Frigate with Hailo for Object Detection on a Raspberry Pi

Key Takeaways

• Hailo‑8L runs Frigate detection at ~12 ms latency
• PCIe fix requires force_desc_page_size=4096 modprobe
• Hailo AI HAT+ costs around $70, cheaper than Coral
• Low CPU usage enables multiple cameras on Raspberry Pi
• Supports both HAT+ and M.2 versions for SBCs

Summary

Jeff Geerling details how to pair Frigate NVR software with Hailo‑8 or Hailo‑8L AI coprocessors on a Raspberry Pi 5 or CM5. He outlines driver installation, Frigate configuration, and a PCIe driver tweak to resolve a max_desc_page_size error. After the fix, the Hailo detector runs at roughly 12 ms per inference with minimal CPU load, even across two cameras. The solution offers a cheaper, lower‑power alternative to the legacy Google Coral TPU for edge video analytics.

Pulse Analysis

Edge AI is reshaping how small‑scale video surveillance systems process data, and the Raspberry Pi remains a popular hardware foundation due to its affordability and community support. While Google’s Coral TPU once dominated the niche, its discontinuation left a gap that Hailo’s AI coprocessors are now filling. Hailo‑8 and the lower‑cost Hailo‑8L integrate via PCIe or dedicated HAT+ modules, delivering up to several TOPS of compute in a form factor that matches the Pi’s power envelope. This shift enables developers to retain the Pi’s low‑power advantage while gaining inference speeds previously reserved for more expensive platforms.

Implementing Hailo within Frigate involves three core steps: compiling the Hailo‑8 driver from source, configuring the object detector in Frigate’s YAML file, and addressing a common PCIe descriptor size mismatch. The latter requires creating a modprobe configuration that forces the descriptor page size to 4096 bytes, then reloading the driver. Once resolved, users report inference times near 12 ms per frame and CPU utilization that barely registers, even when monitoring multiple camera streams. These performance figures translate to smoother video recording, faster alert generation, and the ability to scale camera counts without upgrading the host hardware.

The broader market implication is clear: affordable, high‑performance edge AI is becoming accessible to hobbyists and small businesses alike. By leveraging Hailo’s cost‑effective modules, organizations can replace legacy Coral TPUs, reduce hardware spend, and future‑proof deployments with a platform that supports both HAT+ and M.2 form factors. As edge computing demand grows, solutions that combine low power draw, rapid inference, and flexible integration will likely set the standard for next‑generation NVR and smart‑city applications.