Facial recognition, what is the real deal?

[ssnapier 0]

We have all seen the Hollywood version of this stuff, but what is the deal when it comes to using this stuff in the real world? How close does the subject need to be, are there minimum resolutions? How tough would it be to catalog people... let's say employees, or contractors... and setup alerts if one of those people entered an area?

 

With all of that said, what software/hardware gets the job done and what should I avoid? The 3VR stuff looks very slick, but I think I would like to look at something that I could add to an existing system instead of replacing it with an entire 3VR solution.

[jhonovich 0]

Getting facial recognition to work reliably in the field is hard, especially if you are trying to set up alerts on people.

 

Typically you are looking at 5 foot wide FoV for SD/analog or up to 10 foot wide for megapixel. These are much narrower FoVs than one normally uses in surveillance.

 

You also need to be concerned about sunlight and shadows. For instance, if there is a window behind the subject or a door is being opened to the outside, this can cause major problems. Details of the face may be sufficient for a human to guess a known person but not enough for a computer.

 

Moreover, angles are an issue. The subject needs to be positioning fairly directly in front of the camera. This means no 12, 13 or 14 foot high cameras (common in the real world) unless you are using a 50mm lens and trying to capture someone 50 feet away (uncommon).

 

You also hit issues with scale. The busier the facility being surveilled, the higher the likelihood of false alerts.

 

Face recognition is very impressive on a single camera controlled demo but has lots of operational issues and restrictions when someone tries to roll it out on the field.

[Sky1 0]

I don't have it on my video surveillance system but I do have it on my photos database program called PMB. I can select a photo of say, my wife and thumbnails of her from all other photos on my hard drive are previewed.

 

If I take a picture of someone who is already catagorized, recgnized and labeled will pop in there. Sunglasses or a hat can fool it easily.

 

Check out D.O.T.S, I put a post with a link on it here on this forum/

[ssnapier 0]

Getting facial recognition to work reliably in the field is hard, especially if you are trying to set up alerts on people.

 

Typically you are looking at 5 foot wide FoV for SD/analog or up to 10 foot wide for megapixel. These are much narrower FoVs than one normally uses in surveillance.

 

You also need to be concerned about sunlight and shadows. For instance, if there is a window behind the subject or a door is being opened to the outside, this can cause major problems. Details of the face may be sufficient for a human to guess a known person but not enough for a computer.

 

Moreover, angles are an issue. The subject needs to be positioning fairly directly in front of the camera. This means no 12, 13 or 14 foot high cameras (common in the real world) unless you are using a 50mm lens and trying to capture someone 50 feet away (uncommon).

 

You also hit issues with scale. The busier the facility being surveilled, the higher the likelihood of false alerts.

 

Face recognition is very impressive on a single camera controlled demo but has lots of operational issues and restrictions when someone tries to roll it out on the field.

 

This is EXACTLY what I needed, thanks!