- Screenshot 2025-02-18 at 10.05.58 AM.png (235.76 KiB) Viewed 4690 times
Static analysis cannot be this bad
Static analysis cannot be this bad
This is two nights in a row that I've had to turn off push notifications on my driveway camera. At one point last night it was firing critical notifications about once every 30 seconds firing on the car. you see below even though that car had not moved *all day*. I understand firing the first time but when it sees that the car is _not_ moving it should not fire any more. I've got continuous AI analysis turned on. Any ideas?
Re: Static analysis cannot be this bad
We really need more information for this as most of us do not quite understand it. Can you ask Support for better instructions ?
Forum Moderator.
Problem ? Ask and we will try to assist, but please check the Help file.
Problem ? Ask and we will try to assist, but please check the Help file.
Re: Static analysis cannot be this bad
How are you triggering? Using BI motion detection of external via the cam?
The only thing I can think of is that if you are using BI motion detection it for some reason thinks there has been motion, which then triggers the AI analysis which then detects the car?
For the alerts run the clip through the motion detection to see what happens?
The only thing I can think of is that if you are using BI motion detection it for some reason thinks there has been motion, which then triggers the AI analysis which then detects the car?
For the alerts run the clip through the motion detection to see what happens?
Re: Static analysis cannot be this bad
The car doesn't have to move. The image of the car just has to change. Like with headlights and such.
That is what CP-AI is for.
That is what CP-AI is for.
-
dohabandit
- Posts: 3
- Joined: Wed Jan 19, 2022 7:19 am
Re: Static analysis cannot be this bad
After tinkering with this quite a bit, here is what I have settled on:
I like having the CodeProject AI server doing some analysis, I like it gathering plates and faces, and I like the objects it detects showing up in the alerts list so I get a more contextual idea of the alert. What I don't like is false positives, and I also don't like false negatives and the recording of events stopping due to a bad AI verification.
Running AI processing continuously burns too much resources, and the confidence score level has to be set pretty high to avoid false positives. So I have turned that off. So how to get AI processing to happen still? Well in the camera settings Alert tab you have option to trigger off of AI. I uncheck that option and just trigger off of motion, audio, group, and ONVIF events.
The next option is to "confirm the alert with AI". So if you enable this, once actual motion is detected, AI processing will CONFIRM it and if the AI processing doesn't detect a car, person, bus, whatever, the alert will be untriggered, thats bad. So what to do? Well, I set the min-confidence to 1%. This will basically prevent the AI from blocking the motion detection from triggering the event. You also get the benefit of being able to have the AI zoom into the objects, save the face images to a folder, burn the markup onto the alert images, etc.
So when I watch my camera status pages I now see my GPU use at 0% unless it is processing a triggered event. This would probably clear up your issue where you are triggering on your parked car all the time. I can post some screen shots if you'd like.
I also noticed that you have a reolink camera. Are you using the autotrack PTZ camera version?
If you are, there are some good tricks for those to improve function. I have two of them watching the front of my house. I have one set to use the AutoTrack stream as main, and mainstream as sub. On the other camera I have the sub stream as main (always zoomed highdef) and autotrack stream as the sub. This gives a much better fidelity image, wider angle view, and one camera in the main camera viewer window will show the AI auto-tracking which is quite good in the ReoLink PTZ camera.
I like having the CodeProject AI server doing some analysis, I like it gathering plates and faces, and I like the objects it detects showing up in the alerts list so I get a more contextual idea of the alert. What I don't like is false positives, and I also don't like false negatives and the recording of events stopping due to a bad AI verification.
Running AI processing continuously burns too much resources, and the confidence score level has to be set pretty high to avoid false positives. So I have turned that off. So how to get AI processing to happen still? Well in the camera settings Alert tab you have option to trigger off of AI. I uncheck that option and just trigger off of motion, audio, group, and ONVIF events.
The next option is to "confirm the alert with AI". So if you enable this, once actual motion is detected, AI processing will CONFIRM it and if the AI processing doesn't detect a car, person, bus, whatever, the alert will be untriggered, thats bad. So what to do? Well, I set the min-confidence to 1%. This will basically prevent the AI from blocking the motion detection from triggering the event. You also get the benefit of being able to have the AI zoom into the objects, save the face images to a folder, burn the markup onto the alert images, etc.
So when I watch my camera status pages I now see my GPU use at 0% unless it is processing a triggered event. This would probably clear up your issue where you are triggering on your parked car all the time. I can post some screen shots if you'd like.
I also noticed that you have a reolink camera. Are you using the autotrack PTZ camera version?
If you are, there are some good tricks for those to improve function. I have two of them watching the front of my house. I have one set to use the AutoTrack stream as main, and mainstream as sub. On the other camera I have the sub stream as main (always zoomed highdef) and autotrack stream as the sub. This gives a much better fidelity image, wider angle view, and one camera in the main camera viewer window will show the AI auto-tracking which is quite good in the ReoLink PTZ camera.