Google artificial intelligence

[t_andersen]

Hi guys,

this has been touched upon in another thread but I start a new thread because I think that this issue has a future for BI. The BI motion sensor is as good as it gets with conventional methods but many of us get false alarms from tree shadows moving in the wind etc, etc.

Google has released some neural network models for image recognition, that can be trained to look for cars, people, dogs, - in principle anything. They have also released a development tool called "Tensorflow". The most interesting models are "Inception v3" and "Mobilenet". Using Tensorflow, the models can be trained to look for almost anything you can think of. You need to install Python and then Tensorflow. I have got it running and it seems to be fast enough to be useful for BI.

There are instructions for a simple setup here:
TensorFlow For Poets

And an example of the capability here:



I have no idea how one could interface to BI.

If others are thinking along the same line, then pls give your thoughts so we can see if we can find a way through. There is a lot happening in this field right now.

 

[psycik]

A friend of mine has done something similar with AWS - Detectatron

Some of the thinking may be applicable.

 

[Matt L.]

Neat. I want this for a home security and automation system I am working on with arduinos, BI, cameras, locks... etc.

 

[Q™]

Hmmm. So SkyNet has released the beginnings of the engine which eventually matures into the T-800's optical recognition system. What could possibly go wrong?

 

[jrf]

In that other thread I added into this also:

Script to Detect Colour, Make and Model of Car from IP Camera

I can tell you...yes so far after only a day. Huge improvement. I see lots of rain and snow storms in my area/fog/etc. Makes my cameras go trigger happy. The only logical next step is to actually analyze what's in the image. I do like the idea above "Detectatron" however that is after the fact stuff. Which can be quite useful in addition to something that is a bit more realtime. I can see analyzing an immediate image for an object. Then later submitting video files to be analyzed for further activity. However I suspect that will be difficult since some of our cameras see areas we could care less about activity in (the road) as every car passing by would trigger as interesting. I can see this technology really taking off in the next few years and it should be pretty darn exciting for us! Sighthound is definitely not fool proof. But for me it's a start! I'm going to have to graduate to Tensorflow and I'm sure it's more accurate.

It's possible that someday in the future BlueIris could tap into a subscription service and use that as an alert method....