I have been running Frigate for quite a while now. The Google Coral TPU has long been the “gold standard” for edge processing object detection. It’s efficient, cheap, and handles MobileNet models like a champ. But as my camera count grew and my patience for false positives wore thin, I knew I needed more horsepower.
I recently pulled the trigger on the Hailo-8 AI Acceleration Module, and the results have been night and day.
The “MobileNet” Ceiling
The Google Coral is great, but it’s largely optimized for MobileNet. While fast, MobileNet can be “jittery” with detections. To get higher accuracy, you need to run YOLO (You Only Look Once) models. Trying to run a modern YOLO model on a Coral just isn’t practical, although there is a quantized version coming with version 17.
The Hailo-8 boasts 26 TOPS, significantly outperforming the Coral’s 4 TOPS. This extra headroom allowed me to switch my primary detector to YOLOv9.
Speed Meets Accuracy
Since the switch, two things have happened:
- False Positives are Gone: YOLOv9 is significantly more “intelligent” than the lightweight models I was running before. It handles occlusions and difficult angles with ease. If Frigate says there is a person in my driveway, there is actually a person there. Frigate separates what it detects into alerts and detections, the detections used to be filled with false positives. Now they only show detections I don’t have alerts set for.
- Incredible Throughput: Even with the heavier model, my Inference Speed is sitting at a crisp 16ms on 27 cameras. For a model as complex as YOLOv9, that is blistering fast.
Is it worth it?
If you are running a handful of cameras and MobileNet works for you, the Coral is still a fine choice. But if you want to eliminate false triggers and move to state-of-the-art detection models like YOLOv9, the Hailo-8 is the single best upgrade you can make for your Frigate server. One important detail though, I did notice a drop in night detection quality until I switched to the Frigate+ trained version of the YOLO model.
