What is video analytics ?

Since the late 90s, the digitization of content and the progress of computer CPU power has made possible the real-time processing of video images to extract interpretations (what do we see in the image, what's there, who goes where, etc.). First performed in black and white, and improved later for color information, these algorithms have begun to emerge from research laboratories in this period.

They led to build workable solutions, primarily for monitoring industrial processes in controlled lighting, then road traffic, later for monitoring persons and objects, and more recently for more sophisticated applications such as target classification, on fly face detection or facial biometrics (identification). Technologies behind these solutions come from very different approaches, and are difficult to mix because of the yet still limited processing time. The field is now moving forward and split in multiple applications for Global Security.

The issue is critical: it is the automatic interpretation of video content (image information) that enables the increasing use of CCTV cameras on a large scale. Manual surveillance of dozens of screens changing scene every 10 seconds has demonstrated to be completely useless: nothing is anymore seen on screens by a human after one hour. Only video analytics makes it possible to deploy usefully multiple surveillance cameras: information is filtered, analyzed, sorted, and only relevant alarms are reported to the concerned security officers. The video walls that were installed in years 80 are now to be removed.

As an example of the specificity of the algorithms, license plate recognition relies on some on fly shape detection (plates), with  improved optical character recognition techniques, enriched with a knowledge of license plates over all the available countries that bring consistency rules, for filtering wrong inscriptions that are not plates, and possibly help recognizing some plates by the knowledge of the expected pattern.

Likewise, detection of abandoned object has some proximity with move detection, but detection of an abandoned bag behind a crowd that continuously hides and uncovers the bag needs more specific algorithms. 

In the field of video analytics related applications, many players have emerged and propose very heterogeneous offers.

We at EVITECH have concentrated our efforts since our first developments in 2002 on the intruder detection application. An intruder is a visible target (visible on the image of the camera) that moves in the image and goes from outside a site toward inside the site. Born from a military project, EVITECH's first concern was -and is still- to answer at the best reachable level to this question "is there somebody in my backyard ?" 

We wanted to guarantee 100% detection of any intrusion which would be visible on the image, whatever the size of the human (inside the covered range), whatever the speed, whatever the light conditions.

Because it was a real challenge to guarantee this, it took time to reach it : we developed specific algorithms for night accommodation of sensitivity, for fog situations, for long range with noisy cameras, etc. We recorded hours of videos in very various conditions (countries, cameras, weather, seasons, ...), in order to have a large basis for progressing in the algorithms.

But finally, we succeeded and we are now able to guarantee 100% detection of any on-screen visible target.

Because the Operating System, as well as these algorithms consume an appreciable CPU time and computer memory, it is necessary from time to time to stop the system and restart it. This short break, even if it takes only one minute every 3 months, leads us to a detection probability which is something like 99.99... %.

From this basis, we developed two main offers :

• Jaguar, our civilian product,
• Eagle, our military product.

Once this was reached, we explored the other possibilities of applications that would be accessible with the technology :

• immobility detection,
• reverse walk detection,
• speed estimation/control,
• travel path analysis (duration, length, etc) for loitering detection,
• real size estimation and control,
• behavior analysis,
• fire start detection over thermal related cameras,
• oil and gas leaks detection,
• ...

and, after validation, we decided to propose a unique video content analysis solution that would enable simultaneously all these applications in one sole package. This choice was guided by the idea that the user should have control on every parameter of the solution, and freedom to mix them, so he keeps the hands over his configuration, and can reach the best performance, as well as manage complex controls, such as detecting a car that drives slowly but does not stop for a control at a barrier.

Furthermore, the idea was to provide the user with a report, fully automatically, which enables him/her to reproduce a configuration on a virgin product installation.

Later, we worked on false alarms rates, in order to lower the false alarms rates at their minimum. We have now written a camera installation guidebook, which, when taken into account, enables our users to reach less than 0.5 false alarm per day.camera.

Here is a white paper (in french) on video analytics : Evitech_AViRS08_04.pdf