[tool] [tutorial] Free AI Person Detection for Blue Iris

Update: This doesn't seem to be an AI Tool issue, or even an "issue", really. I notice in BI that I have my motion sensor "min. object size" set to like 250. So by the time the little dog gets to the middle of my yard, the dog is smaller than the min object size. It was simple. :)

So getting back to AI Tool for a moment, hopefully this is a simple question to answer.

Background: I have a single camera, I'm running the VorlonCD mod v1.67.8.35473, recording the sub-stream 24x7, sending sub-stream jpg's to a folder for Deepstack and AI Tool to process, and then sending a trigger to the HD stream to start recording if a person is detected. (I should just put all that in my sig! :lol: ). That's been working pretty good so far.

I'm now experimenting with detecting dogs and cats (just checked the boxes in AI Tool last night).

Fortuitously, a little doggy roamed into my yard this morning, and I got an HD recording of it. However, the recording stopped after 23 seconds, even though the dog was clearly still right in the middle of the camera's view, but walking away. And admittedly, it's a little dog...maybe weighs 5lbs at the most.

So, the question is: why did BI stop recording while the dog was still roaming around in the yard?

Thanks!
Check to see what your break time is set to on your HD cam. This is what defines how long to record after motion stops.
 
Update: This doesn't seem to be an AI Tool issue, or even an "issue", really. I notice in BI that I have my motion sensor "min. object size" set to like 250. So by the time the little dog gets to the middle of my yard, the dog is smaller than the min object size. It was simple. :)

So getting back to AI Tool for a moment, hopefully this is a simple question to answer.

Background: I have a single camera, I'm running the VorlonCD mod v1.67.8.35473, recording the sub-stream 24x7, sending sub-stream jpg's to a folder for Deepstack and AI Tool to process, and then sending a trigger to the HD stream to start recording if a person is detected. (I should just put all that in my sig! :lol: ). That's been working pretty good so far.

I'm now experimenting with detecting dogs and cats (just checked the boxes in AI Tool last night).

Fortuitously, a little doggy roamed into my yard this morning, and I got an HD recording of it. However, the recording stopped after 23 seconds, even though the dog was clearly still right in the middle of the camera's view, but walking away. And admittedly, it's a little dog...maybe weighs 5lbs at the most.

So, the question is: why did BI stop recording while the dog was still roaming around in the yard?

Thanks!
I have found an issue in BI that can cause miss triggers. Seems that no matter what min and max size are set it can fail to see a moving human 10 ft away in certain areas :wow: I have been trying to solve the problem for some time and even set the trigger to 0 seconds. The camera is A HIKVISION 8 MP with the trigger stream set for 1280 x 720. I have finally solved the issue by by changing the trigger sensing algorithm to SIMPLE. Apparently SIMPLE mode is the algorithm that was used in version 4. All cameras are now even detecting birds reliably.
 
  • Like
Reactions: Vettester
I have found an issue in BI that can cause miss triggers. Seems that no matter what min and max size are set it can fail to see a moving human 10 ft away in certain areas :wow: I have been trying to solve the problem for some time and even set the trigger to 0 seconds. The camera is A HIKVISION 8 MP with the trigger stream set for 1280 x 720. I have finally solved the issue by by changing the trigger sensing algorithm to SIMPLE. Apparently SIMPLE mode is the algorithm that was used in version 4. All cameras are now even detecting birds reliably.

Glad you figured your issue out!! It's funny you are moving in the reverse of what most are trying to accomplish.. ignore dogs and birds to see people only. Haha!


Sent from my iPhone using Tapatalk
 
Getting these every morning now:

[11.09.2020, 06:50:04.329]: Error: File watcher error: Too many changes at once in directory:C:\BlueIris\aiinput.

I was thinking it was the cameras switching from night to day, but it's always just before 7, and they've already switched at that point.

I'm using the VorlonCD mod, and may switch back to see if that makes a difference.

Sent from my Pixel 3 using Tapatalk
 
I have found an issue in BI that can cause miss triggers.

Thanks for posting this! I have been fighting with motion detection for awhile now and when I seen your post I went in to see what algorithm mine were set at. Originally I had them all set to use gaussian and to my surprise this algorithm had been deprecated and they had all changed to use edge vector. I guess that’s what I get for not reading all of the release notes before updating. Anyway, I‘ve changed all my cameras to use the simple algorithm and everything is working much better now.

Thanks again!
 
Getting these every morning now:

[11.09.2020, 06:50:04.329]: Error: File watcher error: Too many changes at once in directory:C:\BlueIris\aiinput.

I was thinking it was the cameras switching from night to day, but it's always just before 7, and they've already switched at that point.

I'm using the VorlonCD mod, and may switch back to see if that makes a difference.

The error would have happened with the original GP release, but just not reported in the log. Simply an overload of new images being created in that folder, and the file watcher cannot handle them all.

If you dont have a ton of cameras, set each AI camera in BI to use its OWN clip folder, then set each camera in AI tool to monitor only that folder. (And remove the base input folder from the AITOOLS settings tab).

Also, in BI make sure the AI Camera > Record tab > Only when triggered is CHECKED, and that you dont record snapshots for any longer than needed. (I have mine set to 3 seconds whenever something is triggered for example).
 
  • Like
Reactions: barnyard
Glad you figured your issue out!! It's funny you are moving in the reverse of what most are trying to accomplish.. ignore dogs and birds to see people only. Haha!


Sent from my iPhone using Tapatalk

AITool is normally set for person detection but I used bird activity (feathered variety) to test. In the attached photo my camera will fail to detect me walking around the building on the patio in the V5 default detection mode. In simple mode it will detect me looking out the window! False triggers have increased by about 10% but AITool handles it just fine. I realise this is hard to believe but this is my experience. I was thinking of sending Ken at BI a note as this behavior is pretty bizarre.
 

Attachments

  • Garden_2020-09-11_162939.jpg
    Garden_2020-09-11_162939.jpg
    604.4 KB · Views: 72
Thanks for posting this! I have been fighting with motion detection for awhile now and when I seen your post I went in to see what algorithm mine were set at. Originally I had them all set to use gaussian and to my surprise this algorithm had been deprecated and they had all changed to use edge vector. I guess that’s what I get for not reading all of the release notes before updating. Anyway, I‘ve changed all my cameras to use the simple algorithm and everything is working much better now.

Thanks again!
Your welcome. I'm beginning to believe that edge vector mode might actually be broken. It will be interesting to see if other foum members have experienced the same issue.
 
Here is the latest update for the VorlonCD mod:

  • MQTT support!
  • In my last release I didnt mention that I 'borged' the 'Copy alert images to folder' feature from akmolina28 's fork - Thanks!!
  • TEST buttons for actions dialog
  • Overhaul of Deepstack url queuing system when you have multiple url's specified in settings. [ Thanks for testing and holding in there @Tinbum ]
  • Lots of updates to the 'Dynamic' Masking feature and its details dialog box [ Thank you @classObject ! ]
  • The custom draw mask dialog box has a lot of nice changes. Transparent masks as you draw, the ability to specify rectangle, etc. (This replaces the complicated process for creating your own camera masks on the first page of the monster IP Cam Talk thread.) [ Thank you @classObject !! ]
  • May @GentlePumpkin be with us. Amen. :)

Log issues and feature requests up on VorlonCD/bi-aidetection

 
Last edited:
New version working fantastically!
Question, would it be possible for Telegram to send basic info with the image like camera name, and what it saw?
Example:
Person spotted at FrontDoor

Is that doable?
Would be super useful to see what the alert is on my smart watch without having to tap on the message, and open the picture.
Just a thought!

Sent from my Pixel 3 using Tapatalk
 
Here is the latest update for the VorlonCD mod:

  • MQTT support!
  • In my last release I didnt mention that I 'borged' the 'Copy alert images to folder' feature from akmolina28 's fork - Thanks!!
  • TEST buttons for actions dialog
  • Overhaul of Deepstack url queuing system when you have multiple url's specified in settings. [ Thanks for testing and holding in there @Tinbum ]
  • Lots of updates to the 'Dynamic' Masking feature and its details dialog box [ Thank you @classObject ! ]
  • The custom draw mask dialog box has a lot of nice changes. Transparent masks as you draw, the ability to specify rectangle, etc. (This replaces the complicated process for creating your own camera masks on the first page of the monster IP Cam Talk thread.) [ Thank you @classObject !! ]
  • May @GentlePumpkin be with us. Amen. :)

Log issues and feature requests up on VorlonCD/bi-aidetection


What is the best way to upgrade?
 
Fresh install of buster lite (with desktop). Then did the update and upgrade process.

i followed the instructions here - Using DeepStack on Raspberry PI - ALPHA

via SSH i ran:
wget
and then i had to bash install-deepstack.sh and it worked.

started it up, entered my license key. plugged in NCS2. Thats it.

Just wanted to follow up here to see if your experience running Deepstack on a raspberry pi with an Intel NCS2 has changed after using it for a while? I’m getting ready to pull the trigger on a raspberry pi 4 and a Intel NCS, but wanted to see how it performs when multiple images are sent to it.


Sent from my iPhone using Tapatalk
 
  • Like
Reactions: Eatoff