4 ms·

Are you saying that the lack of redundancy is as a "major advantage to vision-only?" Because integrating those redundant systems (sensor fusion) is hard? > all

Are you saying that the lack of redundancy is as a "major advantage to vision-only?" Because integrating those redundant systems (sensor fusion) is hard? > all of your sensor systems need to notice a danger...before you can react I'd be really surprised if they implemented it this way. I thought the whole pitch behind lidar was that if vision doesn't see the person walking in front of you, lidar can still see the moving person shaped thing and you can still act.

letitbeirie 5y ago
You're still left with the two-sensor problem though: if lidar says "pedestrian" but vision says "clear," you don't know if it's a false positive in lidar or a false negative in vision. If your safety reaction is consequence-free, no problem. If it isn't (like slamming on brakes in traffic), you really want a 3rd sensor so you can do 2oo3 voting.
- misterdabb 5y ago
  Idk why people have the preconception of two (or more) sensors each deciding on an outcome indepently.. Suppose you have a color sensor and a form sensor and you have to identify an orange (fruit). Clearly in conjuction the two sensors will be way better, than assigning everything with color orange orange and everything round an orage.
  - letitbeirie 5y ago
    It's for safety reasons so that the system doesn't enter an undefined state if one of those sensors fails, which it eventually will. If the signal goes from "black nothing" to "orange round" you know you went from no object to an orange, but what if your form sensor breaks and it goes from "black nothing" to "orange nothing?"
    - misterdabb 5y ago
      You just made a point for redundancy ie more sensors... Sensor failures are independent events and need to be correctly identified no matter what. To take Lidar and Vision, both say something about distance and form of objects. Together they can achieve better performance. If one of them fails (and failure is identified) it defaults to the other sensor only and will be somewhat worse, but should at least allow it to pull over and warn you about sensor failure. Also failures are much easier to identify when you have a baseline reference of one or more other sensors, in isolation much harder.
- 6gvONxR4sf7o 5y ago
  What if you have multiple cameras and multiple lidar sensors? You get into statistical modeling immediately and I imagine that’s how the production systems like these work. If 3 cameras and 2 lidars detect a person, the model says X, which translates to such and such FPR and FNR.

letitbeirie 5y ago

You're still left with the two-sensor problem though: if lidar says "pedestrian" but vision says "clear," you don't know if it's a false positive in lidar or a false negative in vision. If your safety reaction is consequence-free, no problem. If it isn't (like slamming on brakes in traffic), you really want a 3rd sensor so you can do 2oo3 voting.

misterdabb 5y ago
Idk why people have the preconception of two (or more) sensors each deciding on an outcome indepently.. Suppose you have a color sensor and a form sensor and you have to identify an orange (fruit). Clearly in conjuction the two sensors will be way better, than assigning everything with color orange orange and everything round an orage.
- letitbeirie 5y ago
  It's for safety reasons so that the system doesn't enter an undefined state if one of those sensors fails, which it eventually will. If the signal goes from "black nothing" to "orange round" you know you went from no object to an orange, but what if your form sensor breaks and it goes from "black nothing" to "orange nothing?"
  - misterdabb 5y ago
    You just made a point for redundancy ie more sensors... Sensor failures are independent events and need to be correctly identified no matter what. To take Lidar and Vision, both say something about distance and form of objects. Together they can achieve better performance. If one of them fails (and failure is identified) it defaults to the other sensor only and will be somewhat worse, but should at least allow it to pull over and warn you about sensor failure. Also failures are much easier to identify when you have a baseline reference of one or more other sensors, in isolation much harder.
6gvONxR4sf7o 5y ago
What if you have multiple cameras and multiple lidar sensors? You get into statistical modeling immediately and I imagine that’s how the production systems like these work. If 3 cameras and 2 lidars detect a person, the model says X, which translates to such and such FPR and FNR.