Projects like those are always ongoing. The problems are extremely difficult.

The basic issue is that the input to the program is video — even relatively low resolution, security camera video represents an enormous amount of information. Say you’ve got 320×240 video, with just 8 bit grayscale at 15fps. Poor quality video, but it’s still more than a megabyte per second. As input data, you’ve basically got a 3-dimensional array that gives you the pixel intensity at a given screen coordinate in a given frame.

Suppose you want to look at 5-second segments of video to extract something like a person’s “walk”. You’ve got a 40 megabit input vector representing that 5-second segment and from that have to determine the gait. It’s a huge problem.

So, typically, the research efforts go into various strategies of picking out “features” — they effectively compress the multi-megabit input vector down into some vastly smaller space. The gait-recognition system you linked to represents the gait as three parameters, which they call an “eigengait” (the process by which they extract it basically boils down to finding eigenvectors in a really big vector space).

If you’re looking for a specific person, with a known gait, you’re almost done—you just run this algorithm against lots of clips until it finds one that presents a gait close enough to the subject’s known gait.

On the other hand, if you’ve got a clip of an unknown person, who you’re trying to identify by gait, you’ve got a whole new problem at hand. You need to have compiled a database of gaits to search.

They’re fascinating technical problems, but they aren’t like throwing together a web page, that’s for sure. They require a lot of specialized knowledge and virtually all of it ends up as custom software that’s really only usable for the specific project at hand.