Turning a Standard CCTV into a Smart One with Hermes Agent 🎥

Published on
·
Time to read
3 min read ☕
Post category
aiopencvllmautomationcctv

I just turned my standard CCTV into a smarter one using Hermes Agent. The real moat of OpenClaw, Hermes, and similar agents is that they can power anything reachable on your local network, as long as those devices expose APIs.

In my case, a basic Ezviz IP camera on my home WiFi is connected to Hermes through an RTSP stream.

The idea is simple: combine static image metadata with AI judgment for multi-layer detection. Or, in less fancy words: save LLM tokens whenever deterministic checks are already enough 🙈.

Surveillance monitor regions of interest for gate and carport detection

The monitored frame is split into regions of interest: one area watches the gate bars, while another watches the carport surface.

Layer 1: Gate Open or Closed Detection

I use OpenCV Sobel edge analysis on a specific region of interest around the gate.

When the vertical metal bars are visible, the edge ratio spikes. In my setup, a ratio above 1.8 means the gate is closed. When the gate slides open, the ratio drops below 1.0 because the road texture or darkness replaces the visible bars.

Layer 2: Carport Occupancy

For the carport area, I measure brightness mean and connected-component blob ratio.

A dark mass covering the concrete means my car is already parked inside. In practice, brightness below 100 means occupied, while brightness at or above 100 means the carport is empty.

Layer 3: Smart AI Judgment

AI judgment only runs when OpenCV confirms that:

  1. the gate is open, and
  2. the carport is empty.

At that point, I invoke the GLM-4.6V vision model to check whether my car is parked outside behind the gate, based on its color and appearance.

This is the layer where static metadata alone is insufficient. The nice part is that it works without training on tons of car images.

The Result

The final flow becomes a context-aware greeting and alerting system.

True positive: the gate is open, the carport is empty, and the vision model confirms my car is waiting outside.

If the carport is empty and no car is detected outside, the gate was probably opened by someone with unauthorized access. Hermes blasts the infamous "faaahh" sound through the CCTV speaker and sends me a WhatsApp notification.

True negative: the deterministic checks and AI judgment agree that this is not the welcome-home scenario.

But if my car is detected outside the gate, it plays "Hala Madrid y Nada Más", my favorite club anthem, as a welcome-home greeting.

This is basic, but very convenient to monitor your "dapur MBG" employees' workload, no? 🗿✌️

Jokes aside, imagine running a local vision model with unlimited tokens judging what it sees in real time. You could monitor everything the AI can see.