I am creating a Unifi camera setup at home, as I do have G4 and G5 cameras without advanced AI I was thinking about passing streams to Frigate and use RADXA NPU to process them, anyone achieved that?
I have seen people got good results on for example google m.2 coral - which got 4 TPU, B5+ achieves 6TPU so technically should make him even better choice.
Anyone have advices?
Yes, we did similar with the GenICam stream from industrial to RK3588 via network pipeline and did the inference.
Any special requirements, or stubborn things to fight to get it running or should be quite straight forward?
If the G4, G5 has built-in ISP then use network pipeline in inference via mpp, or rga would get the better inference results. Are these G4, G5 camera’s are conventional pixels or bayer? later would be hard as per my experience.
I am awaiting them, I won’t need AI for G4 only for G5, and once I have them I will check if they have MPP/RGA, but from what I can see they should have them. Not sure about pixel/bayer thing. I will find out once they here.
My goail is to have two things:
G4+G5 cameras - unifi protect, basic AI on all cameras, notifications alarms etc.
only G5 cameras(outside) - rock 5b+ running frigate for additional AI processing, maybe some alarms etc. nothing massive but I might want to add few bits that are missing in unifi basic AI.
If that wont work then not a big deal, but would be nice to have it 