These were nice early in the TensorFlow evolution, for things like Frigate...
But even CPU inference is both faster and more energy efficient with a modern Arm SBC chip, and things like the Hailo chip are way faster for similar price, if you have an M.2 slot.
I haven't seen a good USB port alternative for edge devices though.
The big problem is Google seems to have let the whole thing stagnate since like 2019. They could have some near little 5/10/20 TOPS NPUs for cheap if they had continued developing this hardware ecosystem :(
>The big problem is Google seems to have let the whole thing stagnate since like 2019.
Google's flightiness strikes again. How they expect developers (and to some degree consumers) to invest in their churning product lines is beyond me. What's the point in buying a Google product when there's a good chance Google will drop software support and any further development in 5 years or less?
At my last day job, https://github.com/google-coral/libedgetpu/issues/26 this was the last nail in the coffin that got us to move away from coral hardware. This was when we were willing to look past even the poor availability of the hardware during the peak chip shortage.
> As per https://www.tensorflow.org/guide/versions , Can we assume that libedgetpu released along with a tflite version is compatible with all the versions of tflite in the same major version?
> Hi, we can't give any guarantee that libedgetpu released along with a tflite version is compatible with all the versions of tflite in the same major version.
I'd also keep an eye on the Rubik Pi 3 as it looks specifically designed to compete with the Pi 5 + AI kit and should provide a faster, cheaper, more efficient option. They're only just starting to ship and no support in Frigate yet, so just something to consider if you're not in a hurry to build a system.
I tested frigate on this combination and I must recommend against it. The Raspberry Pi 5 lacks hardware accelerated video encode like it's predecessor the Pi 4. Due to this limitation, a Pi 5 will struggle with more than a few cameras, even with the 26 TOPS Hailo.
Frigate is surprisingly not that cpu intensive with if you have Coral.
I got repurposed HP G2 SFF desktop with old i5-6500 cpu running proxmox with bunch of VMs and LXC containers including frigate.
I am passing both coral USB through to frigate container for object detection and passing intel's gpu through for video decoding.
With 10 cameras continuously recording, corals inference cpu usage is about 12%, frigate CPU usage is about 5%, although a service called go2rtc which is used by frigate to read the cameras streams and restream them to frigate takes up about 15% of the cpu.
Overall my cpu usage fluctuates below 30% on that entire machine devoted to more than proxmox.
I did run the watt calculation on that machine and it was something reasonable, dont recall it right now
I'm running frigate on an SFF computer I got off eBay for $100 with an i7-8700. It averages around 14 watts. OpenVino for object detection and intel-qsv hardware acceleration preset.
Those OAK cameras start at $400. Frigate is free, and any of the options for hardware acceleration it's detectors will be cheaper. I don't think OAK cameras are a good choice for home surveillance.
I'm looking to upgrade my home surveillance setup, currently running Arlo Pro 2 cameras. They work fine, but I'd prefer higher resolution and to avoid saturating my internet upstream with frequent video uploads.
My needs are pretty much the same as people who buy camera bundles from big box stores. I want reliable motion detection for intruders, deliveries, and visitors, and the ability to watch videos recorded in the past couple of weeks.
Take a look at Tapo or Kasa devices (both TP-Link products).
I recently got a few to try out, and expressly chose them because they do motion and sound detection on device and also support microSD for local recording.
I've only had them a few weeks so can't speak to any slow-showing pain points, but so far both the video doorbell[1] and two inexpensive cameras I purchased[2][3] to test out have been awesome.
I set up an automation so they record continuously when I leave and when home to record on detection only (motion for all three and sound for the Kasa camera) to try to be economical on wearing out the SD cards. But for me personally knowing I'll likely have those go out on me and need replaced was an ok trade off for the convenience, and probably a wash financially because everything I wanted happens locally whereas I kept seeing them gated behind a subscription plan when looking at other options.
There's also an option in the app to enable them to speak stream locally to a NAS or NVR via RTSP if you want to do that with them. So I can eventually set that up for more reliability when the eventual SD burnout occurs, and scratch my tinkering itch with things like streaming it to Frigate and testing it out vs the native detection features, without any actual presasure to need to since it all just works as is.
The doorbell is what I was originally needing the continuous local recording and on-device object detection for. The two cameras were bonuses I threw in to grab a few of their inexpensive models to try out while I was at it. And so far for about $100 in total I've been impressed. Key word being so far – they're still recent enough I might be in my honeymoon phase with them and just don't know it yet.
Yeah the $30 2k camera you linked seems good, but I worry about local card storage, because if someone steals the camera you have no evidence who did it!
I wish the camera could stream to Frigate/whatever but stream empty delta frames when nothing is detected.
Oh nice thanks for digging that up. I currently run Frigate (along with Home Assistant) on my HP Prodesk 400 with a 8th gen Intel i5 and the Coral USB.
I wonder if it would run better with that Hailo-8 on my Pi5 with 4GB.
Finding anything that's not hundreds of dollars to go from Thunderbolt to actual PCIe slot or M.2 slot is tough though.
There are a couple solutions, but then you have to have hardware that has a Thunderbolt port as well... and those aren't everywhere, especially on cheaper computers and SBCs.
But even CPU inference is both faster and more energy efficient with a modern Arm SBC chip, and things like the Hailo chip are way faster for similar price, if you have an M.2 slot.
I haven't seen a good USB port alternative for edge devices though.
The big problem is Google seems to have let the whole thing stagnate since like 2019. They could have some near little 5/10/20 TOPS NPUs for cheap if they had continued developing this hardware ecosystem :(