Video analytics performance is all about accuracy of detection.
Any technology requiring calibration or tuning has an accuracy problem. I’m surprised this isn’t mentioned more often.
Why do you need to tell a system what the size of a human or a vehicle is before it will work accurately? The reason is simple: The technology can’t distinguish humans and vehicles from other false alarms accurately enough.
A bird doesn’t look much like a person. A tree branch doesn’t either. Humans don’t have a problem telling the difference. But this is what calibration is doing – it is trying to eliminate these kinds of false detections.
The only reason that VideoIQ’s technology needs no calibration or tuning to work reliably in even the toughest outdoor environments is because it is accurate enough. The difference in accuracy is astounding compared to all others. It’s not just a little bit better. You won’t often find such a big discrepancy in a field with so many players.
Even the best of the best analytics products require size calibration to work reliably. Comparing uncalibrated systems against the VideoIQ iCVR shows how dramatically better the accuracy is. Even the best systems aren’t close.
When 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. They are unusable outdoors without calibration. That’s how bad their detection accuracy is.
Eliminating the need for calibration and tuning is not just about ease of set-up. That’s a huge benefit, but accuracy becomes an issue everywhere: Where you can use it, who can install it, how many problems will it generate.
For example, over a year ago, one of the largest video analytics companies advertised partnering with a company who wanted to use their technology for protecting a particular residential and commercial application. It was supposed to be a nationwide roll-out. There was one limitation: The system needed to be installed by untrained technicians and therefore calibration could not be used. So, they limited the range of detection, turned down the sensitivity and tried to detect only static scenes, mostly areas of cement – not where other moving objects might be present, such as bushes or trees, etc. In other words, just looking at an area of cement with everything else masked off.
Even those limitations weren’t enough. Shadows from nearby trees would often fall into the detection areas. Reflections from the sun or headlights would cross into the regions of interest. Leaves, paper or dead branches would cause false alarms. After some extensive testing, it became obvious the technology simply wouldn’t work.
The question isn’t how well does the technology work when installed and continually re-tuned by a trained professional, but how well it performs when anyone uses it.
It all comes down to detection accuracy.
March 27, 2009 at 10:51 am
I see that people are identified as vertical objects and cars as rectangular objects. “Other” unknown objects are identified by a yellow box. What do you do with a crawling human? It still “looks” like a human, its just horizontal.
March 27, 2009 at 11:39 am
Liam,
There are companies that sell what they call video analytics that work they way you describe: They identify a person by a vertical object and a car as a horizontal object. These technologies will always give you very poor performance.
VideoIQ’s technology does not work this way. We use the overall appearance of the object and look for a lot of distinguishing features to recognize what a person looks like or a vehicle looks like.
For example, with a technology that identifies people as a vertical group of pixels, what happens when they walk behind a desk or a park bench and the camera can’t see their legs? Their technology will not be able to recognize them as people anymore. This is why some of these technologies don’t work well indoors, where people are often occluded and blocked by desks and furniture.
VideoIQ’s technology doesn’t use this approach, but recognizes what people look like by a number of distinguishing features that are unique to people. If you go to our web site, http://www.videoiq.com, you will see a video clip of a person behind a blowing tree branch. We do not false alarm on the moving branch, and can still detect the person behind it. This would never work if our technology was only recognizing vertical groups of pixels.
Now, to your question about crawling people. We do not include crawling people in our template of people. So, if a person is crawling, they won’t be detected as human. We plan to add a template for crawling people in the future, but have not seen much demand for it yet and we have a long list of other things that people tell us is more important. So, it is low on priority.
If you have a very high security, high risk site you want to protect and need to catch even people crawling, then we suggest you use the “suspicious object” alarm, which as you say puts yellow boxes around them. That works fine. It will also detect animals and sometimes will trigger on moving tree branches from time to time. So, you will get more false alarms, but you won’t miss anything this way.
If you need more assistance with this, feel free to write me directly or you can contact VideoIQ’s tech support.
I hope this helps answer your questions.
Thanks.
Doug.