Webcam Stereo Vision Pt.1

A pair of Quickcam Pro 4000s arrived yesterday, and despite the knowledge that they would work simultaneously, I spent most of last night grappling with them (in between reading bits from The System of The World).

After a bit of experimentation, I realise now that the 8.xx-series of Quickcam drivers are terrible; try as I might, I could not expose any controls other than 'auto'. I'm not sure if installing the whole Logitech suite would fix this, but there's no way I want all that junk on my system. So back to the good old reliable 7.30 drivers I went (nicely still available on Logitech's FTP server).

These drivers aren't without some problems - with the exposure set really low, bright colours and their surrounds become greyscale, whereas what surrounds that remains in colour. Probably not so great for colour tracking, but I'll worry about that later. Further, for some reason, at random, one of the cameras will drop to 2fps. Not that useful, and again I'll have to spend some time figuring out if that situation can be recovered from without a reboot as installation stuff generally has to be able to run for around 12 hours without breaking. For the moment it doesn't matter too much, as the shutter glasses tend to stop working after an even shorter interval, and that problem most definitely does require a reboot.

But when the suns and moons are aligned and things deem themselves fit to work properly, then it works even better than I'd hoped - with many thanks that the QCP4000 doesn't suffer from the same terrible barrel-distortion that afflicted its predecessors, as well as a bit of trigonometry, pretty accurate marker position estimations can be obtained. It's like System B1 but in 3D.

Unfortunately, it all falls to bits as soon as lateral motion is introduced; as I'd expected and forgotten a while ago, one camera almost imperceptably lags behind the other due to minute differences in startup time and so on. The end effect is that one image moves ahead of the other, the horizontal distance between the two marker images changes, and the depth estimate then borks up completely.

How this is counteracted will depend entirely on the nature of the sync issues - should it be constant, then our life is made easy as once the value is known, we can project backwards in time so that the image that is ahead is back where it should be. Should it be random, then we will yell and shout and probably use a whole bunch of smoothing 'til it works. Either way, it will be documented here.

Sidenote: everyone should all be getting full-feeds, if you're seeing truncated posts, it's possible that your feedreader (bloglines included) is subscribed to the old feed and is experiencing some problems with the redirection. Unsubscribing from that one and subscribing to the new one should fix it.