Tracking with Yolo4

[John L. Poole]

I finally made time and ventured into video analysis with AI.  The biggest point I learned last night was that processing a video using a GPU made available through Google's Collab was about 2400% faster than my own Jupyter notebook sitting on an AMD Bulldozer with no GPUS.  11 second video analyzed in 24 seconds vs. 1 hour 38m 60s.  Guess who is going to build an AI processing machine.

Any rate, I toss this email into the moonfire-nvr users group to see if there is interest in performing AI analysis of videos.  My goal include having segments that have been triggered as something happening with the Motion project, being analyzed by AI to determine if its a car, pedestrian or animal.  I'm okay with sharing my Jupyter notebook if there is interest.  I'm new this area, so I'm a bit clumsy at it... but, I'm learning... real fast.

I submitted a photo of my dogs and was surprised to find the frisbee classification in the Yolo model:

--

John Laurence Poole
1566 Court ST NE
Salem OR 97301-4241
707-812-1323

[Scott Lamb]

I definitely want to support ML-based analytics in Moonfire. In the meantime, I'm running Frigate in parallel with Moonfire. Combined with an arcane Home Assistant + Telegram setup, it does a nice job of notifying me when someone's in my driveway or at my front door.

On Wed, Apr 13, 2022 at 9:23 PM John L. Poole <jlpo...@gmail.com> wrote:

I finally made time and ventured into video analysis with AI.  The biggest point I learned last night was that processing a video using a GPU made available through Google's Collab was about 2400% faster than my own Jupyter notebook sitting on an AMD Bulldozer with no GPUS.  11 second video analyzed in 24 seconds vs. 1 hour 38m 60s.  Guess who is going to build an AI processing machine.

If you want to do inference only with a pre-made model, I recommend looking into a Coral accelerator. The USB version is $60; the M.2 versions are $25. One may be enough for all your camera streams. One caveat is that while it does the inference, you still need a powerful enough machine to do the H.264 decoding and scaling on all the video streams you want to send to it.

If you are investing effort into setting up ML, I encourage you to check out this open source ML training data + model project that is specifically meant for CCTV/NVR applications: https://gitlab.com/Curid/TF-CCTV Its author is looking for folks to help sorting the existing public data set, and maybe contribute new training data taken from security cameras.

Any rate, I toss this email into the moonfire-nvr users group to see if there is interest in performing AI analysis of videos.  My goal include having segments that have been triggered as something happening with the Motion project, being analyzed by AI to determine if its a car, pedestrian or animal.  I'm okay with sharing my Jupyter notebook if there is interest.  I'm new this area, so I'm a bit clumsy at it... but, I'm learning... real fast.

I submitted a photo of my dogs and was surprised to find the frisbee classification in the Yolo model:

--

John Laurence Poole
1566 Court ST NE
Salem OR 97301-4241
707-812-1323

--
You received this message because you are subscribed to the Google Groups "moonfire-nvr-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to moonfire-nvr-us...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/moonfire-nvr-users/5532df63-ed4b-992c-3c40-0d52937a14cb%40gmail.com.

--

Scott Lamb <https://www.slamb.org/>