Maybe this forum is not even the best place to ask but I also don’t know of a better forum. The thing is that in my opinion the best home security system is a system that alerts you before the intruder has even reached your doors/windows. There are of course several ways to achieve that but the most interesting way to me seems to be a camera with AI to detect persons.
However it frustrates me that even in 2025 with all the advances in AI, it seems there’s still no camera on the market that can detect a person on a camera feed as good as a human can. And I know it’s not easy to program but there’s now AI that for example can detect skin cancer more reliable than a human being can. So is reliable person detection really that impossible?
Anyway so far I’ve tried 2 systems. One is a costly Axis system with a thermal camera and I’ve tried the way cheaper Unify G5 camera. They both seem to suffer from the same issues: they both have mistaken animals like cats for human beings. And the (way more expensive) Axis even has detected tree movement as human beings.
So I was wondering, are there cameras out there that are really good at this? What are your experiences?
When speaking security camera you should understand DORI
With DORI concept in mind I can say frigate works well for me with camera at about “Observation”’distance for people, Cars and my large dogs.
If you literally want to identify a person — capture picture of Jimmy and reliably have software say “yup, that’s Jimmy — you need to be at “Identify”. You can unreliably do it at “recognition” but expect a lot of mistakes in high traffic environment. Doorbell camera usually allow this since person is really close but if sunlight behind person or face is in shadow it will not work well.
AI cant make miracles. The better the input the better the result. High cost camera isn’t needed but you should work to achieve image that will allow you to achieve needed goal. You can send image to AI image and get good description of scene now but you must ensure you get info in scene that allow output you need.
Camera placement and lighting is first step to good design. Get that right and you will have less issues
It even gave it a higher probability of being a person when indeed a person was there.
I’m sure that there are camera’s out there that wouldn’t have made that error.
In the end you need a good fast algo + learning data all on chip for fast recognition so I’m not sure that I’d agree with your statement that a high cost camera is not needed. Then again, I’m not expert here.
So I was hoping to get some recommendations from users here.
I’m actually surprised that there don’t seem to be any websites or youtube channels dedicated to testing this. How well do the AI algo’s work for each camera? That’s what’s pretty important these days.
I’ve been googling all this a bit today and there actually seem to be 3 classes here:
AI algo on the camera itself
AI algo in the cloud
AI algo on local PC
That last option also seems interesting, haven’t really looked into this yet but this definitely looks interesting: The Ultimate Home Surveillance System – Free Local AI Person Detection – The Hook Up or this: https://frigate.video/
Then again that guy in that video boasts that the algo is 98% accurate and only 1 in 50 alarms is false, but that’s actually still too much in my book. I mean a human can easily do that WAY better, I’m sure AI should be able to be better too.
There is also ability to send the image you have above to AU engine local or remote to process image and provide description. This can be done from HA directly or using various NVR software products like frigate.
I am no expert but that’s a model problem.
Object Detection takes known images and create some behind the scenes magic that allows it to look at another image and say “yeah, I think that’s a cat”. But that same magic allows it to think it is a person. Better models and more powerful processing allows better result. Object detection lacks context. That cat looks like a standing person (2 legs, chest and arm) in the distance. It has no context to judge the objects distance and determine that is unlikely. Some AI models add context and allow a better judgement of the object.
The rotation of camera could effect result. I would generally align it so wall on left is vertical and not angled. If placed the way I describe the cat is horizontal and may not identify as person but I don’t know cuz it all magic. I do know you will get less flaring from light reflecting off wall at night and overall better image in darker areas if you reduce the flaring
Also if you don’t need to watch sky I would decrease that in image as well. I would still be above tree in background to ensure I catch person walking on rooftops.
