Megapixels for Facial Recognition

[MagnumVP 0]

I'm attempting to determine whether I need a 1MP (Axis P1344-E) camera or 3MP (Axis P1346-E) camera for my needs.

 

I need to be able to identify individuals walking by the camera at night about 30 feet away. The lighting around the area is standard street lighting.

 

At what distance does a 1MP camera need to be upgrade to a 2 or 3MP?

 

I'm sure this is a more complex issues with f and lux of the camera, but what about "rule of thumbs".

 

Thanks

[Soundy 1]

Distance isn't a factor as much as the area covered - how large an area will these people be in? If you're covering a door, gate, gap in a hedge, or some other area where you can limit where the person is (aka "choke point"), you can zoom in tightly on that area and get recognizable faces with pretty much any analog camera. If you're needing a wide lens to cover a larger area, then higher resolution may be needed.

 

Axis has a really good article on the topic here: http://www.axis.com/edu/identification/resolution.htm

[MagnumVP 0]

I have roughly 3 areas that I would like to monitor. Area 1 and 2 is roughly 30 feet across where the furthers is about 50 or so.

[mike_va 0]

I'm attempting to determine whether I need a 1MP (Axis P1344-E) camera or 3MP (Axis P1346-E) camera for my needs.

 

I need to be able to identify individuals walking by the camera at night about 30 feet away. The lighting around the area is standard street lighting.

 

At what distance does a 1MP camera need to be upgrade to a 2 or 3MP?

 

I'm sure this is a more complex issues with f and lux of the camera, but what about "rule of thumbs".

 

Thanks

I have both those. I prefer the 1344 day/night at night. The 3MP does not seem to help as much as you think at night. I would do 2x 1344 vs. one 1346. Also check out the 1602, which supposedly has great light sensitivity. Progressive scan helps a lot.

 

I've been playing around, here are 3 (cropped) shots of Mintron 1/2", P1344 at 1-30s, CNB BBM-24F at 1-60s - all at about 90-100 yards out. Dark except for IR. The Axis does not provide quite as much contrast but IF you have enough light will give a better picture at night.

 

Realize that you might want to apply a factor to any pixel calcs, most of those I think are probably daytime. I found I needed more at night.

 

With the CNB you can really see the IR reflecting off that car.

[MagnumVP 0]

I need to add some light to the side of my house. Even better I could invest in something that is more discrete like the AXIS T90A21

 

I watched the video of the 1602 and I'm just amazed at that quality at night. The Light Finder feature is something that I need to consider for sure.

 

It would be nice if they had the Light Finder technology in a Megapixel camera. Not sure if I want to wait a year or so for that.

[mike_va 0]

I have several Axis 850nm IR's, I would not consider them discrete. OTOH, my neighbors did not even notice them. Others have. Glowing red quarter size dots...work good though.

[MagnumVP 0]

A little duct/electrical tape to cover the glowing red dots and your GTG.

 

With your 850nm, how much light does it give you? I was just considering the AXIS T90A21 since I'm attempting to light up the area to the side of my yard (See image Zoomed). Not the entire area just about 20-30 feet down the road.

 

Do I need anything more than that?

[mike_va 0]

some more pics I save from last night, you'll want to be zoomed in more than this...

 

note the shadow cast by the IR

[mike_va 0]

A little duct/electrical tape to cover the glowing red dots and your GTG.

 

With your 850nm, how much light does it give you? I was just considering the AXIS T90A21 since I'm attempting to light up the area to the side of my yard (See image Zoomed). Not the entire area just about 20-30 feet down the road.

 

Do I need anything more than that?

 

Take the rating and cut in half. Also try to match the coverage angle to the lenses, helps minimize "white out". The Raytecs are cool in that regard, they have a new series that allows you to fit different "lenses" to the IR. Wider is fine too but you're just throwing light away. If I was doing it over I'd look at 940nm Bosch. I got a good price on my Raytec/Axis, generally around 200 on ebay new. I'm actually designing my own IR units now, bought the reflectors etc. Getting some alum cut and in a week or two should be good to go. I'm trying to hit some pretty extreme distances though.

 

I want twice the light you see plus covert...

[mike_va 0]

The other thing to consider is that is fine to go too big, as they have an power adjustment.

[MagnumVP 0]

Instead of going IR I may go down the the nearest Home Depot/Lowes and see about installing a motion censored floor light that would illuminate the area with enough light to bring it above 1.2 lux and trigger a color image. Maybe not motion censored, but maybe just a floor light that is on throughout the night.

[hardwired 0]

Instead of going IR I may go down the the nearest Home Depot/Lowes and see about installing a motion censored floor light that would illuminate the area with enough light to bring it above 1.2 lux and trigger a color image. Maybe not motion censored, but maybe just a floor light that is on throughout the night.

Motion sensor lights are nice, if you install it near to the camera (but not in the field of view, due to glare), when the light kicks on, people's natural reaction is to look towards the light, enhancing the chances of a good face shot.

[MagnumVP 0]

people's natural reaction is to look towards the light, enhancing the chances of a good face shot

 

That is what I am hoping for. Even if they are running down the street at least they will look up. I only need that split second of them be curious to catch them.

[Sawbones 0]

Visible light almost always beats IR.

 

Only go IR if you have a really good reason... because it's going to cost you 10x the expense to go IR, compared to visible-light floodlights.

[buellwinkle 0]

The rule of thumb for facial recognition that I go by is you need the face to be 80 pixels tall to indentify someone and of course that has to be facing the camera, no harsh shadows, adequate lighting. You can't say how many megapixel you need to achieve that, for example with a 3MP camera, 30' may be too far if you have a wide angle lens, yet 1MP may be fine at 30' if you have more of a telephoto lens.

 

For about the same price or less, check out the ACTi kcm-5211e, it's 4MP, has built in IR lighting, is autofocus (beats manually focusing on a ladder) and has an 18X zoom lens. You can have your high resolution, tweak the focal lenght from the comfort of your PC and even setup a zoom patrol if you want. I'll be doing a review on my blog of the camera in the next 2-3 weeks.

[Razer_SE 0]

For about the same price or less, check out the ACTi kcm-5211e, it's 4MP, has built in IR lighting, is autofocus (beats manually focusing on a ladder) and has an 18X zoom lens. You can have your high resolution, tweak the focal lenght from the comfort of your PC and even setup a zoom patrol if you want. I'll be doing a review on my blog of the camera in the next 2-3 weeks.

 

I'm installing a lot of these cameras right now actually with great results. Using them as plate capture cameras and they are working out very well for me. Using IR at night, in a pretty well lit area to begin with though but still results have been above my expatiation. Still tweaking to get the best results, but I like this camera.

[Stanislav 0]

Please see my article about person identification criteria, may be it will be helpful.

In the other article there is a comparison between megapixels vs shutter speed in low light.

[buellwinkle 0]

Interesting article. In Axis training, they say 40 pixels head size to ID a familiar person, 80 for an unknown person, so they are pretty close with their 35 and 83 pixels respectively. Of course the more the better, imagine at 35 pixels, someone's eye maybe 2 pixels tall, say 3 pixels wide, I would not call that very identifiable identifiable.

 

What happens too is people love sharp crisp images, so they crank up sharpening which takes surrounding pixels and increases contrast around the edge of an object, like an eye and it looks sharper from a distance, but actually loses resolution.

[Stanislav 0]

Interesting article. In Axis training, they say 40 pixels head size to ID a familiar person, 80 for an unknown person, so they are pretty close with their 35 and 83 pixels respectively. Of course the more the better, imagine at 35 pixels, someone's eye maybe 2 pixels tall, say 3 pixels wide, I would not call that very identifiable identifiable.

In reality there are many factors which have influence on the identification (view angle on the face, light, contrast, compression, lens resolution, noise, range of heights, shutter speed for moving persons etc). But it is difficult to take into account all these factors. Only the pixel density (spatial resolution) can be simple calculated.

Because of many influencing factors weren't taken into account, it can't be exact values of the pixel density for all possible cases. Nobody can guarantee that with 40 or 80 pixels per face the face will be identified. But with 80 pixels this is more probably.

35, 40, 80, 83 pixels are very rough empirical data. But we need the data for our work. To get more accurate result we have to adjust these values according to real conditions.

 

What happens too is people love sharp crisp images, so they crank up sharpening which takes surrounding pixels and increases contrast around the edge of an object, like an eye and it looks sharper from a distance, but actually loses resolution.

Sharpening can correct image blurring from insufficient resolution of lens. But it is difficult to choose exact level of sharpening. Therefore in many cases we see images with excessive sharpening. The excessive sharpening becomes an additional distortion which decreases possibility of identification.