasets remapped to 10 classes specific to charger security.

Privacy

Privacy by design

Faces are blurred on site before anything is sent. Mistral can be used as a European alternative.

Night

Automatic IR mode

Automatic detection of night / infrared mode and capture adjustment.

Flexible

One prompt per camera

Model choice and analysis instructions editable from the interface, camera by camera.

Existing setup

Your cameras are enough

Works with RTSP IP cameras. An on-site Edge mini PC (Apple Silicon or AMD64) does the work.

Roadmap

The first, on-site phase is validated.

  • Multi-camera capture, YOLO detection + ReID tracking
  • Triggers, polygon zones, blurring, night mode
  • LVLM analysis validated on a real camera
  • Multi-channel alerting: email, SMS, Teams, Telegram…
  • Multi-site cloud platform and dashboard
  • On-site actuators: siren, lighting

Stack

  • Python 3.12
  • OpenCV
  • GStreamer
  • Ultralytics YOLO26
  • BoT-SORT ReID
  • YuNet
  • Gemini
  • Mistral
  • Flask
  • TimescaleDB
  • Redis
Next project
Audit Ascenseur