2. Create a backend that queries each camera for their current image.
3. Run the images through some object recognition neural network (fastai has some great object recognition tutorials).
4. Make your backend stitch together the images from each camera to make a video of what was being recognized when objects of interest were detected (each camera will create its own video)
5. Create a nice little UI/App for everyone to view the movies with some filtering in place (time of day, object(s) of interest).
I want to create something like this as well. All of the pieces seem to be pretty straightforward to me except for figuring out power and network connectivity. I guess each neighbor would have to configure the device to use their SSID... and hope someone doesn't steal the raspberry pi and get the SSID credentials from memory... Maybe just make the neighbors use a separate network?
https://github.com/openalpr/openalpr
Also at: http://doc.openalpr.com/opensource.html
And a ready-to-go version you can buy/subscribe based on their stuff:
Friends I have who live there are thrilled with the system although the Seattle police seem to have little interest in the system telling the residents that they are only interested in seeing footage that shows suspicious activities or is directly related to a crime.
https://www.flocksafety.com/
Porch pirates are so bad in my new neighborhood that I admit to wanting one of these on my dead-end culdesac and have investigated what it would take to build such a system myself.