Tuesday, March 25, 2008

Is video analytics for real yet?

My short answer is yes. My long answer is still yes it is, but under certain conditions and for a price.

Pixel based video analysis has been around for years. Available for free, we have all referred to it most commonly as simple motion detection. Video Content Analysis, also known as Intelligent Video Analytics (IVA), jumps beyond analyzing pixel changes and seeks to identify discrete objects in a scene, track them through it and draw conclusions about their behavior. Video content analysis 'knows' the difference between the near-random rustling of leaves in a tree-line versus a group of people walking together along the sidewalk underneath them. Pixel-based algorithms don't have the slightest chance of accomplishing this, which is why they are rarely used outdoors. However they have very high success rates indoors in quiet corridors where nothing interesting ever happens, normally.

So if IVA is ready for prime-time, then why isn't it all around us? There are a bunch of factors at play here. Surprisingly the first is not reliability nor price; it is experience. How can someone sell and install IVA unless he has learned the hard way from previous installations? And if that expertise has to come from the manufacturer then that automatically drives up the cost per channel because of the increased cost of sale, which is the second factor. Prices vary from several hundred dollars to $4,000 per camera depending on the architecture and the kinds of behaviors you want to detect. Government, airports, military and critical infrastructure may be able to justify it, but what about the Elementary School that wants to stop people parking in the fire lane in front of the school? I could put reliability as the third factor, but really IVA works reliably under certain conditions - and it is experience that tells you what those are. So, I'm going to lump reliability with experience.

My vote for the third factor is total cost of ownership, commonly known as TCO. If you are doing video analytics on a bunch of liquid nitrogen-cooled PC servers that have more quad-core processors than an octopus has legs, then according to published data from the Meta Group, SAP and Gartner, the TCO for a PC is 1-5x the capital investment per year. So a $10,000 server costs another $10-50K per year to keep the little LEDs flashing on the front. Extrapolate that for 3 years to see how much these analysts estimate the system really costs you. I'm not saying that server-based analytics is bad; indeed it can do things that embedded analytics cannot; however, at least know the true cost of the system - not only in initial capital but also the maintenance, UPS power for the servers as well as the cooling, which has to be beefed up to cope with all the servers, and finally the rack space. Yes, IVA has a very real environmental footprint, but it hasn't hit most of us in the face yet because we have so little experience, because it hasn't become mainstream yet.

If Experience, Initial Price and TCO are preventing the mass market adoption of IVA, what has to change? Time will give us experience. And in my opinion moving the intelligence to the edge away from centralized PC servers, and detecting 'the most common' kinds of behavior using algorithms embedded in the edge devices will deliver much more acceptable TCO figures, drive up volume and consequently drive down unit price, bringing intelligent video analytics to the masses. And just like simple video motion detection, that's where it's needed most.

1 comment:

John Honovich said...

Hi Dr. Bob,

Great blog and nice article on video analytics.

I think analytics at the core will be the most powerful means to reduce the cost and complexity of deploying analytics.

1. Centralized analytics allows the reuse of existing cameras (smart cameras or appliances requiring buying/installing new hardware)

2. Centralized analytics leverage the growth in today's rapid growth in multi-core CPUs.

3. Aggregating analytics for multiple cameras in a single central unit allows for statistical multiplexing, enabling more cameras to be served with less dedicated resources.

4. Centralized servers are simpler to manage because it eliminates integration and simplifies upgrades.

I do believe centralized analytics has shortcomings (such as bandwidth load and problems dealing with decoding more advanced codecs such as h.264). However, for the mass market that wants simplicity and low cost, performing the analytics centrally such as within the dvrs will be a powerful economic force.

I wrote more about this at: http://ipvideomarket.info/review/show/9