very interesting.. lets see how others would comment on this.

Well not a webcam... it would be too wide-angle to get an accurate reading of the pupil unless it was positioned RIGHT IN FRONT of someone's head. Maybe there's some easy way of mounting one like that.

I think the plug-in would be tough to write, because you might see that the eye is looking left, or looking right... but you don't have any clue what the eye is looking at. You'd need exact positional data of...
- Person's head
- Monitor size
- Monitor distance
- Monitor height

...just blah blah blah details you'd never be able to collect in a convenient fashion. And you'd probably have to lock the person's head in a vice from that point forward too.

Not to say real eye tracking facilities lock people's heads in vices. But they're not picking up the data with webcams.

I like the website I saw where they register the mouse movements. Don't know the name of the site, when I find it I will place it here. When we are able to analyse the eys movement, we can throw away our mouse as well :-)

I don't understand what kind of data you're actually trying to collect here...

I think they are trying to collect this kind of data
http://images.google...038;sa=N&tab=wi

They don't use normal webcams.

sort of Iris reading?

Iridology something?

Update:
-Sites buy eyetracking research to improve their usability and conversion rates.

-Eyetracking camera's are very expensive (think $10.000)

-It is possible to do eyetracking with a normal webcam by callibrating. Than the application asks you to look at several dots and analyzes how your eyes move to be able to track your eyes.

@Gordon and Kevin: have a look at the paper http://www.people.co...4/PedersenSpivey.pdf
They make use here of normal webcams... (like I said ;) )

A 'trick' they use is to record video images that are processed 'offline', so not on your own computer.

This may be very hard to accomplish, but you have to realize that the rewards are very lucrative. Imagine that regular eyetracking research pays for testpersons to come to a test site with expensive gear.

You're right I didn't read the paper and it IS very interesting.

That they do the processing off-line isn't the issue (they probably haven't worried about optimizing yet). Its just accuracy. I'd have to check out one of those commercial products they mention to see if its usable in the real world... because webcam rez is so low, to get real results I suspect even the commercial packages need to be used in very controlled environments.

I'll move my vote up as its more feasible then I first assumed.

All webcams are not the same low resolution. You could pay people with better cameras a higher amount to participate. Because they would be providing better data.

I don't have a clue about the technology, I can see how website owners would be interested. Very interesting. I like it.

Thanks for clarifying what information you're after, and thanks for the technical details. It's nice to find someone who's actually got as far as thinking about the technical feasibility of the idea. However, duplicating the system described in Pederson and Spivey is going to be a challenge. Anyone up for some nice neural network programming?

Most of the professional system I have seen use: multiple IR lights top and bottom, IR cameras, a normal camera for recording the subject’s perspective and at times a camera to record the facial expressions.

It is not just a single web cam.

Thx Kevin_Cox, that's even more hardware than I was expecting.
LarsBell, agreed. But 640x480 is the max for high quality. I've tried 1.3 megapixel webcams, and currently they are crap. Too much video noise.

I really think there is SOME info that can be gleamed from webcam tracking of eye movement, but in the near future, it will never be processable into X-Y screen coordinates.

As Goosie points out, Many people move the mouse cursor as they read, and there are some tracking apps for that. Maybe a browser toolbar which performs a heck of a lot of data mining could be effective. But that sounds like a Y2K biz model, not a 2007 one. And once you lose the heads-in-front-of-webcam, then people will try game the system with macros to collect whatever incentive you are offering.

No to say that you can't gather some data. It is just very vague and not very good compared to a professional system.

About mouse movement: http://www.clickdensity.com

You may have privacy issues with this. Also, not everyone has the "automatically install plug-ins" feature turned on. I think it may be a hit and miss thing and more of a miss, especially when people find that you are "watching" them.

I think the mouse tracking would also be more of a miss. Who moves their mouse cursor as they read a web page? If anything, I move it right out of the way so I can read what's on the page.

With all due respect: it sucks that all negative comments of people are so dominant. They overshadow the improvement of the idea.

Quotes like: "it will never be processable into X-Y screen coordinates" are just untrue (do a search on youtube) and don't add much to the idea.

I don't care that people think my idea sucks (maybe it does), but I think it doesn't even get a chance here. What's the added value of putting it on CH?

I strongly believe in the idea.

I have had the same idea a few years ago and I carried investigation on the topic : it is possible to track eye movements with a standard 640x480 webcam. There are *MANY* research papers available on the topic (I can share them on request).

The requirement is to distinguish Areas of Interest on a webpage, and an overall resolution of at least 16x16 different regions on a screen is already meaningful enough in most applications.
With a few technical tricks, it is possible to achieve a 32x32 resolution, and with other tricks, 64x64 is achievable.
This is way enough to find out about areas of interest on most webpages.

The other requirement is to be able to detect a fixation of at least 100ms, and with a 30fps cam, that gives us at least 3 images to work on. Given the fact that most webcams decrease framerate to deal with poor lighting conditions, let's drop this to 2 images.
It means that a regular 640x480 webcam, at a distance of 50cm from a 17" screen and the optics which makes the face cover around 40% of the screen, gives us enough pixels for two eyes to detect 64x64=4096 different zones on a screen with a temporal resolution of 100ms, using the proper signal processing.

The processing involves head 3D modeling to compensate head movements, and different algorithms to detect the eye movement.
Some promising results have been achieved using Starburst algorithm, and some other smart signal processing.
It is even feasible to avoid user calibration.

The technical solutions are out there already, waiting to be implemented.

Too, given the low cost of the study, it is possible to reach a large size of sample (=people looking at webpages).
It is thus possible to increase the accuracy of the results.

Other leads to make the results even more meaningful and useful to designers are:
- to conduct a study using a proper scenario
- to carefully select the sample of crowd and conduct statistical study on their gaze behavior
- to conduct a multivariate analysis on Areas of Interest
- to process the raw gaze data to rank the AoI, give a scanpath, compare relative attention, etc etc ...
... etc

There IS $ to get here guys.
I'm working on it, and you're welcome to join ;)

I think all the negative comments are exactly right that you will never get the same quality of data with a single webcam that you can get with a professional eye tracker setup. However, I don't think that decreases the value of the system, you just have to recognize those limitations up front and target this project to address the problems it's capable of solving. I think the negative comments are an important part of defining that niche.

I was just reading a paper detailing a usability study of a website. It was a very broad study so they didn't have the ability to do full eyetracker studies of each participant. Basically they were using server logs as their primary source of data to measure how much time each user spent on each page with the goal of determining how much information the users were able to process. If the users' visits were very short, they assumed they didn't process much info and moved on quickly. If the users' visits were very long, they assumed they got distracted by something else. Both of those assumptions are legitimate given the type of data the study had to deal with, but they both introduce substantial margins of error.

A system like the one described here could provide a very good measure of how engaged the user was on each page. Did they scan the page and actively process info? No, you won't get specific eyetrack data on what they read and what they dwelled on, but you will get an overall measure of whether they were processing data or sipping coffee. That provides an incredible amount of data for a study whose only alternative is timing data from server logs.