You may be familiar with the classic tale of the boy who cried wolf. As a shepherd, he spent all day watching his sheep graze peacefully, yet on occasion, would cry “Wolf!” to get a villager or two to run down to help. When they arrived, he’d yell “Gotcha!”, laugh, and the poor villager would sulk back to the village, angry that they wasted their time on his joke. Of course, we all know what happened next. By the time an actual wolf came around, and the boy cried “Wolf!”, no one came to his aid. The result? No more sheep, one very satisfied wolf, and an out-of-work shepherd.

If this boy was a video analytics system, he would have what we call a poor signal-to-noise ratio.

Signal-to-noise is one of the general terms used to describe how often an automated alerting system returns a true event versus a false event. It’s also what typically fails these systems in the real world, as system operators learn to ignore alerts because the majority of those received tend to be false. In the surveillance world, automated alerting has primarily been used for video analytics and surveillance system “health check” programs that check assets for operation and uptime. Here’s an example of how they fail because of signal-to-noise issues.

Let’s assume I manage an installation of 100 cameras in a medium-sized corporate facility. I assign one guard per shift to sit in an on-site central monitoring station to watch screens, get alerts, and be prepared as a first responder in case of an event.

First and foremost, I have a health check running on my surveillance system. If a camera goes down, the guard protocol is to receive the alert, check the monitor, verify the problem, and fix the problem (or call the surveillance installer to fix it). I have one of the best in breed health check systems. That means that each camera only generates 2 false positives a day. Over 100 cameras, that translates to 200 false positives a day. Let’s also assume that there are 5 real problems mixed in there. This translates to a signal to noise ratio of 1:40. For every 40 false events, there is 1 real event. The result? When the health check cries “Wolf!”, no one responds.

Next, I have video analytics on 20 cameras outside monitoring a virtual tripwire to indicate a perimeter violation. Here’s where it gets even more challenging. Next to and around the areas that these cameras monitor, I have branches swaying in the wind, animals (maybe even an occasional wolf) going to and fro, and the occasional early morning jogger who seems to like jogging along our fence. Therefore, on a daily basis, each camera is sending 50 alerts of a tripwire violation – that’s 1000 a day! And, on a daily basis, the perimeter is never violated. Not even on a weekly basis. Or monthly. Actually, last I remember, there was a break-in…two years ago? You get the picture.

So what is the solution to the problem of “The Camera That Cried ‘Wolf!’”? Unfortunately, many industry experts will tell you that even the best analytics and health check systems are still a ways off from effectively lowering the rate of false positives that they generate. That’s why ReTel is developing applications that will work right now with our proprietary two-layer auditing system to solve these problems and more. The broadest description of what we are developing would be a video noise filter that separates bad from good, returning only signal to the end user.

And our goal is not to replace analytics or health checks, but to make them better and more usable, so that they become acceptable features of an organization’s security and surveillance system. Bias is already creeping into end users’ opinions on analytics and health checks, which can make future adoption difficult – even after all the wrinkles are ironed out. That is unfortunate, because they can be truly useful tools to help manage an organization’s security and surveillance.

After all, when that real wolf comes, you want to make sure that there is someone there to hear the warning!