Let’s be real: many of us steer clear of smart home gadgets—especially security cameras—because of data privacy nightmares. Incidents like the Ring debacle proved exactly why handing over your private surveillance footage to third-party corporate servers is a bad idea. Fortunately, there’s a way to build a sophisticated, privacy-respecting monitoring system yourself. I added an official AI Kit to my Raspberry Pi security rig, and the results blew past my expectations.

It started with a dusty Raspberry Pi 5. I decided to pair it with a couple of old CCTV cameras and a DIY RTSP camera spun up from a retired smartphone. With Frigate and the official AI Kit (the OG module, not the + variant), I turned that Pi into a rock-solid NVR hub with genuine object detection chops.

Why Frigate Was the Obvious Choice

There’s no shortage of Network Video Recorder (NVR) apps in the self-hosted ecosystem, but I went with Frigate for my local security stack. First off, it’s incredibly lightweight, so it doesn’t choke the Pi’s resources when juggling multiple camera streams. Recent updates also made it way more accessible, adding tweaks, toggles, and critical settings right into its minimalist Web UI.

But the killer feature? First-class support for AI accelerators—specifically the Hailo-8L module inside the AI Kit I used. That is non-negotiable for efficient object detection. Frigate remains one of the best NVR tools out there purely on the merit of its comprehensive monitoring feature set.

For the OS, I stuck with Raspberry Pi OS Lite. It’s lean, leaves plenty of headroom, and—crucially—avoids the driver hell you often hit with more barebones distros like Armbian, especially when you’re trying to bind an AI accelerator to a single-board computer.

Spinning Up the Base NVR with Docker

Step one: getting Docker running on the Pi. I used the dockerd-rootless-setuptool.sh install script to ensure containers run rootless—no privileged mode required. From there, I pulled the deploy command straight from the official Frigate docs.

Frigate’s Web UI lets you add new cameras on the fly without manually editing config files. I hooked up two cams—one 720p, one 1080p—to the Pi NVR. After a quick RTSP handshake, both feeds popped up on the Dashboard, ready for live monitoring, snapshot triggers, and video recording tasks.

Injecting AI Muscle with the Hailo-8L HAT

At this stage, the Frigate setup was solid for basic monitoring. But the second I flipped the switch on object detection, latency spiked past 70ms. The Pi’s ARM CPU simply doesn’t have the horsepower for AI inference. Enter the Hailo-8L AI Kit, which offloads 100% of the detection workload.

Mounting the Hailo-8L HAT was smoother than expected. Since I was on the latest Raspberry Pi OS, I first had to evict the pre-packaged Hailo driver that ships with the distro:

bash
sudo modprobe -r hailo_pci

Then I renamed that driver to stop the Pi from accidentally grabbing it for Frigate inference duties. Next, I grabbed the Frigate-specific user installation script:

bash
wget https://raw.githubusercontent.com/blakeblackshear/frigate/dev/docker/hailo8l/user_installation.sh

Made it executable (chmod +x user_installation.sh), ran it, then hopped into the Frigate Web UI Config Editor to inject the required arguments (mind your indentation!) before restarting the container.

Initially, Frigate refused to play nice with the accelerator. Checking the logs revealed the culprit: the desc_page_size variable. A quick re-read of the Frigate docs and this fix did the trick:

bash
echo 'options hailo_pci force_desc_page_size=4096' | sudo tee /etc/modprobe.d/hailo_pci.conf

After that, Frigate finally recognized the AI accelerator.

Tuning Object Detection in Frigate Config

With the AI HAT online, I circled back to the Config Editor. I added detect and record blocks with the necessary snippets for each camera.

To keep the system from choking, I dialed the detection FPS down to 8. I also defined a track > object block targeting animal, person, and vehicle. Even with an accelerator, resources are finite—optimization isn't optional.

To maximize throughput, I set up Zones to restrict object detection to specific regions of interest, rather than wasting cycles analyzing the entire camera feed.

After adding extra cams and refining masks, I stress-tested detection. Person detection is near-flawless, though it occasionally hallucinates a garden hose as a snake. As long as you don't get greedy with frame rates, this setup runs flawlessly across all my security cams.

The best part? Unlike cloud NVRs, every byte of data stays on my local network. Total privacy, zero subscription fees. I even bridged Frigate with Home Assistant for rich, actionable notifications using community blueprints.

Reference

https://www.xda-developers.com/i-added-an-ai-kit-to-my-raspberry-pi-security-camera/