DeepStack How to alert on Delivery carriers

[wittaj]

You need to turn off the object travels X pixels and the resets after X% - those are used to eliminate false triggers and you have the opposite going on.

I do not have either checked for any of my cameras.

What is the makeup of your computer? 22,000ms is a long time. That could be an indicator of an overloaded machine and thus why it is missing. Most are under 200ms

 

[AlphaBlueIris]

Great question, I had Deepstack, it was giving good results but to be honest I never looed into closely to those values. I installed CodeProject but something is wrong the installation, As we speak, I'm removing codeproject and reinstall it again. My computer is a Windows 10 with 8 CPU.

 

[wittaj]

What i number and generation and how much RAM?

 

[AlphaBlueIris]

It is a VM with 8 CPU and 16 GB, I can increase the memory if needed as well as the CPU.

 

[AlphaBlueIris]

I think the reason why it is taking so long is that it is going thru all the pt. Is this a normal settings that all of you use?
This is a different camera, with this one, I'm looking specifically for people going thru the pathway.

The other camera that should get the delivery truck is currently not triggering, maybe my resolution for this year is to give up

 

[JNDATHP]

Anything over a few hundred ms suggests to me that DS is timing out because it is not working.

Before CPAI, I had DS and it would suddenly stop working and time out. Perhaps if you were to remove DS and the Nvidia drivers, etc. and do a clean install of CPAI, it will work as expected.

 

[AlphaBlueIris]

thank you all, I also think that I'm looking for too many things, I should only look for people.

 

[truglo]

Could you please confirm this is what I read from your post?

In your second uploaded image, zone A appears to be blank? I did mention a lot of details so understandable it may have been overlooked...

In your zone setup, leave zone A as is.

So again, zone A should still be a strip covering the street where the vehicles are stopping. Since you have zone A blank, you should not expect it to trigger... no need to give up for the year.

Also, I'm with wittaj WRT your AI make times being really long; 20sec is an eternity for AI. This could be due to limitations with your VM... perhaps BI is unable to make use of hardware acceleration that could speed up AI? IDK to be honest if HA is even used for AI, but I think anything more than like 1sec is going to cause problems with missing rapid fire events (like a truck passing by seconds after a car triggers it). So yes, adding items to "to confirm" could help reduce make time, but it won't help if the AI took 20sec and found nothing. Are you sure you have delivery.pt installed on your system, and cpia/bi are configured to use it? Instead of waiting forever for more deliveries, you can test if it works using any downloaded image of a delivery truck, and analyze it in the cpai explorer (open a browser on your BI machine to

http://127.0.0.1:32168/vision

). This webpage allows you to interact with the AI server... test it with sample images to make sure it is able to confirm for example, USPS.

 

[AlphaBlueIris]

Truglo, I did give up but you are bringing me hope again . From the I itial reading I assumed area a was going to be blank my mistake. I will rebuild it again. While I was using deepstack, I never had issues, while with this one, blueiris crash all the times. I was sure the delivery.pt was in that folder but I will verify it again. I will get it up and running this Friday but do not go away
Kidding apart, thank you for all your help, learning a lot.

Sent from my SM-S906U using Tapatalk

 

[truglo]

Hmm, so your VM was running smoothly with DS before, but having issues with CPAI? I wish I knew more about the back end on these AI servers to help with that. Hopefully it's just something that needs to be setup for the VM to work efficiently with cpai. Did you run both BI and DS on the same VM before? AFAIK, cpai does leverage intel or cuda HA... maybe it's just some driver missing in the VM.

 

[AlphaBlueIris]

Same vm, the only difference is the codeproject vs deepstack. Last week, I even reinstalled codeproject from scratch but I got the same results. The only difference is that I am using the custom pt files, from codeproject web portal, it seems to me that every time it analyze a picture, it goes thru the entire set of them.

Sent from my SM-S906U using Tapatalk

 

[truglo]

The only difference is that I am using the custom pt files, from codeproject web portal, it seems to me that every time it analyze a picture, it goes thru the entire set of them.

I attached images of everything relevant to ai on my system in hopes it helps. Note the directory where pt files are stored (red square in 2nd image) must match the "Use custom model folder" in BI AI settings (red square in the 3rd image).

I also have "Use mainstream" checked (first image), together with "High" selected in AI settings (2nd image). These are 2 settings that increase CPU usage, and may not be needed in most cases. I'm just running it like that to test/prepare in anticipation of the new BI plate integration with OCR. I heard that will need mainstream images at least. It also seems hard to believe LPR can be effective without it... even my human eyes have a hard time reading substream caps.

 

[AlphaBlueIris]

It looks like that instead of a beer, I own you a case of it, thank you again for doing it. I will report back soon.

Thank you

Sent from my SM-S906U using Tapatalk

 

[truglo]

I just noticed in my first pic I have zone G unchecked in the cam AI settings. That is likely to add to confusion... it works out OK without zone G in my case (because zone A covers all interesting areas for AI for this cam), but it also essentially disables the zone G trick. In your case you do have to check zone G in AI, otherwise logos will probably get blacked out.

 

[AlphaBlueIris]

I'm very excited to make it to work. As soon as I have access to bi I will get my hand dirty. I will post some pictures so you know if I'm doning it right or I'm missing something.

Sent from my SM-S906U using Tapatalk

 

[AlphaBlueIris]

I'm going thru what you have posted so I understand what you are suggesting. Your below is actually brilliant. Instead of looking for a truck, I could look for any person traveling in my walkway and if any, then I will do what you have described. I think it is a very clean solution since the only people that are going thru the walkway are the delivery people, so I could then check for truck as well at that point. I think that the only thing that I'm need to change from your original suggestion is that Zone A would have marked just the walkway and Zone G everything.

[edit: On a sidenote, I don't see a street side mailbox. If you just want to know when you actually get deliveries, you could instead trigger the clone on the walkway. It can still confirm the logo on the van for alerts, assuming the delivery person that triggered it is parked in view. This could give you the same result but with less processing for each passing vehicle.]

I'm also looking for why AI is taking so long. the added complexity is the VM but from CodeProject web portal, it shows that I'm using GPU but I could be incorrect.

 

[truglo]

Ah gotcha, yeah I think that walkway zone A method will work out better for you. I don't know much about how cpai uses gpu, but I understand cuda support is fairly new and proper working configs are widely known at this point. I was playing around and checked cuda in BI AI settings, and nothing crashed... cpai portal showed cuda on the green bar... but there were red highlighted logs indicating cuda errors. AI kept timing out, and eventually BI crashed out... but I was surprised such an obviously wrong setting would go so far. There are no nvidia drivers or cards in my BI pc; the code is really lacking sanity checks, like a greyed out cuda check box when there are no cudas found.

 

[AlphaBlueIris]

I think pictures are worth a thousand words, I think we are getting close here.
One note, while I'm looking at those pictures, 2023-01-06 11_57_41-.png does not have the ability to testing and tuning, it is off. not sure what makes it active.

 

[truglo]

Not 100% sure, but I think AI test/tune gets greyed out if you are viewing a live feed of a cloned non-recording camera.

That last image is way too green, hehe. Use the 3-dot menus to turn off/disable everything except "Object detection yolo". Only yolo objects should be green. The scene, face, etc are going to absolutely kill your confirm times, and are in general not helpful for security cam use. I bet once you pare that down to just yolo objects, your confirm/unconfirm times will drop drastically (perhaps <500msec, vs 5000msec).

Also yesterday I discovered you can uncheck "Use default object detection" in BI AI settings (2023-01-06 12_03_33-.png). Then you can remove "objects:0" from all of your cameras. I think this is a cleaner configuration personally, but not sure it makes any difference performance wise. On another side note... confirm times will be longer after a reboot due to models getting loaded into RAM... so after a reboot/AI restart, wait until all of your used models have been triggered at least once before assessing performance.

Also... dang too bad that door frame is covering the logo... that could be why delivery.pt didn't find it. If that continues to be a problem, you may have to trigger on the road instead (get your AI shots before that door is opened).

 

[AlphaBlueIris]

Thanks for the speedy response. Not sure what I'm doing wrong, I disable them but they start again. My computer is hunted