CPAI Coral Detections not working

[smith123]

Hello, I have the coral TPU setup and running and working on a raspberry pi. I have tested it and am getting back good results, but the it seems like blue iris isn't taking the results correctly? I am not sure how to describe it but when I check the AI debugging I am not seeing the person icons, just circle crossed out logo. See attached image.

Any ideas on what I am doing wrong?

 

[mike32162]

Are you sure you can use the custom models with Coral?
On my main screen in the settings (where it gives launch defaults) custom is greyed out and MobileSSD is what it is there.
On your Codeproject server (click the launch button from the same AI settings window in BI settings) is the Coral TPU enabled and running? Start it if not. Is Yolo disabled? Someone correct if I'm wrong, but I don't think you can use them both without conflict, and when I disabled Yolo, detection times decreased drastically.
While you're on that screen, look at the status log and see what it is doing.
To make finding what you want easier, turn off static object detection. You're constantly pinging the AI.
Other notes (but this doesn't seem to be an issue):
Make sure windows services has codeproject set for automatic startup. If you're running BI as a service, change it from automatic to delayed. CP has to start first. Reboot it (or stop the 2 services, then launch CP, wait, launch BI) and see what happens.
Your times in the coral.png file seem way high. Wondering if your system is timing out, perhaps caused by one or more of the issues above.
Edit -- go to the status tab (where you can see the log, cameras, drop AI files, etc...) click on log, and see if you're getting AI timeouts.

 

[MikeLud1]

Are you sure you can use the custom models with Coral?

Custom models is not supported with the Coral module

 

[mike32162]

Custom models is not supported with the Coral module

That's what I thought. Look at his detection times. That doesn't seem normal, so something must be mis-configured, and those times are likely causing timeouts.

 

[smith123]

Are you sure you can use the custom models with Coral?
On my main screen in the settings (where it gives launch defaults) custom is greyed out and MobileSSD is what it is there.
On your Codeproject server (click the launch button from the same AI settings window in BI settings) is the Coral TPU enabled and running? Start it if not. Is Yolo disabled? Someone correct if I'm wrong, but I don't think you can use them both without conflict, and when I disabled Yolo, detection times decreased drastically.
While you're on that screen, look at the status log and see what it is doing.
To make finding what you want easier, turn off static object detection. You're constantly pinging the AI.
Other notes (but this doesn't seem to be an issue):
Make sure windows services has codeproject set for automatic startup. If you're running BI as a service, change it from automatic to delayed. CP has to start first. Reboot it (or stop the 2 services, then launch CP, wait, launch BI) and see what happens.
Your times in the coral.png file seem way high. Wondering if your system is timing out, perhaps caused by one or more of the issues above.
Edit -- go to the status tab (where you can see the log, cameras, drop AI files, etc...) click on log, and see if you're getting AI timeouts.

Okay, Thanks. I will do some testing tomorrow. What I am doing is just trying to test 1 camera with coral to make sure it works well and fast. So I have YOLO running on 1 machine and Coral on a different. So in my Blue Iris AI settings I have my normal code project YOLO and I am using the override feature to direct this one camera to CORAL. I am guessing blue iris is saying I have "these custom models" and coral doesn't see them. I also noticed that I am using face detection and that is probably timing out at 15 seconds. Not sure how if I can turn that off/on for an individual camera?

 

[smith123]

That's what I thought. Look at his detection times. That doesn't seem normal, so something must be mis-configured, and those times are likely causing timeouts.

I am running my coral on a raspberry pi 4. I am running code project ai via docker. I am wondering if the power supply isn't good enough or something, because I have ran test pictures through it and get 600ms or higher. I have seen people posing about times in the double digit ms.

 

[mike32162]

I am running my coral on a raspberry pi 4. I am running code project ai via docker. I am wondering if the power supply isn't good enough or something, because I have ran test pictures through it and get 600ms or higher. I have seen people posing about times in the double digit ms.

I would try taking the TPU out of the pi and installing it directly on the BI machine running CP, unless it is USB, then I'd order the correct M.2 Coral for your machine. I went the route of a cheap PCIE adapter card to hold the TPU.
My response times where <60ms, now about 150 because I changed the model size. Without the TPU, my CPU usage was spiking to levels bogging down the system. Fixed that by using IVS detection first on cameras in a windy area so they wouldn't constantly ping, but I did not like the fact that it went that high. For S&G, I tested the Coral without IVS on a windy day and with constant pings, it barely effected the BI system, however, still don't want to do that because if Coral is busy dealing with a flood of false alerts, a simultaneous request from another camera could time it out. My first line of defense was fine tuning IVS and BI motion detection where I'm using it to send what is necessary over to AI for processing.
I don't have a dedicated GPU to compare to Coral, but this cheaper and more power efficient route convinced me not to bother.
Even if it is working the way you have it configured, I think your response times are so high BI is timing out waiting for it. You could change the timeout variable in the settings screen, but still, those response times are so high I'd find the root cause.

 

[smith123]

I would try taking the TPU out of the pi and installing it directly on the BI machine running CP, unless it is USB, then I'd order the correct M.2 Coral for your machine. I went the route of a cheap PCIE adapter card to hold the TPU.
My response times where <60ms, now about 150 because I changed the model size. Without the TPU, my CPU usage was spiking to levels bogging down the system. Fixed that by using IVS detection first on cameras in a windy area so they wouldn't constantly ping, but I did not like the fact that it went that high. For S&G, I tested the Coral without IVS on a windy day and with constant pings, it barely effected the BI system, however, still don't want to do that because if Coral is busy dealing with a flood of false alerts, a simultaneous request from another camera could time it out. My first line of defense was fine tuning IVS and BI motion detection where I'm using it to send what is necessary over to AI for processing.
I don't have a dedicated GPU to compare to Coral, but this cheaper and more power efficient route convinced me not to bother.
Even if it is working the way you have it configured, I think your response times are so high BI is timing out waiting for it. You could change the timeout variable in the settings screen, but still, those response times are so high I'd find the root cause.

Great, so I figured it out. Problem is I am using Face Detection on my normal AI system and it is turned off on my raspberry pi version of code project AI. When I inspect it there is a error code -1. If I turn on Face Reignition in Blue Iris everything works fine. If you have any ideas on how to run both at the same time, that would be great. For now, I think I will disabled Face Detection and see how things go. Thanks for everyone help, Hopefully someone else will find this information useful.