With any major technology shift, there are often discrepancies between what the public imagines and what is practical.
Facial recognition, for example, had people trying to find faces in a crowd from a watch list shortly after 9/11, which was completely beyond its capability at the time. The technologists knew this, but some companies still encouraged projects like this, creating a serious negative reaction.
We’ve seen disappointment in some applications with video analytics, such as baggage left behind detection in an airport, which is far more complicated to make work practically than is generally presented. We at VideoIQ have avoided even offering it for this reason.
This makes education important. The good news is that most in the industry are trying to set better expectations on what is realistic.
However, there is one area that is shockingly out of whack.
Of the five largest camera companies in the security industry – all well known names by any integrator – except for Axis, they all claim to offer Video Analytics options with some of their IP cameras.
Most people assume this means that they are offering the kind of detection and performance as the largest and most well established names in the Video Analytics space. Technologies like VideoIQ and other early entrants were, after all, the ones who established what Video Analytics means to the industry. However, the software these camera companies provide and the advertisements they run suggest that they have the same capabilities. It is one of the biggest lies I’ve seen in the industry for a long time.
I can certainly see why they would want customers to believe they are offering products in the same class, but the technology isn’t even close. And it is clear that they know it, since they not only use the term Video Analytics when referring to their technology, they also call it Intelligent Video Motion Detection, or Adaptive Motion Behavior, or Motion Detection, etc.
What they are really offering is what is known as Advanced Video Motion Detection, and this is dramatically different from true Video Analytics.
For example, none of these AVMD technologies can distinguish a human or a vehicle from anything else, except by size. In other words, they only recognize a blob of moving pixels, and if the size of that blob is about right then and only then will it be detected. This falls far short of true object type detection.
Secondly, these technologies cannot work accurately in scenes with highly dynamic backgrounds, such as blowing bushes or tree branches, or rippling water: In other words, the types of changes and movements you find in typical outdoor applications.
You have to dig deep into their manuals to find the truth. Here are warnings from one of the big five:
- Movements may falsely be detected if there is: a reflective metal background, glass (glazed building frontages), water as a background
- Large areas of reflected light can also cause spurious motion detection
- A constant background is necessary in order to detect motion reliably
- A person walking front of a hedge that is moving in the wind will very probably not be detected
Another of the companies added these warnings:
- The ideal scene selection is one with light traffic and a clean background
- If heavy traffic or a busy background is unavoidable, place the monitoring zone or trip wire in a relatively stable area
- Avoid crowded scenes where people move in all directions or stand in one place for long periods of time
One of the top camera companies offered no warnings at all. However, the largest camera company in the industry gave the best list to show the limitations of Advanced Video Motion Detection. They say their technology might not work if the:
- Camera is shaking.
- Depth of object is too long.
- Object is too big or too small.
- Fluorescent light is flickering.
- Too many objects are moving.
- Weather condition is extremely poor.
- Movement of object is too fast or too slow.
- Object is moving directly toward the camera.
- Dirt, drip, or splash is on the dome cover of camera.
- Luminance level of image is too low (During nighttime, etc.)
- Outside light (sunlight, headlights, etc.) enters the shooting area.
- Luminance level of shooting area is subject to change (outdoors, by the window, etc.)
Real Video Analytics technologies can work under all of the conditions listed above. There are of course limits, and there are still false alarms, but you will see about 10X – 100X times as many false alarms in typical outdoor environments with AVMD, and many times more missed detections as well.
Advanced Video Analytics systems are designed to:
- Detect colors, contours, shapes and movements of humans and vehicles, not just luminance levels
- To ignore fluorescent lights flickering, sunlight, headlights
- Work in bad weather: Rain, snow, hail, sleet, fog
- Ignore camera shaking, dirt and drips on dome or lens
- Adapt to changes in environment automatically
If used indoors away from windows, a good AVMD system should work fine. Motion detection has always been usable indoors, and the new Advanced Motion Detection systems are slightly better. However, outdoors is another story.
These companies should be up front about these limitations, and they should be making it absolutely clear that AVMD is not in the same league as the technologies from VideoIQ and other true Video Analytics providers.
The problem with all of this is that the term, Video Analytics, is not being used carefully. People are using it to describe everything and anything, including traditional motion detection, license plate recognition and facial recognition. My suggestion is that industry start clearly differentiating: The term Video Analytics should only be applied if the system can provide:
- True object type recognition, not just pixel blob detection
- Able to discriminate objects of interest from highly dynamic background movement
- Automatically adjusts to changes in the environment
- Can track objects through the field of view
If it can’t do all of these, then it should be called AVMD.
The terms, License Plate Recognition (LPR) and Facial Recognition (FR) are clearly recognizable. There is no need to lump these into the term, Video Analytics. It only confuses what Video Analytics stands for, and leaves us with no clear way of distinguishing the type of system described above. Clarity is important, and so is honesty.
What is fascinating about this lie is that it isn’t the Video Analtyics companies who are perpetuating it. Usually you find the start-up companies with the breakthrough technology who get carried away, hoping for more attention. But in this case, the biggest camera companies in the industry (except for Axis) are the ones trying to ride on the coattails of the Video Analytics leap forward.
There is no excuse for not making these differences clear, and no one should be trying to make it look as if they are selling Video Analytics when they aren’t.
February 6, 2009 at 8:50 pm
Well said! It’s nice to see you hit a major problem right away.
Bosch is the worse here. They are literally the poster boy for irresponsible video analytic marketing. Their phony video analytic challenge at the trade show sets the standards for that.
I think this is a big problem for the industry – because it guarantees large customers are going to continue to have negative experiences that will poison market expectation.
February 6, 2009 at 9:50 pm
Thanks, John.
I’m surprised it has gone this far.
It hurts everyone to create this kind of confusion.
Doug.
February 9, 2009 at 1:45 am
An interesting article!
I think a good benchmark to convert qualitative measure to quantitative one is really the key to solve this kind of confusion.
February 9, 2009 at 12:25 pm
Doug,
I agree with you that it is important to distinguish analytics based on motion or object tracking from those based on more advanced methods.
If your implying that this is in a Pelco manual let me know and I will modify. Our anlaytics are designed to detect more than motion and they do work outdoors.
February 9, 2009 at 4:31 pm
Shawn,
You are pointing to one of the core issues underlying this problem – the lack of good industry testing that can show the difference in performance.
We at VideoIQ have been looking around for years to find an organization that could step up and fill this void.
Setting up good testing is more difficult than it might seem. I’ll address this in a future blog.
But you are right, with good testing this could show quantitatively how dramatic are the differences in results.
Thanks.
Doug.
February 9, 2009 at 5:18 pm
Dan,
Thanks for your response.
Yes, Pelco is one of the companies I was referring to. But the manual is actually not the problem in Pelco’s case.
If you run a simple Google search for “Pelco Video Analytics,” you will see the Endura Encoder promotes Video Analytics. It opens up to a web page that shows an outdoor scene with trees and bushes in the field of view. It looks just like any video analytics scene. Once those trees and bushes are blowing in the wind, do you think it is going to generate false alarms? Will it be able to correctly recognize a person? What about when there are leaves on the tree? Is that a good example of the type of environment where it will work reliably?
The product spec sheet also goes on in length about the Video Analytics behaviors.
However, this sales material does not warn about the limitations of the technology, or about the fact that it is not the same as the Video Analytics systems everyone has been reading about from other companies, even though it sounds and looks as if it is.
When you go to the manual for your product, you are right, it is right up front about the warnings, which are some of the points I listed in my blog post. Those very points would suggest that the scene with the trees and bushes would not be a good application. That’s how it looks to me, at least.
Also, in the manual it is called Adaptive Motion Behavior rather than Video Analytics. So, the manual actually looks pretty good.
Perhaps the marketing department is playing on the Video Analytics image, while the engineering department writes the manual and they are trying to be more technically accurate? I don’t know, but there is a gap that the customer won’t see until after they’ve bought it and tried to use it.
Another thing to keep in mind: Outdoor use is far more difficult than it might seem.
All of the leading video analytics companies spent 3-4 years working through unexpected challenges. Even today, many of those companies have real difficulty in most outdoor applications, even with all the advanced video analytics tools they have. Using only Motion Behavior, I think you will find, falls far short of most users expectations, except in special cases. It might work in a parking garage, where there is little background motion and stable lighting, but this is not what most people think of when you say it works outdoors.
I agree that the important issue here is that Motion Behavior should not be confused with more advanced approaches.
I hope this feedback helps.
Thanks for asking.
Thanks.
Doug.
February 10, 2009 at 11:26 am
Good morning,
I have accidentally found this very interesting article. I live in Poland, Your point of view is not very popular in my country, but I agree with it. People here don’t even wonder about it and I think it is worth of mentioning. I have got a question than: could I translate this text to my mother tongue and publish it on my site (about security too)? Of course with link to Your blog and Your name there. Sorry I am asking here, but I couldn’t find Your e-mail address. Now You know my address, so You can answer the way most comfortable for You.
Thanks, regards.
February 10, 2009 at 12:12 pm
I think that part of the problem is that like with many things, the buzz word becomes a marketing accessory rather than a generalized term. I often see the word “analytics” thrown around to mean many things. I think there are really three segments of analytics. One is the AVMD that you refer to. Another is the real “Video Analytics” you refer to. The last is something in between – the ones that fall in the grey area in between “same old stuff with a new label” and “revolutionary.” Most integrators are largely uneducated about analytics, as are their customers. Their customers don’t often know the difference between the different types of capabilities, but they do know they want more than “just recording.” The integrators typically aren’t going to tell them that what they want, and what they’re being offered, are not necessarily the same thing. There are some pretty clever things you can do with AVMD as you’ve labeled them without spending the premium on “real analytics,” but in many situations if the budget is available, more advanced analytics are the only way to go.
Another part of the problem in terms of creating a large grey area where understanding does not exist is that many hardware manufacturers have relationships with 3rd party software companies that design the analytics. These analytics may exist for purchase directly from the software company, so essentially the hardware manufacturer is a reseller of something they don’t completely control. I’d prefer to purchase such 3rd party applications separately, since in many cases they truly are nothing but a 3rd party application that’s been rebranded or the DVR/NVR manufacturer simply doesn’t advertise that it’s not really theirs. I’d just as soon skip the delusions and have the manufacturer either buy the third party company or at least clearly indicate that the applications are 3rd party so that integrators and consultants don’t have to spend a whole lot of time sorting through the weeds to determine where it’s coming from. The second paragraph is really more of a soapbox than the primary issue, but it’s one that I think also causes problems.
February 10, 2009 at 12:50 pm
Doug,
Thanks for clearing up that we don’t have issues in our manual. Since I’m the product manager I’m responsible for the marketing of our behaviors.
We have 7 different behaviors available on the intelligent encoder. One of which is adaptive motion used mainly for outdoor motion detection.
If the customer is trying to detect people going the wrong way in an airport exit they wouldn’t buy the adaptive motion behavior. They would buy our directional motion behavior.
To address your concern that we don’t give customers enough warning, I’m working on a video analytics design guide that will help customers design analytics for the right application. A lot of this information is already out in the public domain.
When I went to Google to search for Pelco video analytics it brought up a UK company marketing our product. Since we have a lot of resellers around the world it is very difficult to control the messaging. On the Pelco.com web site, and in our marketing materials, we market the 7 distinct behaviors for each customer application.
I would like to respond to our behaviors working in outdoor environments. It is a matter of setting them up properly. Our behaviors will learn the scene including motion from trees over a period of time. We also have a sensitivity setting to reduce the false alarms. I believe the industry experts that say “analytics are not 100% accurate”. If your company or any other company is claiming this, they are not telling the truth. Analytics is a balance between proper configuration (sensitivity) designed for a particular application that is also usable (reduced false alarms).
We have a team of engineers that have been working on Video Analytics for 5 years. I agree with you that people mis-market their capabilities in manuals and websites.
At Pelco we carefully craft our marketing messages to our customers. However they have their own expectations of video analytics from either past vendors or unsuccessful installations. My advice to anyone reading this blog is to have a shoot out for their particular application between video analytics vendors. The proof is in the pudding. You can contact me at domalley@pelco.com for more information.
Thanks,
Dan O’Malley
February 10, 2009 at 1:14 pm
Marian,
Yes, feel free to translate and post this on your site. Adding a link back to this blog and my name as author would be perfect.
I agree that most people are not aware of the issue.
Let me know what you hear back from your readers.
Thanks for helping to get the word out.
Doug.
February 10, 2009 at 1:35 pm
Nick,
Good points.
You’re right – the customers don’t know the difference and no one is telling them. In most cases, I don’t think the integrators realize the difference either. We see that all the time.
I agree with your point about AVMD. On the screen it can look the same. But the performance is not close. And what I find interesting is that in most cases it is not any less expensive. Compared to our products at VideoIQ, for example, it is generally more expensive to buy AVMD.
I also agree with you on the whole licensing issue. Not having developed it, the manufacturers often don’t know the strengths and weaknesses. This can also make it a challenge for them to provide good tech support.
Thanks for the comments.
Doug.
February 10, 2009 at 2:26 pm
Dan,
Thanks for the added information about your intelligent encoder.
Here is a quick comparison with VideoIQ’s video analytics encoder, if you are interested:
You license each of your behaviors separately. Some appear to double the cost of the system, but all our behaviors are included in our encoder at no extra cost, along with the software to monitor alarms and view live video. Plus, you can use all of our behaviors at the same time on multiple regions of interest for each camera.
Our encoder also ships with a hard drive, so you don’t need a DVR. This eliminates most of the bandwidth and prevents the loss of video if the network goes down. Plus, it saves money.
I like the approach you are taking to educate and help the integrators know where and how to use. This is probably the most important thing we can do. This is exactly the right approach to take and I compliment you on doing this.
You are absolutely right that there are no analytics with 100% performance. But there are still big differences in performance, especially outdoors. I’ll put up another post to help show the differences.
Check out my other blog post on the gap in perception: http://spotonsecurity.com/2009/02/06/the-gap-in-perception/ and John Honovich’s article I listed there. We, as manufacturers know how well our equipment can work when set up and used properly, but the question is how easy is it for any integrator to get the same results? There is often a big gap there.
This is often true with new technologies. However, breakthroughs don’t usually become popular until they are as easy to use as the untrained person expects them to be.
We think our products are there now, although, we still see lots of room to get even better.
I agree that side by side testing is sometimes the only good way to see the difference. I’ll second that recommendation.
Thanks.
Doug.
February 13, 2009 at 7:10 pm
[...] blogu Pana Douga Marmana znalazłem interesujący artykuł dotyczący niedomówień związanych z zaawansowaną detekcją ruchu, analizą obrazu, video itp.. [...]
March 13, 2009 at 3:26 pm
[...] it comes to comparing advanced motion detection systems (http://spotonsecurity.com/2009/02/06/the-big-video-analytics-lie/), that sometimes try to pass themselves off as video analytics, the contrast is even more dramatic. [...]
October 20, 2011 at 8:36 am
Buddhist Discussion…
[...]The Big Video Analytics Lie « Spot On Security[...]…
April 2, 2012 at 11:08 am
Foolproofing is a myth for anything that’s entirely “left-brained” but has to work in a domain that needs manual intervention. Doug agrees with that in this insightful and enjoyable article – a must-read for beginners in video analytics!