Video analytics can be run in a variety of ways that have significantly different pros and cons. IPVM has identified 9 fundamental architectures, explaining the tradeoffs of each one herein.
These architectures are:
Run entirely on the camera using the camera manufacturer's own analytics
Run entirely on the camera using 3rd party analytics
Stream from a camera to a recorder using the recorder's analytics
Stream from a camera to a server using 3rd party analytics
Stream from a camera to a recorder than to a server using 3rd party analytics
Run entirely on the cloud, streamed from a camera
Run a combination of camera and recorder/server
Run a combination of camera and cloud
Run entirely on the cloud, streamed from a bridge
Additionally, certain architectures using 3rd parties for analytics or video management will require special integration. We conclude this report by looking at APIs, SDKs and the use of ONVIF to do so.
This is spot on and a great intro for both end users and service providers to grasp what happens in the back end. Often to many times projects w/video analytics are designed and sold incorrectly out the gate. Understanding the platform of the product you choose is key to making it scalable and successful.
Like many of you that have dabbled in video analytics for a while I have learned while I cut my teeth on analytics the hard way.
Maybe a follow up to this article up could be one about Video Analytics Deployments, the other half of how it works and where the partner often fails making the manufacturers products look inferior. IMO partner delivery failure is the biggest reason the technologies are not as widely adopted as they should be in this day and age.
I am surprised that the residential market seems under-served by analytics. Among my 100-home sub-development, dog poo is a recurring theme. Dog posture which results in unsavory byproducts is quite distinctive. My own experimentation in this area has been quite, er, fruitful. Where is the residential marketplace innovation, for goodness sake?
P.S. Thanks for a really great overview of a rapidly evolving area. This is an area that even amateurs can dabble in and really get decent results, but real integration is more elusive. Thanks again!
It's pretty trivial to build a model to detect a person loitering around walking a dog (which likely implies the dog has paused to defacate).
But you would have to combine that with pose detection - if the human picks up the commode, then no alert; if she doesn't, then send alert.
It then depends on the facility manager's tolerance for false positives on this issue. Pose detection is something that can have a fairly high FP rate.
But the biggest takeaway is that most property managers aren't going to want to pay $X per camera / month on several dozen cameras for what's a fairly minor issue, especially since they'll have to send a human to go and clean it up after the detection anyway.