You are reading a single comment by @cg5154 and its replies.
Click here to read the full conversation.
London Fixed Gear and Single-Speed is a community of predominantly fixed gear and single-speed cyclists in and around London, UK.
This site is supported almost exclusively by donations. Please consider donating a small amount regularly.
@cg5154
I didn't say it would be easy... as you said, image recognition is extremely difficult, it's a big topic in computer science research. But I'm the sort of sick puppy who'd enjoy attempting this sort of thing.
At first blush the problem space is fairly small and isolated. I don't need software that will recognize all types and situations of theft. Just the theft of one particular bike (mine) from one specific area (the row of U-shaped whatchamacallits opposite my office).
The software needs to be able to recognize which bike is mine, first of all. If it can do that, then detecting when the bike has been moved more than, say, 0.5 m should be easy. Unfortunately, if somebody's moved it 0.5 m, that means they've gotten through the locks already and I can kiss it goodbye.
So I'd want something to recognize that the theft is in progress. Maybe a human being standing/crouching within a certain distance of my bike for more than some time limit? Or even better, somebody directly handling the locks on the bike?
False positives would be a problem, but given a choice between false positives and false negatives, I'd choose the former. Perhaps I could come up with some sort of statistical method for estimating the likelihood that this is a real theft... for instance, if the person standing near my bike is wearing a hoodie or a tracksuit, then the probability this is a real theft is 99.9%. Given that I work in a somewhat posh area in Belgravia, this is probably a decent assumption :-)
Mind you, I don't actually know anything about image recognition at the moment, so I could be talking utter bollocks...