ExNVR and machine learning

[gBillal]

Bring machine learning to ex-nvr, try a machine learning model (object detection, face detection, motion estimation, ...etc.) The final result may be as follows:

• Read a stream from IP cameras, use the sub-stream if possible (already done)
• Decode the snapshots
• Feed the snapshot to the machine learning model
• Save the results somewhere (or include it in the snapshots)
• Show it on the web UI (there's multiple ways, overlay the results on the video player, encode the video with the metadata hardcoded on the snapshots, ...etc.)

How to include ML in Elixir:

• Pre-trained models and hugging face models with Bumblebee
• Onnx models with Ortex
• Or be brave and build a model yourself using Axon
• Or be super brave and run it using NIF/CNode with C/C++ code :sunglasses: like this lady

[marcoherbst]

My ask on this task is to do the absolute minimum amount of decision making in order to show an end-to-end working solution.

I believe a path to that would be to take inspiration from: https://culttt.com/2023/01/26/video-object-detection-elixir-nx-bumblebee

And replace the MP4 (or even don't !! ... let it run on recordings ??) Use his UI completely as is , in a new tab.

Of course this is only intended as a proof of concept and would not get into the main branch but it would still establish a lot of the basic ideas and get us familiar with the dependencies etc. etc. with, I hope, minimal effort or decision making required.

[marcoherbst]

My current thinking on this: Run the Evercam Big Model at low fps (1 per min?) to detect the presence or abscence of different types of equipment over the timeline.