When we won the ultimate one-pixel-target detection fight

 By P. Bernas & F. Alvez, Inventors.

First developed in 2003 for the Ministery of Defense, EVITECH's Video Motion Detection (VMD) solutions have rapidly demonstrated in 2004 and 2005 their ability to detect and track even very very small and almost invisible targets in the rough, as well as in urban situations. For example, when observing a road in a foggy situation, the EVITECH Eagle video motion detection software was able to detect a several pixels character almost invisible with the human eye on the screen. Also at night, such detection ability proved to be very helpful in dangerous situations (even for 3-4 pixels targets).

The key fact is there : if some bad guy wants to rob your site or make a terrorist action, he will prepare his job first. He will send an animal or a kid balloon in your fance, he will tackle your lighting or your cameras, in order to blind you (or point them out), he will break a very small wire from your roasting..

All these small actions are riskless for him (almost no penalty for playing football on the side of your site), and will provide him measures about your ability to react and maintain your protections active : Did you notice him ? How fast do you drive to the cut fance ? How long do you take for changing a broken light ? How many lights spare do you have ? Do you detect a camera change ?

This is the point where you need tough video motion detection solutions. Once the light is off, at night, once the camera is turned toward the sky, how shall you protect your site from an intrusion ?

EVITECH has done the job for the video motion detection part. We cannot improve your procedures, we cannot repair your fance nor change your lights, but yes, we can detect at night when your light is off (since 2005). We also detect all kinds of camera attacks that your site may suffer (since 2006) : blurring the camera glass with a polish spray, blinding the camera, moving the camera towards some other landscape, cutting the camera wire, etc... Even an event as small as one image loss in the image flow can be detected with our video motion detection solutions and can alert you (like when plugging in a fixed image on the CCTV wire and turning off the camera).

Now in 2008 comes the final fight : the one-pixel-target detection. The question has raised from a military project held in 2007 : when it is possible to detect and track a 2-3 pixels target with EVITECH's video motion detection, how is it possible, when falling to 1 pixel, to have a stable detection ? One pixel targets also include "sub pixel targets", which make their importance. On a 20° FOV camera, all human targets over 900 meters (half a mile) are sub pixel targets, even at one mile or more. If they are almost not visible on a colour image, a thermal camera can make them slightly visible. However, our aim has always be to be able to detect any visible target, 100% of the time,with our video motion detection solutions, like a human would detect it when looking carefully at the video.

The case was difficult, let's say like a clear one pixel shadow over the details of the landscape in a greyscale image. A one-kilometer-away ghost. From a Flir thermal uncooled camera covering a 0 to 1 km rough corridor. One grey pixel over 320x240 all grey levels glittering pixels. Plus one NUC operation resetting the image every 60 seconds (this is specific to thermal uncooled cameras, it's an additional trick).

One additional difficulty for video motion detection is what is called "noise", in the image. On a given landscape, observed with a stable camera, you don't get constant values for pixels even observing a still object. This is because of sensor noise. Sensors receive photons, on very short time frames (under 40 milliseconds, at 25 fps), and count them in order to establish a grey/colour level. On thermal cameras, these are "thermal" photons ("heat" photons that you can't see). Sometimes, in difficult situations such as cold weather, there are not enough photons for overcoming the statistical effects of photons counting. From one image to the next one, pixels get a different number of photons, and then pixels values are unstable. Possibly of 5% unstable, but this can also correspond to the effect of the shade you are looking for.

We had to develop very strong statistical analysis tools, inside our video motion detection algorithms, in order to overcome this difficulty, we investigated a wide area of probabilistics such as Bayes laws, Markov chains, ... The other difficulty is to do the job at real time. Complex theories lead most of the time to complex computations, which take long processing time, when called on the whole image... We searched for a long period. But on the nice day of April 6, 2008, after an incidental remark from one of our customers, who noticed a behaviour difference between two versions, we found it !

Even by hot weather (30°C here in France in May) we can demonstrate far detection (here two pedestrians at 500 meters in a test realized with a FLIR systems SR35 camera, lended by Alfa Visionics, the french FLIR systems reseller).

All this with no more false alarms rates than we had before (~1 per day and per camera when correctly set).

Call us, if you want to borrow us one video motion detection system for a check !

P. Bernas & F. Alvez.