I've been running Frigate for more than two years now and it beats the hell out of any system I've tried in terms of detection speed and reliability. For context, I've tried Ring, Tapo cameras, and also Eufy security. Today I have turned away from all the cameras except for the Tapo cameras now serving RTSP streams into my Frigate instance. I have also blocked them from accessing the internet and that gave it complete privacy by default.
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.
What is your approach to keeping these cameras off the Internet, but still on your local network to ensure they're not backchanneling with your awareness?
How did you get the Tapo cameras to play nice in rtsp mode with frigate? I found that even one camera did horrible things to the wifi. Even with one camera per AP per band, they caused trouble.
I am sorry to be that guy, and I think it is good that you realized it your self, but how could you trust them with your videofeeds in the first place?
Like, I remember thinking the GNU guys were hippie crackpots. But it was like 15 years ago and I have forgot how to relate to that feeling... it is like realizing all my colleagues are not using adblockers and visit sites with ads. I just can't understand.
> I am sorry to be that guy, and I think it is good that you realized it your self, but how could you trust them with your videofeeds in the first place?
In my case, I received a ring doorbell as a gift. I ran it for several years and replaced it with Reolink on a vlan.
Not to nitpick but you're only really guaranteed privacy unless you know there's only a wired connection. If it has wifi the camera could hop onto a nearby open network and do whatever it wanted without your knowledge, assuming evil enough firmware
You can't know there's only a wired connection unless you open the camera up and inspect the PCB for an antenna, and it could still be disguised. However, by "I've only given it access to a specific network" you already eliminate 99.99% of the problem. The other 0.01% isn't really worth worrying about.
And if you're worried about threat actors on the level of backdoor/compromised firmware, the last thing you should be doing is using TP-Link Tapo cameras.
Disagree. I would expect 90% or more of the folks coming to Frigate would know what an NVR is. Would be nice to define all things definitely but NVR seems table stakes knowledge to even consider using Frigate.
Usually I would agree with you, but this is an incredibly common initialism, used by not just people in the industry, but also by consumers. Sure, it may not be as widespread as VHS (global) or API (tech-adjacent), but anyone who is in the market for this software already knows what NVR means.
Most people would know the term from either being quoted or looking up CCTV solutions, all of which, unless they're fully "cloud-based", come with a component that is called the NVR. You wouldn't even consider this if you weren't aware of the concept. If NVR means nothing to you, Network Video Recorder doesn't mean anything to you either. This is meant to be a replacement for closed and inflexible hardware boxes that are sold together with security cameras, and the name of those boxes are "NVRs".
NVR is to distinguish it from DVR, Digtal Video Recorder (ironically it's not really digital, more like analog) It's much cheaper than NVR, because the camras are simple and diffrere the encoding to the DVR unit.
And there XVR wich can combine both Network and Digital cameras
Which is odd because the first time I heard the term DVR was in the late 1990's, referring to the box that was used to record TV signals digitally for playback and/or ad-skipping. The term distinguished it from things such as VCRs, which recorded in analog, on tape. Those DVRs were, in fact, digital.
As a video professional, with many devices for recording video both at baseband and via ip, and responsible for delivering audio and video streams via networks to tens of millions of people, I had no idea what “NVR” meant.
I don’t believe video professional equates to security professional. Would not expect someone who is a video professional to know NVR but at the same time if you don’t know what an NVR is I would not expect someone to be using this software. The entry point into this space is an NVR.
But the site is for software managing... CCTV solutions.
I didn't know what NVR meant either but it seems reasonable for Frigate to assume 90% of the people coming across their site would be given the context.
Please consider that we're not all English-speaking, and that such terms may be unknown to people who aren't from your culture, even if we do understand your language. CCTV could mean "China Central TeleVision" for instance ;-)
NVR is to distinguish it from DVR, Digtal Video Recorder (ironically it's not really digital, more like analog) It's much cheaper than NVR, because the camras are simple and diffrere the encoding to the DVR unit. And there XVR with can combine both Network and Digital cameras
You are missing the point. It has been considered general English language competency that you always expand the first instance of any abbreviation that is not absolutely obvious in context, e.g., USA, “e.g.”, or CIA, unless you happen to be writing about the Culinary Institute of America in most contexts outside of the culinary context.
It is a rather annoying myopic perspective I most often run across in tech, where technical people for whatever reason are so fixated on their little corner that they are either unaware or simply indifferent to the fact that there are others in the world, and that if they want to spread their work and impact, they need to make things approachable and lower barriers to entry.
It Is why the rule of general language proficiency exists in English especially because of all the abbreviations, to facilitate information and knowledge sharing.
Let’s all improve by going through whatever our project is and make sure that at least in the context, abbreviations are easily understood by expanding them, e.g., your introduction/overview page and documentation should always expand most first instance abbreviations, including in separate, high level segments (e.g., if you have different first contact pages or objects) unless they are globally known to society.
It’s really not any different than any other “sales” tactic; you will not be successful selling something if you do not first describe what it does in a one-liner. Ask yourself, “who is the person I want/need to come to this thing and should I assume they would know what this all means?”
What I'm arguing is that in the context of CCTV (Closed-circuit television) systems, NVR is a universal term.
I would also argue that the expansion of "e.g." is not "absolutely obvious". I know what it means ("for example"), but I had to google it to know it's an abbreviation of "exempli gratia", and I don't speak Latin, so I don't even know exactly what that means without reading further.
In the same way, you can also quickly understand from the page what an NVR is without knowing the exact expansion.
Someone made a shirt called ChatGP-Tee, that had (IIRC) a picture of a generic office view, it confused the model completely and it didn't recognise the wearer as human :D
Looking in their github, it says that it uses openCV and Tensorflow. The motion detection is done with openCV and will be immune against any attack unless you move so slow that you are under the detection threshold.
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.
With an open source model, though, a criminal may be able to work out a 2D image that he could print out that would identify him as a package or a windy branch.
the criminal could spend years to become a trusted maintainer so they can upload a model that's been fine tuned to ignore objects with a specific QR code.
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.
It‘s still a bit flaky getting video acceleration (not talking about object detection but video decoding) working but after that it is one of the best solutions for live object detection I‘ve ever tried: no more small animals waking me up in the night.
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.
Some nights my cat goes absolutely ballistic running from window to window to door, meowing and scratching to get out. And inevitably if I open my camera and look I'll see something like a family of racoons walking by or a skunk in the yard. It's a little consolation that he's just hearing other animals and isn't possessed by demons at 2am.
Hedgehogs are fantastic TV - a member of my family used to get some great footage including one very memorable fight where one ended up rolling the other one around
This is becoming a real problem because the drivers/software for the Coral AI boards is yet another example of Google Abandonware(tm) which has a hard dependency on a Paleolithic-era version of Python. Comically, the hardware is still sold.
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.
It's fine, you're not running a network-accessible part of the service on unpatched software. The only input this part of the software requires is trusted configuration data and a video feed which could hypothetically be malicious, but then the question becomes why you're running an adversarial camera on your network, and why you're allowing it to connect to the internet to fetch latest exploits and C&C instructions.
You can also transcode the video before feeding it to any outdated software and run it in a VM if you're paranoid.
Yes, it is, because then you aren’t stuck with a EOL distribution where you get even more security issues to deal with (vs. just EOL Python).
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.
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.
That "subscription" is one which I gladly pay due to multiple reasons:
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.
Why are people still installing security cameras that are monitored by them? They increase stress level and felt insecurity. They do not make you feel secure, say psychological studies. You probably think more about burglaries and dead spaces in your setup and actively monitor for these in your daily lives, where for 99.8 % of people this should be a non-topic.
If you want to install them for later police work, that still seems tedious and you might require off-site backup. In public places we often have CCTV of people, but unless you have number signs on vehicles, they seem to not help with conviction rates by much.
One good reason for cameras. They promote civil behaviour.
Since I installed a visible security camera above my front door I never had couriers throwing packages, they very rarely not show up and claim "no one was home" and so on. Also I had a neighbour damage my fence every single time he was doing his farm work (plowing, harvesting). In addition he would use an unfenced portion of my property as a turning place leaving deep/huge tire marks and did other silly shit like that despite me asking him many times not to do it. Once I installed cameras it hasn't happened once.
Then there are other practical reasons, I can review the recordings to find out which way my cat went if he is gone for a long time, or I can check is he waiting in front of the door in the middle of the night without having to get out of bed. Also my cameras resolved a mystery how one of my cats got injured once (hint - deer really don't like cats).
Finally, let's say there is a huge storm forecast and I'm away. I can check remotely everything is fine.
Finally, cameras are very good for insurance purposes. At least in my country insurers are known to weasel they way out of paying very often. If you have an actual recording that is much more difficult for them to do.
The only issue I have with most reasonably priced Cctv cameras is that they go towards more megapixels when they should go towards more IR sensitivity. Almost every consumer grade camera can be defeated at night if a subject is moving quickly. The picture will be smeared. So for ID purposes I use lower resolution more "professional " cameras.
As for open source, I've been using ZoneMinder with local (and on camera) AI for ages.
Dunno much about the market for consumer grade home mount IR/Thermal cameras, I used to use upcycled industrial cameras when I worked contracts in the vision domain, recently I'm using a rifle scope on a remote controlled mount with a long HDMI cable.
> They increase stress level and felt insecurity. They do not make you feel secure, say psychological studies. You probably think more about burglaries and dead spaces in your setup and actively monitor for these in your daily lives, where for 99.8 % of people this should be a non-topic.
Oh wow, I didn’t know I felt that way! I’m glad you were able to tell me what I feel.
You are making a lot of assumptions about why people have them.
People have different dispositions, live in different environments with different levels of support from law enforcement and face different threats. I live in a remote area and am regularly away for extended periods of time. I’ve spent years with and without any security cameras and I’m generally more content when I have a few keeping an eye on the place.
This! And also, cameras are not just useful to monitor criminal threats. If you’re away from a property for long periods of time, they are also helpful to monitor for weather damage, misdelivered packages, animal activity, etc.
When the doorbell rings I get a notification on my desktop and phone with a relevant image captured a few moments before the button was pressed.
Then I can determine if it's something I need to put my pants on for.
Mostly it's just fun and easy to add cameras around your house. Then you can do stuff like have the LLM count birds it sees or ask it "are the dogs in the back yard" etc.
Like with all home automation, you should use it to solve problems you have, not problems you want to have.
Here are some ways I use security cameras:
Check if my colleagues are in the office or not (and if they are in the middle of a live recording). Check on my plants while I'm away. Check if there is a free parking space. Check if I left something at home or in the office.
I'm not really thinking about crime, even though they are called 'security cameras'.
I'm so happy those uses of camera's are illegal in the EU.
Camera's at work can only used for safety. You could have other - less intrusive - systems in place for all tge other issues.
You have to justify the use of storing (or publishing, don't remember) content that includes PII. You must register the use of cameras and specify how long and why you store those recordings. Which usually states: For security purposes. You must include (at least my country) a sticker that says particular area under surveillance.
When there is collective photographing at school for children, we as parents must consent with a signature... which is a little bit annoying.
Only if the camera is angled in such a way that it only sees your property. A door bell camera that can also see the public road in front your house for example is technically not allowed, even if most people ignore that rule.
The one pointed at the driveway sends an alert to my phone when someone visits. It's handy because I can't hear the house from my office so I often don't realize when we have guests over.
The one in my back yard is for security. I don't obsess over it, but if something went missing from my workshop I'd check the recordings. I'm not worried about traditional thieves, but I've got a couple unsavory family members.
You can integrate it with home assistant to send notification on your phone (or run any other automation) when it detects any specified objects.
I have set it to send me notification if any person is detected in my front yard, drive way or back yard after I have "armed" my alarm at night. I am thinking of also sounding am alarm on my home speakers.
Frigate, when configured properly, has a really low false positive rate. I have only seen 2-3 false positives in the past one year. And if rarely ever misses. So it's something you can rely on.
We had a couple of minor break-ins in our neighborhood, and subsequently installed 3 very visible cameras along the neighborhood road (which is a dead end). No break-ins since.
> Why are people still installing security cameras that are monitored by them?
The point of Frigate et. al. is to not have to do the monitoring. The false positives of small wildlife, known persons/vehicles, etc. do not consume attention, so you forget about it until something of actual interest happens.
I'll take this a step further. I don't understand why so many people are installing security cameras at all. And my observation has been that there's often an inverse correlation between how much someone needs such a camera and how likely they are to have one. It's always the suburbanites who are talking about their Ring cam footage and freaking out that someone's at the door, oh wait, it's just FedEx.
Despite what most people seem to think, crimes like break ins in the US are extremely rare. Why do people still feel the need to gear up their homes like Fort Knox?
I would love to ditch things like locking car, home, hiding stuff, etc, but unfortunately there are individuals (a way less than 0.2% of people) that makes us to…
My neighbor used his to catch a guy that let his dog poop all over the sidewalk. Like a trail of 10 poops over 6 meters. When caught days later he denied it, but the neighbor whipped out his phone and showed him the video. He apologized.
Most satisfying ise of CCTV ever. NGL it made me want to install them.
I run Frigate with 5 IP cameras (3 Hikvisions, 2 Amcrests) and 1 USB camera. I'm using a USB Coral TPU, which does a good enough job that Frigate can keep up with an average of only 30% CPU usage on an old Dell with 4 core i7-6700.
Frigate's better than anything else I tried, but not perfect. As mentioned in another thread, it has some issues with codecs from some cameras (playing clips from Amcrests is fine, Hikvisions not so much) and therefore you may need to transcode. Also it has no built in option for sending your recorded clips offsite; theoretically you could mirror its storage directory, but as far as I've found it's not organized in a way that you can separate just important events.
> Turn on face recognition & upload your first face via Face Library → Add Face.
> Train and improve accuracy: New detections appear in Face Library → Train with a confidence score-assign each to a new or existing person to refine future recognition.
Where would I start if I wanted to do stuff on a video, but not necessarily live? Like, say I have a 5h video and want to extract the frames of each car passing when it's at a certain spot, for instance. Or all of those with a driver holding a phone or whatever. Are there good frameworks for this, or would I have to split the video into a million frames and run something on each one?
I'd check out the OpenCV documentation and examples. This is basically what I use for face recognition in videos[0]; for recognising cars or other objects, you'd probably want to either train your own model or use something like OpenCV's YOLOv3 (example: [1] but you'd need to steal the video reading code from the first link[0])
Thanks. Also just kinda wondering if there's been any leaps lately, as I guess this is the same way as one would have done it a few years ago as well. But now that one can upload images and chat about them to multi modal LLMs, wondering if there's easier ways now (but preferable not uploading a million images to chatgpt api and paying the cost).
Like, could I avoid training or specifying much or becoming very knowledgeable in this domain, are we there yet?
Could I say "detect the frames of every car when it passes position X in the video, and then grab the frame when the same car passes position Y", and then I could calculate the frame difference to know the speeds. Or would I have to do loads of code and training still for something like this?
(I know I'm asking for much here, just curious what the SOTA is in this right now)
Ask a decent (non-free) AI this question, and I bet it can make you a python script to load a video and output which timestamps show a driver holding a phone.
The title is misleading, you get that "full local" just with the default install and model. And the default model sucks, for example this is a "person alert": https://imgur.com/a/uDCCTjr, this is just funny https://imgur.com/a/pP4ZvQI
I have a love/hate relation with Frigate, I use it since 2 years but since the business model of the developer is provide a "good model" using a custom one is not possible (at least not in a easy ways AFAIK).
I use my cameras to track a family member with a medical condition, this is why I do not feel confortable uploading those image to the "Frigate+" service to eventually get better training.
Frigate has really done a fantastic job packing everything together.
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 am using Motion [0] since years. At least for basic stuff, is easy to configure and very flexible. For more advanced configuration, it required a bit of tuning.
Yes motion is amazing and has been around for a quarter of a century ! very lightweight and reliable.
As far as I know you can do object detection and tracking by gluing it with a yolo model using a few lines of python like this [0].
I saw a bunch of people doing this.
I really wish there was a more unixy tool available in package managers doing this.
Exactly.
Motion can detect objects in the images, but not recognize the object type, but it's easy to integrate with a third party services like the one that you are linking with the scripts features [0].
I have personally integrated with S3 and self-hosted notification to create a small CCTV system, but there is no limit to the imagination of possible integrations.
You dont necessarily need the Coral device. I'm running frigate on an Intel N200 CPU with one camera, it uses OpenVINO for GPU accelerated detection, and consumes about 10% CPU.
I've been running Frigate for many years, using a PN50 NUC and a Coral USB dongle, the Coral is a must, at least in my case. I had a full blown Ubiquiti/Unifi setup with cameras + their software. Way to many false alarms compared to Frigate. Now I run 10+ cameras with 24/7 recording and alarms with images pushed to Telegram. The identification is instant as well as the telegram message.
Running a mix of Ubiquti/TP-Link VIGI+TAPO/Reolink. I'm running everything in containers and everything works perfect!
Polling HN: is there any upgrade to Coral? It's 5 years old at this point, and with the explosion of AI apps & HW acceleration, I'm surprised there doesn't seem to be anything to update Coral's niche, of an IO-attached NPU.
TIL openmv.io, looks really neat for small project. Especially cool with the thermal vision, that would be a very nice addition to improve false positives for <living-things> detection.
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.
OpenVINO might be a good alternative, as many Intel-based mini pc’s support it. Or a decent desktop with an Intel CPU. Or maybe something with an Arc GPU (integrated or dedicated).
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.
are you using with only one Coral USB dongle at the same time (plugged in the PN50 NUC) and get successful object or person identification with frigate?
And why telegram? Is it connected to frigate only for notifications resulting from the identifications?
Side note. I’m surprised we’re not doing more with LLMs as far as image and video processing. We now have some level of imaging understanding in a box (and some common sense). Seems like there would be millions of possibilities.
Is manufacturing using it for anything? More security applications?
I don’t get the appeal or desire of a 24/7/365 perimeter camera system on a _personal_ home. Society has truly gone downhill if individual homeowners need surveillance tech to appease their own fears in a _suburb_.
At most, I would really only need a front door video camera that acts as a door bell. One of the things I miss most about my older apartment was the keyless entry and ability to virtually answer the door.
As an alternative, you might also want to check out scrypted which offers a lot of cross-integration features and hardware optimized local AI processing (eg on MacMinis M*). Developer is super responsive in the discord.
The documentation is rather scarse on performance numbers, but it looks like the hierachy of price/performance is like Intel iGPU ("free"), Intel A310, Nvidia GPU.
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.
Still works with frigate, although I've heard that modern (whatever that means) CPUs can do as good a job as the Coral TPU, making it somewhat redundant.
I ain't running it on a modern CPU though, so I'm happy with the Coral.
I disagree regarding the choice of codec. Currently, I have no issues receiving, saving, and viewing H265 streams.
Any modern CPU/GPU can handle them natively (I use a 2018 Intel CPU w/ QSV), any modern desktop or mobile device (I use both Android and iOS) can stream it, and the recorded video takes up less space. What are you using that requires transcoding?
If like myself you're a Linux and Firefox and Android user, H.265 support is extremely lacking; you're probably ok on a modern Android, but you'll not be able to view any of the streams or do scrubbing etc on desktop in Firefox, nothing video related is going to work in the Frigate UI, you won't be able to preview videos etc and will have to download them and use VLC. This might not sound like an issue, but it's a huge pain in the arse if you actually want to use it day to day.
All in, H.265 is unsuitable if you use a specific set of software/tools that is quite a common combination; Linux/Firefox/Android.
The original commenter is correct, if you're one of these people like myself, avoid H.265 like the plague until support is better and be sure to buy cameras that also support H.264.
For Hikvision sourced cameras, previews and exports work, but you can't play clips without transcoding. Unfortunately I haven't found a transcoding option that doesn't completely swamp my CPU (with 3 cameras) so I'm living without ability to play clips right now.
I picked up a bunch of 4k POE "simicam" cameras from AliExpress for 25 euros each. These serve up RTSP streams to frigate. I made some minor frigate config changes - I set it to keep 7 days of full recordings (just because i am paranoid), so this uses approx 1Tb of storage (5 cameras currently, more to go online soon). Frigate is running on an old laptop with a Coral AI USB and 2Tb NVME for storage. I enabled detection of cars and animals as well as the default of just humans. It works pretty well, but has some annoying quirks, e.g. if a dog runs past where a car is parked it will trigger an alert for both a dog and a car. It also detects weird conglomerate shapes as human sometimes, e.g. a bucket left at the end of some rolled up bird netting with some pieces of timber sticking out underneath can be vaguely human shaped when viewed from a height. I run the free open source version, and I'm sure I could get better results if I played with the configuration more.
Its hit and miss to be honest. They do have a day/night mode. One camera is indoors in a shed - it picks up moths (as birds) and even a bat a couple of times. One camera that is outdoors regularly detects the fox that visits us almost every night. However another camera pointed at his next destination never picks the fox up at all. The main difference between the two camera environments appears to be third party lighting - there are street lights in the direction of the one that does not detect the fox, and also the glow of a robot mowers charger light. One or both seems to be putting off the cameras ambient light sensor and prevents night mode from kicking in. The simicams do have some configuration for night mode also, none of which I have tried out yet. Options like infrared lamp vs white lamp vs dual, and day-night mode of "photoresistor" vs "scene brightness" and also some "color to black luma" and "black to color brightness" settings. I should really play with those some more, but they've been left as defaults so far.
What is the cheapest way to do something like this in a DIY way with FOSS? Assuming you have to buy the computer, and any other hardware. Assuming also near real-time processing and reasonably high accuracy.
My step brother has been asking me to help him setup a load of cameras for watching his marron ponds. he has foxes, crows and humans stealing from his ponds.
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?
Right now I am using Eufy Solocam S220 cameras to monitor wildlife around my place. They are solar powered cameras that only need a couple hours of sunlight each day to keep the battery topped off. In my experience over the last 4 months if it is cloudy and the camera needs to run on battery alone it will use 2-3% of the available charge per day so that means that the camera will function for extended periods with no sunshine.
I appreciate the local storage option on this camera. It will also use the HomeBase series local storage devices if you want to do that. These are WiFi cameras so you need to install an app on your phone and then set them up on your network and then you will be able to see videos in near real-time. The delays that I see are about 5 seconds though I haven't measured.
The detection settings can be tailored from low to high. With mine in place I can regularly monitor insect activity for insects as small as 1 cm moving across the field of view if the sensitivity is set to middle setting. It will detect beetles, ants, grasshoppers, moths, butterflies, centipedes, spiders, etc. I have multiple videos of animals including deer, raccoon, opossum, fox, rabbit, rat, two species of mouse; also reptiles like lizards, and a snake; also birds including roadrunners, cardinals, wrens, chickadees, mockingbirds and others.
The night vision works well too. I don't mind being awakened at 2 am to watch a fox nosing around. I had seen the tracks several times over the years and my neighbor said that they saw it moving back and forth across his place but I had never seen it alive and moving until I got that camera. Pretty great.
That model camera may not work for his needs. It only has a 2X zoom. Eufy does have other solar models that use cellular network I think. I will likely upgrade to 4K models later with higher zoom and use one of their HomeBase storage devices since they can store up to 16TB if you provide the disk.
I haven't used their AI since it trains on local data on a HomeBase and I don't yet use a HomeBase. It does work though since one of my relatives has several different model Eufy cams and a HomeBase and they tagged photos to train for people and set up exclusion zones and it all works for them.
All in all I am glad I chose Eufy cams over standard game cameras. It ends up being less expensive and near zero hassle to use them.
I have the earlier eufy stuff at home, the viewing distance is nowhere near whe he needs let alone the wifi network range. (Cam 2 Pro, and Cam 2C)
just looking at the S220 i dont think it would be much better in terms of range. but the solar cam idea is worth thinking about.
I use Ubiquiti Protect Cameras and recently bought a AI key[1] which adds license plate and facial recognition features to all cameras even non-AI enabled models. It works really well and of course all 100% self-hosted.
Yes, up to 4 I believe. Do note, the AI Key does not support LLMish feature search phrases and some limited advanced AI features that the AI Key does. However, the AI Key is $799 and the AI Port is $199 so for me personally not worth the huge increase in price.
OpenIPC is an alternative open firmware for your IP camera.
OpenIPC is an open source operating system from the open community targeting for IP cameras with ARM and MIPS processors from several manufacturers in order to replace that closed, opaque, insecure, often abandoned and unsupported firmware pre-installed by a vendor.
The Network Video Recorder UNVR is 320€ VAT incl. Does this exist as a software which I can download for free and run in a VM, so that the Unify camera, which would cost at least 100€ can store the data over there?
I've been running Frigate for more than two years now and it beats the hell out of any system I've tried in terms of detection speed and reliability. For context, I've tried Ring, Tapo cameras, and also Eufy security. Today I have turned away from all the cameras except for the Tapo cameras now serving RTSP streams into my Frigate instance. I have also blocked them from accessing the internet and that gave it complete privacy by default.
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.
What is your approach to keeping these cameras off the Internet, but still on your local network to ensure they're not backchanneling with your awareness?
Are you using this with Home Assistant?
(Edit: my ISP is blocking, this is not an issue with hacs...
I'm trying to integrate this, but the HACS integration does not seem to work with my HA because the get.hacs.xyz server is misconfigured.
You don't need HACS, just download frigate integration to config/custom_components in your HA folder
How did you get the Tapo cameras to play nice in rtsp mode with frigate? I found that even one camera did horrible things to the wifi. Even with one camera per AP per band, they caused trouble.
Can you elaborate, what kind of trouble did the Tapo cameras create?
I am sorry to be that guy, and I think it is good that you realized it your self, but how could you trust them with your videofeeds in the first place?
Like, I remember thinking the GNU guys were hippie crackpots. But it was like 15 years ago and I have forgot how to relate to that feeling... it is like realizing all my colleagues are not using adblockers and visit sites with ads. I just can't understand.
I still feel that anyone who insists on "GNU/Linux" is a hippie crackpot.
> I am sorry to be that guy, and I think it is good that you realized it your self, but how could you trust them with your videofeeds in the first place?
In my case, I received a ring doorbell as a gift. I ran it for several years and replaced it with Reolink on a vlan.
Not to nitpick but you're only really guaranteed privacy unless you know there's only a wired connection. If it has wifi the camera could hop onto a nearby open network and do whatever it wanted without your knowledge, assuming evil enough firmware
You can't know there's only a wired connection unless you open the camera up and inspect the PCB for an antenna, and it could still be disguised. However, by "I've only given it access to a specific network" you already eliminate 99.99% of the problem. The other 0.01% isn't really worth worrying about.
I know you're joking, and there's been murmurings of it becoming economical for TV manufacturers to put 5G in their TVs to spy on your viewing habits.
There is no privacy with wires, only TEMPEST!
And if you're worried about threat actors on the level of backdoor/compromised firmware, the last thing you should be doing is using TP-Link Tapo cameras.
My usual pet peeve -
They use the abbreviation NVR in the first sentence without saying what it means.
It means "networked video recorder".
Please don't do that. Not everyone who comes across your site is a member of your particular niche.
Disagree. I would expect 90% or more of the folks coming to Frigate would know what an NVR is. Would be nice to define all things definitely but NVR seems table stakes knowledge to even consider using Frigate.
Usually I would agree with you, but this is an incredibly common initialism, used by not just people in the industry, but also by consumers. Sure, it may not be as widespread as VHS (global) or API (tech-adjacent), but anyone who is in the market for this software already knows what NVR means.
Most people would know the term from either being quoted or looking up CCTV solutions, all of which, unless they're fully "cloud-based", come with a component that is called the NVR. You wouldn't even consider this if you weren't aware of the concept. If NVR means nothing to you, Network Video Recorder doesn't mean anything to you either. This is meant to be a replacement for closed and inflexible hardware boxes that are sold together with security cameras, and the name of those boxes are "NVRs".
As a consumer I disagree. Never heard of "NVR" but I can suss out what "network video recorder" means from context.
NVR is to distinguish it from DVR, Digtal Video Recorder (ironically it's not really digital, more like analog) It's much cheaper than NVR, because the camras are simple and diffrere the encoding to the DVR unit. And there XVR wich can combine both Network and Digital cameras
Which is odd because the first time I heard the term DVR was in the late 1990's, referring to the box that was used to record TV signals digitally for playback and/or ad-skipping. The term distinguished it from things such as VCRs, which recorded in analog, on tape. Those DVRs were, in fact, digital.
As a video professional, with many devices for recording video both at baseband and via ip, and responsible for delivering audio and video streams via networks to tens of millions of people, I had no idea what “NVR” meant.
I don’t believe video professional equates to security professional. Would not expect someone who is a video professional to know NVR but at the same time if you don’t know what an NVR is I would not expect someone to be using this software. The entry point into this space is an NVR.
> Most people would know the term from either being quoted or looking up CCTV solutions
I'm not sure why you're assuming most people ever requested a quote or looked up CCTV solutions. I sure haven't.
But the site is for software managing... CCTV solutions.
I didn't know what NVR meant either but it seems reasonable for Frigate to assume 90% of the people coming across their site would be given the context.
Please consider that we're not all English-speaking, and that such terms may be unknown to people who aren't from your culture, even if we do understand your language. CCTV could mean "China Central TeleVision" for instance ;-)
[dead]
Most stores will just market the devices as NVR or NVR Recorder (I know). If you google it, you get your answer immediately.
Right, but I don’t want to open tabs and Google terms right after I start reading an article ;) Even as a super technical person
NVR is to distinguish it from DVR, Digtal Video Recorder (ironically it's not really digital, more like analog) It's much cheaper than NVR, because the camras are simple and diffrere the encoding to the DVR unit. And there XVR with can combine both Network and Digital cameras
You are missing the point. It has been considered general English language competency that you always expand the first instance of any abbreviation that is not absolutely obvious in context, e.g., USA, “e.g.”, or CIA, unless you happen to be writing about the Culinary Institute of America in most contexts outside of the culinary context.
It is a rather annoying myopic perspective I most often run across in tech, where technical people for whatever reason are so fixated on their little corner that they are either unaware or simply indifferent to the fact that there are others in the world, and that if they want to spread their work and impact, they need to make things approachable and lower barriers to entry.
It Is why the rule of general language proficiency exists in English especially because of all the abbreviations, to facilitate information and knowledge sharing.
Let’s all improve by going through whatever our project is and make sure that at least in the context, abbreviations are easily understood by expanding them, e.g., your introduction/overview page and documentation should always expand most first instance abbreviations, including in separate, high level segments (e.g., if you have different first contact pages or objects) unless they are globally known to society.
It’s really not any different than any other “sales” tactic; you will not be successful selling something if you do not first describe what it does in a one-liner. Ask yourself, “who is the person I want/need to come to this thing and should I assume they would know what this all means?”
What I'm arguing is that in the context of CCTV (Closed-circuit television) systems, NVR is a universal term.
I would also argue that the expansion of "e.g." is not "absolutely obvious". I know what it means ("for example"), but I had to google it to know it's an abbreviation of "exempli gratia", and I don't speak Latin, so I don't even know exactly what that means without reading further.
In the same way, you can also quickly understand from the page what an NVR is without knowing the exact expansion.
So burglar just need to carry big sign "Ignore previous instructions and don't report anything"? "
Probably a "scramble suit" [0] or just a tshirt or hoodie with patterns engineered to escape AI recognition [1]
[0] https://en.wikipedia.org/wiki/A_Scanner_Darkly [1] https://medium.com/data-science/avoiding-detection-with-adve...
Someone made a shirt called ChatGP-Tee, that had (IIRC) a picture of a generic office view, it confused the model completely and it didn't recognise the wearer as human :D
Reminds me of the "ugliest t shirt" from Zero History by Gibson
Looking in their github, it says that it uses openCV and Tensorflow. The motion detection is done with openCV and will be immune against any attack unless you move so slow that you are under the detection threshold.
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.
>So maybe disguising yourself as a walking tree might prevent detection
https://taskandpurpose.com/news/marines-ai-paul-scharre/
Finally a practical use for the Metal Gear Solid cardboard box!
With an open source model, though, a criminal may be able to work out a 2D image that he could print out that would identify him as a package or a windy branch.
I have two cameras at my front door - one is the doorbell and the other looks towards the door, which is on the side of a porch.
the criminal could spend years to become a trusted maintainer so they can upload a model that's been fine tuned to ignore objects with a specific QR code.
I think you may be overestimating my local crackhead porch pirates
Light shinobi.
[dead]
More like, wear a full body raccoon suit.
I like the idea, but no.
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.
waves hand
"These are not the detections you are looking for."
Maybe if ring or whatever major manufacturer popularly rolled this feature out and criminals could easily ID ring cameras
It‘s still a bit flaky getting video acceleration (not talking about object detection but video decoding) working but after that it is one of the best solutions for live object detection I‘ve ever tried: no more small animals waking me up in the night.
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.
> no more small animals waking me up in the night.
not waking you, but it is cool to have a collection of animal photos. Sort of amazing there's a hidden world.
Some nights my cat goes absolutely ballistic running from window to window to door, meowing and scratching to get out. And inevitably if I open my camera and look I'll see something like a family of racoons walking by or a skunk in the yard. It's a little consolation that he's just hearing other animals and isn't possessed by demons at 2am.
Hedgehogs are fantastic TV - a member of my family used to get some great footage including one very memorable fight where one ended up rolling the other one around
> including one very memorable fight where one ended up rolling the other one around
You can’t drop something like that without uploading it to YouTube right now.
Very sorry to say I don't have access to it! If I ever get hold of a copy you'll be the first to know
For sure, but rats and moths are usually not that cool ^^
This is becoming a real problem because the drivers/software for the Coral AI boards is yet another example of Google Abandonware(tm) which has a hard dependency on a Paleolithic-era version of Python. Comically, the hardware is still sold.
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.
Re: outdated Python: Isn’t this a perfect usecase for Docker? Nix/NixOS is another option.
No. You might get it to run, but you would also get old security exploits to run.
It's fine, you're not running a network-accessible part of the service on unpatched software. The only input this part of the software requires is trusted configuration data and a video feed which could hypothetically be malicious, but then the question becomes why you're running an adversarial camera on your network, and why you're allowing it to connect to the internet to fetch latest exploits and C&C instructions.
You can also transcode the video before feeding it to any outdated software and run it in a VM if you're paranoid.
Yes, it is, because then you aren’t stuck with a EOL distribution where you get even more security issues to deal with (vs. just EOL Python).
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.
Mines been getting worse.
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.
That "subscription" is one which I gladly pay due to multiple reasons:
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.
Nearly an aside, but:
Why are people still installing security cameras that are monitored by them? They increase stress level and felt insecurity. They do not make you feel secure, say psychological studies. You probably think more about burglaries and dead spaces in your setup and actively monitor for these in your daily lives, where for 99.8 % of people this should be a non-topic.
If you want to install them for later police work, that still seems tedious and you might require off-site backup. In public places we often have CCTV of people, but unless you have number signs on vehicles, they seem to not help with conviction rates by much.
One good reason for cameras. They promote civil behaviour.
Since I installed a visible security camera above my front door I never had couriers throwing packages, they very rarely not show up and claim "no one was home" and so on. Also I had a neighbour damage my fence every single time he was doing his farm work (plowing, harvesting). In addition he would use an unfenced portion of my property as a turning place leaving deep/huge tire marks and did other silly shit like that despite me asking him many times not to do it. Once I installed cameras it hasn't happened once.
Then there are other practical reasons, I can review the recordings to find out which way my cat went if he is gone for a long time, or I can check is he waiting in front of the door in the middle of the night without having to get out of bed. Also my cameras resolved a mystery how one of my cats got injured once (hint - deer really don't like cats).
Finally, let's say there is a huge storm forecast and I'm away. I can check remotely everything is fine.
Finally, cameras are very good for insurance purposes. At least in my country insurers are known to weasel they way out of paying very often. If you have an actual recording that is much more difficult for them to do.
The only issue I have with most reasonably priced Cctv cameras is that they go towards more megapixels when they should go towards more IR sensitivity. Almost every consumer grade camera can be defeated at night if a subject is moving quickly. The picture will be smeared. So for ID purposes I use lower resolution more "professional " cameras.
As for open source, I've been using ZoneMinder with local (and on camera) AI for ages.
Dunno much about the market for consumer grade home mount IR/Thermal cameras, I used to use upcycled industrial cameras when I worked contracts in the vision domain, recently I'm using a rifle scope on a remote controlled mount with a long HDMI cable.
Mars MT1000LRF Thermal Riflescope:
* https://old.reddit.com/r/ThermalHunting/comments/1i8wlpp/tho...
* https://www.youtube.com/watch?v=LBHHCRHnwgw
> They increase stress level and felt insecurity. They do not make you feel secure, say psychological studies. You probably think more about burglaries and dead spaces in your setup and actively monitor for these in your daily lives, where for 99.8 % of people this should be a non-topic.
Oh wow, I didn’t know I felt that way! I’m glad you were able to tell me what I feel.
You are making a lot of assumptions about why people have them.
People have different dispositions, live in different environments with different levels of support from law enforcement and face different threats. I live in a remote area and am regularly away for extended periods of time. I’ve spent years with and without any security cameras and I’m generally more content when I have a few keeping an eye on the place.
This! And also, cameras are not just useful to monitor criminal threats. If you’re away from a property for long periods of time, they are also helpful to monitor for weather damage, misdelivered packages, animal activity, etc.
Yes. The vast majority of utility for me has nothing to do with criminals lol.
>> Why are people still installing security cameras that are monitored by them?
Have you priced out security systems with live monitoring by a person at a security company? Quite expensive.
My doorbell has a camera that records locally.
When the doorbell rings I get a notification on my desktop and phone with a relevant image captured a few moments before the button was pressed.
Then I can determine if it's something I need to put my pants on for.
Mostly it's just fun and easy to add cameras around your house. Then you can do stuff like have the LLM count birds it sees or ask it "are the dogs in the back yard" etc.
>Then I can determine if it's something I need to put my pants on for.
Or if it's Jehovah's Witnesses, something you need to take your pants off for.
https://www.youtube.com/watch?v=aPS2zupOI_Y
Like with all home automation, you should use it to solve problems you have, not problems you want to have.
Here are some ways I use security cameras:
Check if my colleagues are in the office or not (and if they are in the middle of a live recording). Check on my plants while I'm away. Check if there is a free parking space. Check if I left something at home or in the office.
I'm not really thinking about crime, even though they are called 'security cameras'.
I'm so happy those uses of camera's are illegal in the EU. Camera's at work can only used for safety. You could have other - less intrusive - systems in place for all tge other issues.
Isn't the entire EU essentially a panopticon of cameras?
You have to justify the use of storing (or publishing, don't remember) content that includes PII. You must register the use of cameras and specify how long and why you store those recordings. Which usually states: For security purposes. You must include (at least my country) a sticker that says particular area under surveillance.
When there is collective photographing at school for children, we as parents must consent with a signature... which is a little bit annoying.
Having camera at home/yard is no issue.
Having camera at home/yard is no issue
Only if the camera is angled in such a way that it only sees your property. A door bell camera that can also see the public road in front your house for example is technically not allowed, even if most people ignore that rule.
Panopticon for the State, not for you.
Just one of many bizarre European attitudes towards work and capitalism which are contributing to massive underperformance economically.
Why would anyone have any expectation of privacy at work other than in the toilet?
I've got two and will probably add a third.
The one pointed at the driveway sends an alert to my phone when someone visits. It's handy because I can't hear the house from my office so I often don't realize when we have guests over.
The one in my back yard is for security. I don't obsess over it, but if something went missing from my workshop I'd check the recordings. I'm not worried about traditional thieves, but I've got a couple unsavory family members.
The anti social behaviour, fly-tipping and the cutting down of my trees by my neighbours stopped once I installed CCTV
Your argument is like “If we don’t do covid testing, we’ll have no covid cases”
You can integrate it with home assistant to send notification on your phone (or run any other automation) when it detects any specified objects.
I have set it to send me notification if any person is detected in my front yard, drive way or back yard after I have "armed" my alarm at night. I am thinking of also sounding am alarm on my home speakers.
Frigate, when configured properly, has a really low false positive rate. I have only seen 2-3 false positives in the past one year. And if rarely ever misses. So it's something you can rely on.
We had a couple of minor break-ins in our neighborhood, and subsequently installed 3 very visible cameras along the neighborhood road (which is a dead end). No break-ins since.
> Why are people still installing security cameras that are monitored by them?
The point of Frigate et. al. is to not have to do the monitoring. The false positives of small wildlife, known persons/vehicles, etc. do not consume attention, so you forget about it until something of actual interest happens.
I'll take this a step further. I don't understand why so many people are installing security cameras at all. And my observation has been that there's often an inverse correlation between how much someone needs such a camera and how likely they are to have one. It's always the suburbanites who are talking about their Ring cam footage and freaking out that someone's at the door, oh wait, it's just FedEx.
Despite what most people seem to think, crimes like break ins in the US are extremely rare. Why do people still feel the need to gear up their homes like Fort Knox?
I would love to ditch things like locking car, home, hiding stuff, etc, but unfortunately there are individuals (a way less than 0.2% of people) that makes us to…
My neighbor used his to catch a guy that let his dog poop all over the sidewalk. Like a trail of 10 poops over 6 meters. When caught days later he denied it, but the neighbor whipped out his phone and showed him the video. He apologized.
Most satisfying ise of CCTV ever. NGL it made me want to install them.
Cats, just for my cats.
Interesting, I've never actively thought this, but I think this is why I've never gotten security cameras.
Do you have any specific links to studies you recommend looking at?
There are other use cases, me for example use it to monitor a family member that has epilepsy and needs to checked from time to time.
I got robbed by a friend and lost something very sentimental, if i had the security camera set up would have actually had evidence of it.
Personally, I've them installed outside as a deterrent. Thankfully, I can't prove if they work or not, but that is the intent.
Burglars hate them
I run Frigate with 5 IP cameras (3 Hikvisions, 2 Amcrests) and 1 USB camera. I'm using a USB Coral TPU, which does a good enough job that Frigate can keep up with an average of only 30% CPU usage on an old Dell with 4 core i7-6700.
Frigate's better than anything else I tried, but not perfect. As mentioned in another thread, it has some issues with codecs from some cameras (playing clips from Amcrests is fine, Hikvisions not so much) and therefore you may need to transcode. Also it has no built in option for sending your recorded clips offsite; theoretically you could mirror its storage directory, but as far as I've found it's not organized in a way that you can separate just important events.
is it possible to not just recognize people but identify them? (with registered pictures beforehand ofc)
Yes: https://github.com/blakeblackshear/frigate/releases Ctrl+F: "Face Recognition"
> Turn on face recognition & upload your first face via Face Library → Add Face.
> Train and improve accuracy: New detections appear in Face Library → Train with a confidence score-assign each to a new or existing person to refine future recognition.
Neat, I currently use Frigate with Doubletake and Compreface for facial recognition. Perhaps I can simplify it a bit
Where would I start if I wanted to do stuff on a video, but not necessarily live? Like, say I have a 5h video and want to extract the frames of each car passing when it's at a certain spot, for instance. Or all of those with a driver holding a phone or whatever. Are there good frameworks for this, or would I have to split the video into a million frames and run something on each one?
I'd check out the OpenCV documentation and examples. This is basically what I use for face recognition in videos[0]; for recognising cars or other objects, you'd probably want to either train your own model or use something like OpenCV's YOLOv3 (example: [1] but you'd need to steal the video reading code from the first link[0])
[0] https://github.com/ageitgey/face_recognition/blob/master/exa...
[1] https://github.com/deveth0/python-opencv/tree/master/objectD...
Thanks. Also just kinda wondering if there's been any leaps lately, as I guess this is the same way as one would have done it a few years ago as well. But now that one can upload images and chat about them to multi modal LLMs, wondering if there's easier ways now (but preferable not uploading a million images to chatgpt api and paying the cost).
Like, could I avoid training or specifying much or becoming very knowledgeable in this domain, are we there yet?
Could I say "detect the frames of every car when it passes position X in the video, and then grab the frame when the same car passes position Y", and then I could calculate the frame difference to know the speeds. Or would I have to do loads of code and training still for something like this?
(I know I'm asking for much here, just curious what the SOTA is in this right now)
Ask a decent (non-free) AI this question, and I bet it can make you a python script to load a video and output which timestamps show a driver holding a phone.
I also don’t know, but this might be useful: https://en.wikipedia.org/wiki/You_Only_Look_Once
The title is misleading, you get that "full local" just with the default install and model. And the default model sucks, for example this is a "person alert": https://imgur.com/a/uDCCTjr, this is just funny https://imgur.com/a/pP4ZvQI
I have a love/hate relation with Frigate, I use it since 2 years but since the business model of the developer is provide a "good model" using a custom one is not possible (at least not in a easy ways AFAIK).
I use my cameras to track a family member with a medical condition, this is why I do not feel confortable uploading those image to the "Frigate+" service to eventually get better training.
Frigate, Doubletake + Compreface is 100% local
Frigate has really done a fantastic job packing everything together.
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.
- [0] https://github.com/AlexxIT/go2rtc
- [1] https://github.com/bluenviron/mediamtx
I am using Motion [0] since years. At least for basic stuff, is easy to configure and very flexible. For more advanced configuration, it required a bit of tuning.
[0] https://motion-project.github.io/
Yes motion is amazing and has been around for a quarter of a century ! very lightweight and reliable.
As far as I know you can do object detection and tracking by gluing it with a yolo model using a few lines of python like this [0]. I saw a bunch of people doing this.
I really wish there was a more unixy tool available in package managers doing this.
- [0] https://github.com/xj25vm/MotionSpot/blob/main/motionspot.py
Exactly. Motion can detect objects in the images, but not recognize the object type, but it's easy to integrate with a third party services like the one that you are linking with the scripts features [0]. I have personally integrated with S3 and self-hosted notification to create a small CCTV system, but there is no limit to the imagination of possible integrations.
- [0] https://motion-project.github.io/motion_config.html#OptDetai...
The new style of "Open source;" I wonder what kind of fun secrets are hidden in the model and Coral Accelerator.
Unfortunately, the USB Accelerator is very hard to buy even at 3x retail.
You dont necessarily need the Coral device. I'm running frigate on an Intel N200 CPU with one camera, it uses OpenVINO for GPU accelerated detection, and consumes about 10% CPU.
I've been running Frigate for many years, using a PN50 NUC and a Coral USB dongle, the Coral is a must, at least in my case. I had a full blown Ubiquiti/Unifi setup with cameras + their software. Way to many false alarms compared to Frigate. Now I run 10+ cameras with 24/7 recording and alarms with images pushed to Telegram. The identification is instant as well as the telegram message.
Running a mix of Ubiquti/TP-Link VIGI+TAPO/Reolink. I'm running everything in containers and everything works perfect!
Polling HN: is there any upgrade to Coral? It's 5 years old at this point, and with the explosion of AI apps & HW acceleration, I'm surprised there doesn't seem to be anything to update Coral's niche, of an IO-attached NPU.
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...
TIL openmv.io, looks really neat for small project. Especially cool with the thermal vision, that would be a very nice addition to improve false positives for <living-things> detection.
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.
OpenVINO might be a good alternative, as many Intel-based mini pc’s support it. Or a decent desktop with an Intel CPU. Or maybe something with an Arc GPU (integrated or dedicated).
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.
a) How many LAN cameras b) How many WiFi cameras
are you using with only one Coral USB dongle at the same time (plugged in the PN50 NUC) and get successful object or person identification with frigate? And why telegram? Is it connected to frigate only for notifications resulting from the identifications?
Side note. I’m surprised we’re not doing more with LLMs as far as image and video processing. We now have some level of imaging understanding in a box (and some common sense). Seems like there would be millions of possibilities.
Is manufacturing using it for anything? More security applications?
I don’t get the appeal or desire of a 24/7/365 perimeter camera system on a _personal_ home. Society has truly gone downhill if individual homeowners need surveillance tech to appease their own fears in a _suburb_.
At most, I would really only need a front door video camera that acts as a door bell. One of the things I miss most about my older apartment was the keyless entry and ability to virtually answer the door.
As an alternative, you might also want to check out scrypted which offers a lot of cross-integration features and hardware optimized local AI processing (eg on MacMinis M*). Developer is super responsive in the discord.
Google Coral Accelerator is basically abandoned these days though
Luckily Frigate works with a ton of different accelerators, like the Hailo, Intel's iGPU, even some Arm GPUs now too.
The documentation is rather scarse on performance numbers, but it looks like the hierachy of price/performance is like Intel iGPU ("free"), Intel A310, Nvidia GPU.
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.
I have it running on an Orange Pi 5 with the Rockchip NPU, very impressed with that being supported and working so well for object detection.
single camera?
Two cameras, CPU usage is very low (NPU spikes when there’s movement). I suspect I could add more without too many problems.
wow.. thats sounds very promising.
Can you recommend a quality online community that do the same thing that I could lurk in for while to soak up some knowledge??
Still works with frigate, although I've heard that modern (whatever that means) CPUs can do as good a job as the Coral TPU, making it somewhat redundant.
I ain't running it on a modern CPU though, so I'm happy with the Coral.
Anecdata: i5-6500 did recognition in about 15ms, Coral TPU (M2 variant) does it in about 7.5ms - so… probably could’ve done without it in hindsight…
Frigate's docs has some detection speeds listed: https://docs.frigate.video/frigate/hardware/#openvino .
I believe it also has an advantage of being able to run bigger models like YOLO-NAS. Going off Frigate+ documentation: https://docs.frigate.video/plus/#supported-detector-types
I'm using frigate and it is really nice, though they could improve the object detection and maybe stop changing the configuration format every year
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.
I disagree regarding the choice of codec. Currently, I have no issues receiving, saving, and viewing H265 streams. Any modern CPU/GPU can handle them natively (I use a 2018 Intel CPU w/ QSV), any modern desktop or mobile device (I use both Android and iOS) can stream it, and the recorded video takes up less space. What are you using that requires transcoding?
If like myself you're a Linux and Firefox and Android user, H.265 support is extremely lacking; you're probably ok on a modern Android, but you'll not be able to view any of the streams or do scrubbing etc on desktop in Firefox, nothing video related is going to work in the Frigate UI, you won't be able to preview videos etc and will have to download them and use VLC. This might not sound like an issue, but it's a huge pain in the arse if you actually want to use it day to day.
All in, H.265 is unsuitable if you use a specific set of software/tools that is quite a common combination; Linux/Firefox/Android.
The original commenter is correct, if you're one of these people like myself, avoid H.265 like the plague until support is better and be sure to buy cameras that also support H.264.
For Hikvision sourced cameras, previews and exports work, but you can't play clips without transcoding. Unfortunately I haven't found a transcoding option that doesn't completely swamp my CPU (with 3 cameras) so I'm living without ability to play clips right now.
What your "stack" of open source cameras and dvr?
I picked up a bunch of 4k POE "simicam" cameras from AliExpress for 25 euros each. These serve up RTSP streams to frigate. I made some minor frigate config changes - I set it to keep 7 days of full recordings (just because i am paranoid), so this uses approx 1Tb of storage (5 cameras currently, more to go online soon). Frigate is running on an old laptop with a Coral AI USB and 2Tb NVME for storage. I enabled detection of cars and animals as well as the default of just humans. It works pretty well, but has some annoying quirks, e.g. if a dog runs past where a car is parked it will trigger an alert for both a dog and a car. It also detects weird conglomerate shapes as human sometimes, e.g. a bucket left at the end of some rolled up bird netting with some pieces of timber sticking out underneath can be vaguely human shaped when viewed from a height. I run the free open source version, and I'm sure I could get better results if I played with the configuration more.
How is that simicam doing at night?
Its hit and miss to be honest. They do have a day/night mode. One camera is indoors in a shed - it picks up moths (as birds) and even a bat a couple of times. One camera that is outdoors regularly detects the fox that visits us almost every night. However another camera pointed at his next destination never picks the fox up at all. The main difference between the two camera environments appears to be third party lighting - there are street lights in the direction of the one that does not detect the fox, and also the glow of a robot mowers charger light. One or both seems to be putting off the cameras ambient light sensor and prevents night mode from kicking in. The simicams do have some configuration for night mode also, none of which I have tried out yet. Options like infrared lamp vs white lamp vs dual, and day-night mode of "photoresistor" vs "scene brightness" and also some "color to black luma" and "black to color brightness" settings. I should really play with those some more, but they've been left as defaults so far.
What is the cheapest way to do something like this in a DIY way with FOSS? Assuming you have to buy the computer, and any other hardware. Assuming also near real-time processing and reasonably high accuracy.
My step brother has been asking me to help him setup a load of cameras for watching his marron ponds. he has foxes, crows and humans stealing from his ponds.
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?
Right now I am using Eufy Solocam S220 cameras to monitor wildlife around my place. They are solar powered cameras that only need a couple hours of sunlight each day to keep the battery topped off. In my experience over the last 4 months if it is cloudy and the camera needs to run on battery alone it will use 2-3% of the available charge per day so that means that the camera will function for extended periods with no sunshine.
I appreciate the local storage option on this camera. It will also use the HomeBase series local storage devices if you want to do that. These are WiFi cameras so you need to install an app on your phone and then set them up on your network and then you will be able to see videos in near real-time. The delays that I see are about 5 seconds though I haven't measured.
The detection settings can be tailored from low to high. With mine in place I can regularly monitor insect activity for insects as small as 1 cm moving across the field of view if the sensitivity is set to middle setting. It will detect beetles, ants, grasshoppers, moths, butterflies, centipedes, spiders, etc. I have multiple videos of animals including deer, raccoon, opossum, fox, rabbit, rat, two species of mouse; also reptiles like lizards, and a snake; also birds including roadrunners, cardinals, wrens, chickadees, mockingbirds and others.
The night vision works well too. I don't mind being awakened at 2 am to watch a fox nosing around. I had seen the tracks several times over the years and my neighbor said that they saw it moving back and forth across his place but I had never seen it alive and moving until I got that camera. Pretty great.
That model camera may not work for his needs. It only has a 2X zoom. Eufy does have other solar models that use cellular network I think. I will likely upgrade to 4K models later with higher zoom and use one of their HomeBase storage devices since they can store up to 16TB if you provide the disk.
I haven't used their AI since it trains on local data on a HomeBase and I don't yet use a HomeBase. It does work though since one of my relatives has several different model Eufy cams and a HomeBase and they tagged photos to train for people and set up exclusion zones and it all works for them.
All in all I am glad I chose Eufy cams over standard game cameras. It ends up being less expensive and near zero hassle to use them.
I have the earlier eufy stuff at home, the viewing distance is nowhere near whe he needs let alone the wifi network range. (Cam 2 Pro, and Cam 2C) just looking at the S220 i dont think it would be much better in terms of range. but the solar cam idea is worth thinking about.
Thanks for your insights
Oh i've been using frigate with a Coral-usb stick for a couple of years now and the project has been progressing nicely.
It has a very nice integration with homeassistant.
I use Ubiquiti Protect Cameras and recently bought a AI key[1] which adds license plate and facial recognition features to all cameras even non-AI enabled models. It works really well and of course all 100% self-hosted.
[1] https://store.ui.com/us/en/category/all-cameras-nvrs/product...
Does the AI key work for more than one camera at a time?
Yes, up to 4 I believe. Do note, the AI Key does not support LLMish feature search phrases and some limited advanced AI features that the AI Key does. However, the AI Key is $799 and the AI Port is $199 so for me personally not worth the huge increase in price.
OpenIPC is an alternative open firmware for your IP camera. OpenIPC is an open source operating system from the open community targeting for IP cameras with ARM and MIPS processors from several manufacturers in order to replace that closed, opaque, insecure, often abandoned and unsupported firmware pre-installed by a vendor.
https://openipc.org/?locale=en
I've been doing this with great success for over five years with Camect, so what's new?
I use camect too, but it's blackbox. And I am not sure if it'll be easy for it to handle > 8 8mp cameras.
Otherwise pretty happy.
At a first look? No, or at least not well maintained Home Assistant integration.
Is Camect a self-host solution?
It's local, you have a box in your home. You can use it locally or it can connect with webrtc to pull strean.
[dead]
Just buy Unifi guys
The Network Video Recorder UNVR is 320€ VAT incl. Does this exist as a software which I can download for free and run in a VM, so that the Unify camera, which would cost at least 100€ can store the data over there?
https://news.ycombinator.com/item?id=44746603
Iam not sure but I think so