False negatives with AI in BI 6

[n2qew]

I'm working on migrating a BI 5 system to BI 6. Built from absolute scratch, Win 11, 20 / 40 cores, 64 Gig of RAM Nvidia P4000 for GPU.
I'm running into an issue with motion detection / confirmation via AI. I've tried the built in, and CP, and Blue Onyx. The issue is that things of interest are being - generally - accurately identified. The problem seems to be that BI is deciding that they are static. For example, in the attached pix, there's a person walking up the sidewalk. It shows that it knows it's a person, but it false negatives. The "person" ID is flagged with the blue square stating "....objects which do not overlap an area of movement.....". What does that mean? it triggers, AI says person, they are walking, etc. For testing, I've got it doing pre 2 + 10 post with a .333 analyze rate. Alot of those pics show person. You'll note the "eval version " burned in because I'm not ready to move to it until I can get the motion trigger / AI Ironed out. The old version 5 works fine - even when pointing to the Blue Onyx running on the new box for it's AI. So, something is different in 6 that I'm missing.

BTW... If anyone is looking at Blue Onyx, beware of the latest couple of versions as they set off about 20 Virus alerts in Virustotal.

 

[jrbeddow]

First thing to try is to lengthen the times between each analysis to at least 500ms or preferably 750ms (I would even experiment with 1 second), especially given that your average analysis time is running 337ms (potential for backlog of analyses). Also, no person is going to cross this scene in less than 1/3 second, and more importantly the AI needs to see a difference in each analyzed "frame/image", so better to lengthen that a bit. Head on approach like this is always harder to get right compared to scenes where people move across the field of view.

 

[n2qew]

First thing to try is to lengthen the times between each analysis to at least 500ms or preferably 750ms (I would even experiment with 1 second), especially given that your average analysis time is running 337ms (potential for backlog of analyses). Also, no person is going to cross this scene in less than 1/3 second, and more importantly the AI needs to see a difference in each analyzed "frame/image", so better to lengthen that a bit. Head on approach like this is always harder to get right compared to scenes where people move across the field of view.

Thats probably the key to it. Default was .750 1 pre and 2 post. It's not explained in the docs as to how the AI motion detect works. I guess that the "motion" is pix to pix and not 1st pix to last pix however many that is. This definitly acts differently than BI 5, and there's more options. Any opinions as to what to set the "Object tracking: Scan at least" to? default is 10, Is longer better, or is default?

I've got 18 local cameras on the new box so far, and 9 more remote cameras will be on it as there are on the BI 5 system. The new box with BI 6 on Win 11 CPU hovers around 17%. The GPU is fairly beefy, and turns the analysis around quickly, and as you can see, it's not getting backed up.

 

[jrbeddow]

I have my cameras setup with 3-4 pre-trigger images and 6-8 post-trigger images and that generally works quite well, but I can't claim that those are optimum for every situation, you might need to experiment. Definitely keep the box checked for reanalysis on continued motion triggers.
Sadly, the object tracking module remains a mystery box for me: I haven't seen the utility in tinkering with it, and I still get occasional static cars labeled as moving objects.