Part 3: Exploring Custom AI Models in Frigate

Lucas Calje

8 min read

·

Oct 31, 2025

--

2

Listen

Share

Press enter or click to view image in full size

Now that Frigate and Home Assistant are up and running, it’s time to explore how different AI models perform. These tests were conducted on a Raspberry Pi 5 (16 GB RAM) running Debian Lite and a Hailo 8L accelerator.

This article shows how to switch between object detection models, verify that they are working correctly, and measure their performance in Frigate. It focuses on understanding how the system behaves with different detector configurations rather than comparing detection accuracy, which depends on the specific scenes and objects you monitor.

Frigate uses different default object detection models depending on the active detector backend, such as:

detectors:
  hailo:
    type: hailo8l

When running on the CPU, Frigate relies on an efficient SSD MobileNet V1 TensorFlow Lite model located at /cpu_model.tflite inside the container. The model operates on 300 × 300 input frames, making it lightweight and compatible with any processor, though its accuracy is limited for smaller or more distant objects.

When a Hailo-8L accelerator is available, Frigate automatically switches to a hardware-optimized YOLOv6n model with a 640 × 640 input, compiled into a .hef file (located in /config/model_cache/ inside the container).
This setup provides significantly higher…