Tuesday, December 8, 2009

Choosing cameras using the DCRI method

Before you can choose the camera that will work for you, you will have to define your expectations for the camera. Take a moment and think about what you want the camera to do for you, and what the camera is actually for. Put simply, there are four operational levels use to classify camera performance, and they are detection, classification, recognition, and identification, known as DCRI.

Imagine two security guards looking at the feed from a camera on a monitor in a control room somewhere. Bob nudges Mike and points to the screen.

Hey, Mike. Looks like movement on camera six.

What is it, Bob?

Beats me, Mike. You think we should go check it out?

This is detection. Bob can see that there is an unnatural level of movement at the edge of the camera’s vision, and has been alerted that there is a situation that may require further attention. Bob stares at this particular camera all day, every day, and knows there is movement where no movement should be, but doesn’t know if he should be concerned.

As the subject moves closer to the camera, the next level is classification.

Hey, Mike. Looks like it’s a person, moving closer to the building.

Who is it, Bob?

I don’t know, but I don’t think anyone’s supposed to be in that sector, Mike.

At this point, you can tell if your subject is a vehicle, a person, or an animal. You can’t tell much about the subject, but you at least know what kind of subject you’re dealing with.

As the subject moves closer to the camera, the next level is recognition.

Hey, Mike. We have a white male, wearing a blue shirt and khaki pants, in Sector G… doesn’t look like our subject is carrying anything. Subject is walking at moderate speed. Go and check it out.

We can get some broad details of our subject, but the subject is still a little fuzzy. We cannot recognize unfamiliar faces. We cannot recognize scratches, dents, or other small distinguishing marks on vehicles. We can tell what type of clothing a subject is wearing and we can determine what body type a vehicle has. We may be able to recognize familiar people, but that is because we recognize people we know well using cues other than facial features. For example, we often recognize people by the way they walk, or the way they hold themselves, or their particular clothing or hairstyle. Keep this in mind when testing a camera.

The final level is identification.

(Speaks into walkie-talkie) Hey, Mike, this is Bob, I can see him on the camera now. It’s Joe, the new guy. Let him know he’s supposed to be in Sector F. Over.

This is the level of detail we need if we are to prove that a specific individual committed a specific action in a court of law.

In addition, identification can be broken down further to three sub-levels of performance: General, Forensic, and High. General, as we said, can be used to identify the facial features of an individual performing an action. Forensic is a higher level of detail, allowing you to make out small elements such as license plate numbers. High gives you an extremely fine level of detail, such as is required to make out the denominations of specific bills, and is mainly used in cash counting rooms, casinos, and in industrial processes.

How do we determine the distances involved in the DCRI system? It depends on two main factors, the performance of the camera and the performance of the recording device. In this example, we will assume Mike and Bob are looking at a high quality analog camera with a decent lens and good lighting, being fed through a DVR recording at 1CIF (pronounced “sif”), or 352 x 240 pixels, which is still the most common recording spec, as it allows the video to be stored most efficiently. Assuming a camera mounted no lower than 9 and no higher than 12 feet from the ground, identification ends at about 12 feet, recognition ends at about 24 feet, classification ends at about 32 feet, and detection will end at about 42 feet. This is not scientific, as many other factors will play a part- not least the visual acuity of the operator- but are rather rules of thumb.

Back in the olden days, security professionals had only one way to increase the chances of identification, and that was to use creativity to mount the cameras in spots where the likelihood of the subject coming close to the camera increased. While this is still the best and most effective method, it isn’t always possible. Technological advances have now given us another option- increase both the resolution of the camera and the resolution of the recording device. For higher end analog recorders, we now have the option of recording at D1 resolution, which is 704 x 480 pixels. This will give you identification at 22 feet, recognition at 32 feet, classification at about 42 feet, and detection at about 52 feet (at this point we get scalloping of the image due to using an analog monitor, so doubling the resolution will not double the DCRI lengths).

If we want to go digital, we have the option of using a new operational concept we call dPTZ, or digital pan-tilt-zoom. Using a megapixel camera will give us an image with more data in it. We can then take that image, and blow it up (after the fact), picking out details we could never see with the naked eye, just like you’ve seen on a thousand TV cop shows. Sure, megapixel cameras are much more expensive than standard cameras and they take up a lot of storage space. However, you can get better ROI (return on investment) by pairing megapixel cameras with rectilinear lenses. Rectilinear lenses are basically super wide angle lenses which use fancy optics so that there is virtually zero distortion. This allows you to cover enormous areas using a single camera, and because you are using megapixel, you can drill down into the image, and your Identification and Recognition thresholds are much, much higher.

Of course, we all want the highest resolution possible, but that gets expensive quickly. The smartest way to specify surveillance product is to 1) define the purpose of the surveillance system, 2) define what our expectations are for each camera, and 3) figure out the optimal placement for each camera. Remember, a low resolution camera in the right place is often all you need, and much cheaper than a high resolution camera that still isn’t close enough to see what we need it to see.