Otherwise pretty happy.
[1] https://store.ui.com/us/en/category/all-cameras-nvrs/product...
It has a very nice integration with homeassistant.
Iam not sure but I think so
For basic needs go2rtc [0] or MediaMTX [1] can be enough. But once you need some form of intelligence on top AFAIK unfortunately there is no unixiy tool that can take a stream and easily define and apply a model on it. You will have to code something in python.
I ain't running it on a modern CPU though, so I'm happy with the Coral.
I'm explicitly leaving out the Coral TPU, since it's been reported that the newer Intel CPUs (Core Ultra) seem to provide the same performance with it's iGPU.
Tensorflow for the object detection doesn't do any OCR thus written instructions dont work. However, according to the website the system has a limited list of objects it detects. So maybe disguising yourself as a walking tree might prevent detection.
"These are not the detections you are looking for."
They have a two-stage approach, first motion detection with - I think - OpenCV and then afterwards object detection of zones of interest with different object detection models, depending on your hardware.
It supports Coral TPU, Halio Accelerator and most GPUs. I think AMD is still the worst, since ROCm is not available on iGPUs.
Afterwards, they provide/support models like edgedet (Coral), YOLO-NAS, YOLO, D-Fine or RF-DETR.
They also offer paid access to a specially trained version of YOLO-NAS where you can also train your own images.
[0] https://en.wikipedia.org/wiki/A_Scanner_Darkly [1] https://medium.com/data-science/avoiding-detection-with-adve...
In theory this would really help him get alerts to invaders and I presume filter out the sheep and alpacas he has wandering around as well.
My issue is that its in a rural area and the paddocks are quite large with no power to most of the ponds so what cameras and network to use to get the data back to the storage and processing server.
Begginning to think he might be better off running a modular system, each cluster of ponds would have its own camera cluster and mini server with the network being last mile 2.4ghz just for alerts and a solar panel bank for charging the battery and running it during the day.
What would I get away with here? N100 mini device? processing maybe 6 cameras?
If you want to start just remember to avoid h.265 cameras so you don't need to transcode since few clients and browsers support it.
Running a mix of Ubiquti/TP-Link VIGI+TAPO/Reolink. I'm running everything in containers and everything works perfect!
For on-camera AI, I'm aware of OpenMV https://openmv.io/ and their recently-kickstarted N6 & AE3 https://www.kickstarter.com/projects/openmv/openmv-n6-and-ae...
But for surveillence, it's usually the sensor/camera quality that is the most important. I've struggled hard to find an affordable IP camera that can actually handle both shutter speed + quality in order to for example read license plates.
Disclaimer: I didn’t try it yet but the last rabbit hole regarding OpenVINO comparisons looked too good to be true and it seems Frigate supports it too. Win-win.
Eufy Security started showing advertisements about their new products whenever I tap on a motion detected notification. They prioritize their ads over your own security which is ridiculous. Not just that, some of their clips stored in their cloud storage would never open despite the fact I used to pay them my membership fees every month. They were also caught storing passwords and other security credentials in plain text. Thanks to them, they were the primary motivation for me to move away from using those proprietary platforms and look for something self-hosted.
I got Frigate running on my old hardware with hardware acceleration enabled via RX 550 GPU and detection is always under one second. I wrote a small app that uses Frigate API to grab screenshots and send me notifications via Telegram and Pushover. It's been very self-sustainable for two years now. I only had to restart the service two times in all of this time. I am also using some tunneling from my VPS into the locally hosted Frigate running on my home server and it's just been flawless. Thanks to this amazing project.
My stack was sanic, opencv, and I was randomly experimenting with the onnx speech models. Detection even on an rpi5 can be cut down to 1.3s on average - though constantly drawing bounding boxes (and lord help me, multiple faces) made it bad.
xiconfjs•3h ago
P.S.: I‘m also supporting them with a yearly? subsciption to train the „A.I.“ model against false positives I provide which increased the accuracy even more.
m463•2h ago
not waking you, but it is cool to have a collection of animal photos. Sort of amazing there's a hidden world.
xiconfjs•2h ago
danparsonson•2h ago
wiseowise•44m ago
You can’t drop something like that without uploading it to YouTube right now.
danparsonson•9m ago
alias_neo•2h ago
Been running about 2-3 years, was mostly fine before but now I get constant false positives from the children's garden toys, scooter left in the garden, pirate flag waving etc.
I don't submit false positives for privacy reasons but I'm looking at trainingy own model. I've got years worth of positives/negatives to train on.
sugarpimpdorsey•1h ago
In so many words if you expect to use the Coral boards you are stuck on EOL versions of Debian/Ubuntu - which have terribly old video drivers and missing kernel GPU support. There's a good chance your modern GPU - even well-supported Intel ones - won't work.
Imagine buying new hardware in 2025 whose software still required Windows 7.
Cyph0n•1h ago
smokel•1h ago
Cyph0n•44m ago
Also, what kind of “security exploits” would an outdated Python result in if the Python interpreter itself isn’t serving a network port or accepting arbitrary user input in general?
I assume Frigate itself isn’t running the web app on the same Python version - it’s likely just the Coral SDK that requires an outdated Python version.
dns_snek•39m ago
You can also transcode the video before feeding it to any outdated software and run it in a VM if you're paranoid.
zeroflow•26m ago
1. It supports the developers(s) 2. The price can be directly attributed to cost for training 3. You can keep the models you trained during your subscription indefinately
That's pretty much the opposite to AgentDVR. I don't need hosted services for remote access or push notifications - I can do that myself. But if I want to abide the license terms, I need to purchase a monthly subscription for remote access over my own VPN.