The problem isn't computing power, it's algorithm accuracy. If there are thousands of people who are a close-enough match to your face, then when they see your face in a random crowd at a train station, they don't know which of those people you are.
To fix this they'd probably need other heuristics, which for all I know they're already using. An easy one would be to narrow it down by location, though sometimes that'd be wrong. A more sophisticated method would be to build a model of human movement, keep track of locations where they have high confidence it's you (e.g. in pics of you in small groups with your friends) and combine these into an estimate of the probability of each candidate match being at the train station.
There are companies who are selling tech that will recognise a person rapidly walking past in a conference crowd (had it demoed to me), so yes the tech is definitely capable now
To fix this they'd probably need other heuristics, which for all I know they're already using. An easy one would be to narrow it down by location, though sometimes that'd be wrong. A more sophisticated method would be to build a model of human movement, keep track of locations where they have high confidence it's you (e.g. in pics of you in small groups with your friends) and combine these into an estimate of the probability of each candidate match being at the train station.