My security cameras now describe what they see—and not a single frame leaves the house

A smart home enthusiast has detailed how he built an AI-powered surveillance system that runs entirely on local hardware, with no video frames ever leaving his home network. The setup combines TP Link security cameras with a homemade ESP32-CAM module, which he assembled after realizing he was short on ready-made cameras.

The backbone of the system is Frigate, an NVR server running on a Raspberry Pi 5 with 8 GB of RAM. To boost motion detection performance, the author paired the single-board computer with a Raspberry Pi AI HAT accelerator he had bought earlier. According to his testing, this combination handles multiple 1080p streams without breaking a sweat, significantly cutting the processing load on the CPU.

Home Assistant serves as the central hub, managing not only the cameras but also a range of other smart devices in his home, including smart plugs that monitor his home lab and DIY weather stations built from cheap ESP32 boards. The platform coordinates automation and delivers notifications when motion is detected.

The core challenge was implementing two features: automatic alerts whenever movement appears on specific cameras, and a short text summary describing what the footage contains. Rather than relying on expensive cloud-based AI models, the author chose a self-hosted approach. He integrated a local language model that analyzes captured frames and generates concise descriptions of what happened.

When motion is detected, the system extracts the relevant video segment and feeds it to the language model, which produces a human-readable summary of the scene. All of this happens on local hardware, ensuring that no sensitive footage is transmitted to external servers.

The author notes that the same result can be achieved with alternative tools, such as the MotionEye app within Home Assistant or by connecting cameras directly to the platform. His setup is aimed at users who want intelligent video surveillance without recurring subscription fees or privacy concerns. The entire project demonstrates that modern AI features can be brought to home security systems with modest hardware and a bit of configuration work.