Facial recognition, pixels/m

[robert 0]

Some say you need 120pix/m, some say 140. What do you think? Also - it depends on camera quality, doesnt it ?

 

Also - 120 pix/m - meant from where? From camera with 2.8mm or 10mm lense?

 

Any calculators for this ?

[thewireguys 3]

Here is some good info on Arecont's web site. I plan on proving this when I have some free time.

 

http://www.arecontvision.com/uploads/file/Arecont%20Camera,%20Lens,%20and%20Pixel%20Density%20Chart%20-%20English.pdf

 

http://www.arecontvision.com/uploads/file/Arecont%20Vision%20Pixel%20Density%20Calculator%20-%20IMPERIAL-ENGLISH%20UNITS.xls

[3RDIGLBL 0]

Great links.

 

I have an argumentive question for debate.

 

What determines each category? What determines how many pixels it takes to read a license plate? Is there a clarity filter of some sort or is it just that someone sits there and keeps increasing the pixel density until they could read it. Once they read it do they keep increasing the density until it becomes clear? I mean I just don't understand this very clearly I guess.

 

Is the pixel density listed the least recommended for each category?

[ak357 0]

Great links.

What determines each category? What determines how many pixels it takes to read a license plate? Is there a clarity filter of some sort or is it just that someone sits there and keeps increasing the pixel density until they could read it. Once they read it do they keep increasing the density until it becomes clear? I mean I just don't understand this very clearly I guess.

Is the pixel density listed the least recommended for each category?

 

I think nothing but common sense

have a look at few samples

keep in mind i took them from google

look at the white numbers on blue background for pix count

 

[zmxtech 0]

Its only a piece of the puzzle, things like >

 

Camera res

correct lens ?

shutter speed ?

lighting sunlight ? IR ?

filters ?

angle ?

DVR res

DVR compression

 

tons of factors for a good picture -if your missing any of these "the number plate" or face is gone.

 

Number plate capture is a specialized field like black magic.!

Most states have different types and colors of plates -and people put covers and mud on them.

 

but to answer your N-plate question all the good readers have 16mm + lenses to get in close and synced IR illumination and very heavy IR pass filters. They make a bright headlight look like a candle. and the plate a perfect picture. [the regX is a good example]

 

If its a well lit area a mega pixel camera with the right lens and angle can do a better job again -getting the plate-driver and car !

 

Even if the plates are fake you can see the driver and read the rego sticker !

 

my 2c

[robert 0]

Excatly. But whats average for number plate and face recognition?

[zmxtech 0]

Really moot these days with Megapixel and its only at the far end of the scale.

Guess im a bit elitist because I only deal with mega pixel these days and don't waste time with low res gear.

 

A color cam/IP/ANPR all have different pixels and sensitivities

 

where a color cam might need 200 the ANPR cam can do it with less because of the sharp picture and contrast. At night different again.....

Im sure the boffins have a formula

 

You need to separate faces and plates. [plates are retro-reflective for a reason] >to grab $$$ from us. 61 in a 60 zone shame on you !

For example 'faces' these days have hoodies and BB caps so its the angle of the cam not the pixels.

 

I would not design a system that has a narrow margin of pixels that's the old analog problem really. [where there was 300] even those days a tight shot on the choke point did the business.

 

If your worried about pixels you have the wrong lens or the wrong camera.

 

i feel all pixelated....

 

[megapixel man 0]

Excatly. But whats average for number plate and face recognition?

 

The average to get both lic plates and faces is 50pix/ft or 164pix/m.

 

You can start getting facial rec from around 125pix/m and lic plates from 130pix/m if you have a lossless compression.

 

Compression levels really do matter for what or looking at, you may be able to use a higher compression level for faces (Soft edges) whereas you would want low compression on lic plates (hard edges).

[zmxtech 0]

dont agree with any figures they would only suit one type of camera/lens/DVR/lighting scenario. At night it would be different and ANPR filtered IR systems way out here...

Any lens change or even angle will throw that formula out easy.

 

for sure no compression, again they are all minimum limits with mega pixel who cares ! I used to run no compression on some systems to get every tiny bit of data. >in the analog world.

 

I agree to baseline a camera system yes talk pixels....

 

In a properly set up system with a choke point and the camera strait on

with the correct lens & lighting you should not have to worry at all about pixels. should be able to tell the time from the guys wrist watch as he drives past..

[megapixel man 0]

Forget ANPR, it is a different science with OCR "optical character recognition"

 

Facial rec and lic plate rec is often refered to in CCTV standards in EU and AUST and I am assuming this is what Robert is refering to.

 

The following chart gives a guide to pixels on target.

 

 

 

[@cctvgeek 0]

Thanks for the compilation of six images. Interesting to peruse.

 

I've created for myself an image of a UK number plate at 164px/m. This is 85px wide (164 x 0.52m) and the end result looks much more like your image labelled 98px/m than it does the one labelled 164.

 

Thoughts welcome.