Context-aware analysis
The LVLM tells normal charging apart from an attack. In the scenes tested, normal use of a charger triggered no alert.
Video surveillance that understands what it films and raises an alert when a cable is cut or a charger is damaged.
The problem
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
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
RTSP H.264 / H.265 streams, multi-camera, hot-plug.
Fine-tuned YOLO26 + BoT-SORT identity tracking with ReID.
Polygons around chargers, new person, timer, de-duplication.
Faces blurred on site before anything is sent (YuNet).
Gemini or Mistral: description, verdict and confidence score in JSON.
Snapshot flagged “ALERT” in the monitoring interface.
Why it's different
The LVLM tells normal charging apart from an attack. In the scenes tested, normal use of a charger triggered no alert.
YOLO26s trained on merged datasets remapped to 10 classes specific to charger security.
Faces are blurred on site before anything is sent. Mistral can be used as a European alternative.
Automatic detection of night / infrared mode and capture adjustment.
Model choice and analysis instructions editable from the interface, camera by camera.
Works with RTSP IP cameras. An on-site Edge mini PC (Apple Silicon or AMD64) does the work.
Roadmap
Stack