Object Detection Tuning

[slicknick1337]

Hello! Long time lurker, first time poster here.

I have BI and CPAI on separate machines. CPAI is running in docker on a Linux box, GPU enabled (RTX 3050). BI runs on a separate box, GigE (obviously).

Trying to figure out why object detection isn't reliable, especially at night.

In the troubleshooting process, I have:

• Increased the number of frames taken for AI analysis
• Enabled the dat files
• Turned on main stream analysis (frames are 3840x2160)
• Increased bitrate of cameras for a better quality photo
• Bought a better GPU for faster processing (i didn't think this would fix the issue obviously, but if I was going to throw more at it, I didn't' want to slow down what actually works.)
• Adjusted AI frame capture to "prog" - progressive

Basically, what I'm finding is that faster moving objects (cars, mostly) are not detected reliably. More so at night (I get that it is more difficult, this is why I'm posting) but I have this issue with this camera, and other cameras too with cars, even during the day.

For context:

• Motion zone A is my driveway, which will send me a push notification with whatever the image contains
• Motion zone B is the street, which will simply classify the object as "car" - no notification
• This was taken shortly after midnight
• Residential street, most people drive ~25mph
• There is a street light almost directly above the car in the photo (large reason why the IR / night mode is not enabled)
• Approximately 81 feet from the camera to the street
• Approximately 104 feet from the camera to the car
• Camera is a 4k Amcrest dome mounted on the garage wall above 9-10ft off the ground.

I've analyzed many DAT files and made tweaks, with limited success. Here are some examples.

Here is a daytime example:


Here is a nighttime example:



As you can see, most of the frames are not detected at night - but this happens during the daytime too.

Here is one final example of a canceled alert where nothing is detected:



There are lots of instances where nothing is detected during the day, so this is not exclusively a nighttime issue. However, if I play back the clip with "Analyze with AI" - the car is detected.

DAT File



Same Clip w/ Analyze w/ AI:




What did I miss? Any suggestions?

Thanks in advance!

 

[wittaj]

Classic case of trying to do too much with one field of view.

That area is too small and too far out so it will be problematic. Especially since CodeProject downrezes the image. And BI motion will struggle as well at that small size. The object needs to be a lot larger in the field of view to be effective.

The analyze feature after the fact will always find more than when doing it live.

 

[slicknick1337]

Thanks for the feedback, that is helpful.

Out of curiosity, is there a way to have CodeProject analyze the full res image?

Classic case of trying to do too much with one field of view.

That area is too small and too far out so it will be problematic. Especially since CodeProject downrezes the image. And BI motion will struggle as well at that small size. The object needs to be a lot larger in the field of view to be effective.

The analyze feature after the fact will always find more than when doing it live.