6 ms·

I'm skeptical, especially of the surveillance claims. If you already know who you're looking for and you have lots of footage of that person, you might be able

I'm skeptical, especially of the surveillance claims. If you already know who you're looking for and you have lots of footage of that person, you might be able to train your system to pick up that person's gait from new footage. But that's a far cry from being able to "identify anyone" by their gait, i.e. set up lots of cameras around the city and track individuals coming and going. There are already amazing airborne camera systems that track the motion of every license plate in a city, and even that isn't the same as tracking every car's movement because you still have to know which license plate to follow -- you just can't care about all movement of all cars. I suppose someone will say that you can store all movement in a database, but I'll wager that's just not worth anywhere near the cost. Your computing power would be much better spent on computing large primes. Anyway, it's a cool sci-fi movie premise. You can see the murderer shaving off his mustache, burning off his own fingerprints and then knee-capping himself to avoid detection.

nindalf 8y ago
> you just can't care about all movement of all cars Of course we can. For each license plate encountered by CCTV or cop cars, store a tuple of (timestamp, coordinates). This won't require much disk space, so that's not a concern. Later, it is easy to answer questions like "what places was this car seen and when?" and "which cars were in this vicinity at this time?"
- wwarner 8y ago
  Well, let's test your claim. Say there are 1 million cars in a city. Say your camera system covers 300 square miles (to handle nyc). The question is how long would it take to scan the images from a single moment and capture all license plates. At 5 seconds per plate, that's 57 days. Plenty of time to commit a crime and escape without detection. And again, you could've been doing something useful like capturing some large prime numbers, in order to break the encryption of someone who is actually trying to hide something.
  - somebodythere 8y ago
    Why would it take 5 seconds a plate? They could as well station a licence plate reader at each intersection, and update a key-value store with a given plate's last seen location whenever one is spotted. O(1).
    - wwarner 8y ago
      Parallelization you say? 5 seconds is constant time, and a pretty optimistic one. How many license plates are visible from any given intersection in my town, NYC? Somewhere between 5 and 50. There is always motion, so what should the frame rate be? You probably can't commit much of a crime in 15 seconds, but you can drive an 8th of a mile (at 30 mph), about two intersections. So let's settle on a frame rate of 5 seconds. You still won't have much precision with cameras only at each intersection, so let's put them 10 feet apart, and let's have 4 at each point. Honestly, that's a minimum. At 10 feet apart, for NYCs 6074 miles of road, that's 12,828,288 computerized cameras. That's about $12B. And I still bet that the error rate will be too high to be useful. I'm not arguing that it will never be possible, I'm arguing that the original claim, and the mass surveillance that make people nervous, is not happening today.
- JoeAltmaier 8y ago
  There're already drones flying over American cities that record the city from the air nonstop. When a crime occurs, they review the recording in reverse and see where the bank robber etc drove from (their house), then go to their house and arrest them. No need for anything automatic.

nindalf 8y ago

> you just can't care about all movement of all cars Of course we can. For each license plate encountered by CCTV or cop cars, store a tuple of (timestamp, coordinates). This won't require much disk space, so that's not a concern. Later, it is easy to answer questions like "what places was this car seen and when?" and "which cars were in this vicinity at this time?"

wwarner 8y ago
Well, let's test your claim. Say there are 1 million cars in a city. Say your camera system covers 300 square miles (to handle nyc). The question is how long would it take to scan the images from a single moment and capture all license plates. At 5 seconds per plate, that's 57 days. Plenty of time to commit a crime and escape without detection. And again, you could've been doing something useful like capturing some large prime numbers, in order to break the encryption of someone who is actually trying to hide something.
- somebodythere 8y ago
  Why would it take 5 seconds a plate? They could as well station a licence plate reader at each intersection, and update a key-value store with a given plate's last seen location whenever one is spotted. O(1).
  - wwarner 8y ago
    Parallelization you say? 5 seconds is constant time, and a pretty optimistic one. How many license plates are visible from any given intersection in my town, NYC? Somewhere between 5 and 50. There is always motion, so what should the frame rate be? You probably can't commit much of a crime in 15 seconds, but you can drive an 8th of a mile (at 30 mph), about two intersections. So let's settle on a frame rate of 5 seconds. You still won't have much precision with cameras only at each intersection, so let's put them 10 feet apart, and let's have 4 at each point. Honestly, that's a minimum. At 10 feet apart, for NYCs 6074 miles of road, that's 12,828,288 computerized cameras. That's about $12B. And I still bet that the error rate will be too high to be useful. I'm not arguing that it will never be possible, I'm arguing that the original claim, and the mass surveillance that make people nervous, is not happening today.
JoeAltmaier 8y ago
There're already drones flying over American cities that record the city from the air nonstop. When a crime occurs, they review the recording in reverse and see where the bank robber etc drove from (their house), then go to their house and arrest them. No need for anything automatic.