Built-in AI not detecting (DEBUG ATTACHED)

[electron73]

Hi guys!
Really rare that I get stuck with something but here we go.
Blue iris latest running on Windows10 LTSC VM on a Ryzen 9950X Proxmox, everything works great and snappy, events getting detected properly, persons, vehicles detected properly, round trips for AI in the 50-60ms range. But there is a mystery here.

Please see the attached screenshot. WHY did this event not trigger? A person was clearly detected with 84% and 70% and then 63% confidence in three frames in a row. The canceled alert just says "occupied" and there is a yellow stopwatch.
I have other (confirmed red person) events from the same camera with less confidence or person detected in just one of the frames yet it triggered correctly.
It feels like Blue Iris is on the cutting edge with the software but the docs are ANCIENT.

Does anyone know whats going on here?

 

[hapstabu]

I have been trying to resolve this with settings, but have been unsuccessful. I would advise you to email Blue Iris support with as much documentation and examples as possible.
I have been emailing with Ken, but the more examples he receives might help him put this issue higher on his list.
I have not found any settings that resolve this.

 

[hapstabu]

There is more discussion about this issue in this thread:

Thread 'Blue Iris Version 6.0.1.24 Now Supports Custom Models'

Jan 23, 2026

You can download my custom models from the below link

Blue-Iris-Custom-AI-Models/Custom-YOLOv5 at main · MikeLud/Blue-Iris-Custom-AI-Models

Blue Iris custom AI models. Contribute to MikeLud/Blue-Iris-Custom-AI-Models development by creating an account on GitHub.

github.com

[6.0.1.24] – 2026-1-23
Added
• Support for custom models using the built-in AI. The model must be an object
detection model with an ONNX extension and either the Yolo5, 8 or 10 format.
Anything else may cause hanging or instability; please send the model to support for
analysis if you are unsure.

Fixed
• A potential code ambiguity regarding use of global custom models and camera
overrides

• MikeLud1
• Replies: 163
• Forum: What's New

• 

 

[jrbeddow]

Your timings between each analysis are too closely spaced, thus not properly identifying the motion between each "frame" of analysis. Try changing that setting to every 750ms or even 1 second for relatively slow moving objects like people. Somewhat of a "known issue" with the version of AI that Ken has integrated into BI.

 

[hapstabu]

Not only built in AI, but also CPAI.
And if all I want to do is track people walking, this will work.
Vehicles, not so much.
So Blue Iris is now good at identifying slow objects.

 

[electron73]

I'm so confused. The AI detects a sitting, stationary car in my street every time the street cam gets a motion event. I have intentionally not set the option to require ongoing motion overlap on the AI tab because I just want to trigger a web hook to my home automation to play a jingle every time there is ANY motion by the front door and a person is detected there..
If I understand this right from the sparse documentation, when a camera causes a motion trigger either via on if event or built in motion detector, the frames are sent to the AI for analysis and if a required condition is met ("person") then blue iris will execute the action.
Nowhere is it explained or clear how many frames one should analyze and how far apart the they should be. It sounds like looking at multiple frames gives you a higher chance to detect something which worked perfectly in my example. The person was detected three times. Why didn't it trigger?

 

[jrbeddow]

Have you actually tried changing the time between each analysis to a minimum of 750ms (or 1 second) for this specific camera with a view of an entryway? This is a particularly difficult view, with most of the motion happening in the parallel (or "head-on") direction, making it even more challenging for the AI to see motion if each frame is too closely spaced. You might get away with closer timings with faster moving objects or ones that cross perpendicular to the field of view.
I'd also suggest 2-3 pre-trigger and at least 6-8 post-trigger analyses (I use up to 16 personally). Despite the current general recommendation against it, I also still add an "impossible" object (table, girafe, etc) in the "To Cancel" box: this forces the maximum number of analyses that you have specified.

 

[hapstabu]

My opinion, and then I will stop posting about this.
For about a year and a half or more, I used the same pre post and timing settings. I didn't have to change them since moving to CPAI.
Pre: 4
Post: 10
Time: 500 msec (for years, since Deepstack)
I never saw this behavior. All objects are stationary, even if they are moving.
You shouldn't have to make all these adjustments to allow Blue Iris time to figure out if an object is moving.
There is something going on with the logic that identifies stationary objects. Blue Iris is getting confused.
Email BI support. Ken is out of the office until next week however.