Of course at the time the computing power needed just to do the image tracking was far in excess of what could be carried on his person, so it involved a (possibly pre-WiFi) radio link to a lab network of graphics workstations, and as far as I know the software wasn't doing any kind of AI ad identification, but only matching pre-tagged ad images (or maybe just tracking the physical locations of the user vs the known location of the ads, via GPS + INS + video tracking).
It was nevertheless an exceedingly impressive demo that it has taken quite some time to make a significant improvement on.