Blue Iris is amazing, but I've never had it work right with AI for longer than 24 hours.
Anyone having luck with continuous ai analysis as trigger?
It must be me. Please help.
I live in the PNW with a lot of rain, so when I use motion detection triggers outside at night, I get an amazing number of alerts... with nothing but rain. I tried object detection, but when I fix settings to not trigger from raindrops, a person or car can drive by with no trigger.
Naturally, I turned to the AI features. That was a long path. It may have taken some years off my hairline, but I can now say I am confident operating CodeProject.AI, adding custom models, and modifying config files, but it took a while
Initially, I went for the motion trigger plus AI confirmation. However, I would get blind spots... if there was motion but nothing detected, the trigger would continue to the minimum trigger duration and if something happened then, it would miss it. Then I increased the number of post-trigger images to analyze and the interval between images so that it would last the full length of my minimum trigger duration. That worked, but I was sending 40 post-trigger images every time motion was sensed. My machine could keep up, but it would just stop working. I would see processing in CPAI, but no triggers would be happening on Blue Iris. That was last week. I bought the Blue Iris upgrade as it said AI improvements, but I still haven't got anything stable.
The next thing I tried was the newer "continuous AI analysis" in the trigger tab. The default setting was completely ineffective. Scanning the images every 120 seconds or whatever leaves a lot of room for unrecorded activity. I changed all the settings to one second and now I actually have continuous AI trigger. But even with that, the kicker is if it detects something I don't get the person logo in my alerts, like I did with AI confirmation.
It still misses things all the time, and when it is working it detects things like a bush and says it's a person.
I then turned to AI confirmation for my AI triggers to get the person/car/dog logos in my alerts, which works sometimes. But it kept detecting bushes or my leaf blower as a person. I turned on the static object option and because I am using AI as a trigger at a 1-second analysis interval with a 1-second make-or-break time for sensitive and fast AI triggering, I can't get the static object detection to work. It rejects most things because the make/break time is 1 second. So everything detected is considered static, I think.
I am on a dual socket Proxmox machine with an A1 CPU and all its RAM passed to the Win 10 Pro VM, which is hosting Blue Iris and CPAI. I found some forum somewhere where someone had to increase their RX and TX buffer sizes in both Proxmox and the Windows VM. I have had better luck with YOLOv8, but MikeLud1 only has a few models for it that I could find so I'm using a few I found. I tried just default object detection, which works great... but nothing has been stable. Every time it works right I feel like the 200+ hours + $$$$ spent on this wasn't a waste. Something within 24 hours happens, and I leave my house and drive away. I don't get the push notification I am supposed to get. I fire up my VPN and log in, and the camera wasn't even alerted.
And honestly, even when it does work, it's nothing like I've read and heard about, like very quick notifications. I only have 6 cameras. I will walk from one camera to another, and the cameras won't trigger chronologically. Or it will catch the very end of the action, even with a big pre-recording option set.
Please, someone help. I am doing something wrong. I have to be missing something. I've basically wire-sharked my network and none of my cameras have any crazy packet loss or anything. One camera I have won't go past a .25 i-interval, and I followed the documentation and increased buffer, and it almost does better than my cameras with the ideal 15/1.00 fps.
CPai was giving crazy errors and I fixed it and I can almost make custom models. I think I am doing something wrong in blueiris.
I really want to use ai triggers. I am considering to going back to retry motion trigger plus ai confirmation for the duration of the trigger, but that is a lot of images send during the confirmation.
What do I do?
Anyone having luck with continuous ai analysis as trigger?
It must be me. Please help.
I live in the PNW with a lot of rain, so when I use motion detection triggers outside at night, I get an amazing number of alerts... with nothing but rain. I tried object detection, but when I fix settings to not trigger from raindrops, a person or car can drive by with no trigger.
Naturally, I turned to the AI features. That was a long path. It may have taken some years off my hairline, but I can now say I am confident operating CodeProject.AI, adding custom models, and modifying config files, but it took a while
Initially, I went for the motion trigger plus AI confirmation. However, I would get blind spots... if there was motion but nothing detected, the trigger would continue to the minimum trigger duration and if something happened then, it would miss it. Then I increased the number of post-trigger images to analyze and the interval between images so that it would last the full length of my minimum trigger duration. That worked, but I was sending 40 post-trigger images every time motion was sensed. My machine could keep up, but it would just stop working. I would see processing in CPAI, but no triggers would be happening on Blue Iris. That was last week. I bought the Blue Iris upgrade as it said AI improvements, but I still haven't got anything stable.
The next thing I tried was the newer "continuous AI analysis" in the trigger tab. The default setting was completely ineffective. Scanning the images every 120 seconds or whatever leaves a lot of room for unrecorded activity. I changed all the settings to one second and now I actually have continuous AI trigger. But even with that, the kicker is if it detects something I don't get the person logo in my alerts, like I did with AI confirmation.
It still misses things all the time, and when it is working it detects things like a bush and says it's a person.
I then turned to AI confirmation for my AI triggers to get the person/car/dog logos in my alerts, which works sometimes. But it kept detecting bushes or my leaf blower as a person. I turned on the static object option and because I am using AI as a trigger at a 1-second analysis interval with a 1-second make-or-break time for sensitive and fast AI triggering, I can't get the static object detection to work. It rejects most things because the make/break time is 1 second. So everything detected is considered static, I think.
I am on a dual socket Proxmox machine with an A1 CPU and all its RAM passed to the Win 10 Pro VM, which is hosting Blue Iris and CPAI. I found some forum somewhere where someone had to increase their RX and TX buffer sizes in both Proxmox and the Windows VM. I have had better luck with YOLOv8, but MikeLud1 only has a few models for it that I could find so I'm using a few I found. I tried just default object detection, which works great... but nothing has been stable. Every time it works right I feel like the 200+ hours + $$$$ spent on this wasn't a waste. Something within 24 hours happens, and I leave my house and drive away. I don't get the push notification I am supposed to get. I fire up my VPN and log in, and the camera wasn't even alerted.
And honestly, even when it does work, it's nothing like I've read and heard about, like very quick notifications. I only have 6 cameras. I will walk from one camera to another, and the cameras won't trigger chronologically. Or it will catch the very end of the action, even with a big pre-recording option set.
Please, someone help. I am doing something wrong. I have to be missing something. I've basically wire-sharked my network and none of my cameras have any crazy packet loss or anything. One camera I have won't go past a .25 i-interval, and I followed the documentation and increased buffer, and it almost does better than my cameras with the ideal 15/1.00 fps.
CPai was giving crazy errors and I fixed it and I can almost make custom models. I think I am doing something wrong in blueiris.
I really want to use ai triggers. I am considering to going back to retry motion trigger plus ai confirmation for the duration of the trigger, but that is a lot of images send during the confirmation.
What do I do?
