5 ms·

fog / rain / low visibility (potentially fooling the sensor into thinking there's an object in front of you), cars behind you?

beamatronic 10y ago
That raises an interesting question: It is 2016, do we ( or do we not ) currently have economically feasible technology to place into cars that can reliably and accurately determine when you are about to hit a solid object or not?
- mapt 10y ago
  Define "Solid object". In the lab, opaque textured objects in sufficient lighting are pretty easy to localize by a variety of methods. Reflective objects are quite a lot harder, but there are methods that work. Reflective objects through distortion like rain/fog... harder, but often doable. In low light? Depends. The weird thing with environment modelling in computer vision is that there are a lot of overlapping methods, but sensor fusion of these methods is rarely attempted. Most robots end up with a pair of laser rangefinders or a pair of cameras or something equally simple to program. Creating a high-reliability automated vehicle that can bypass all the edge cases where these things fail is another thing entirely. You want overlapping, inter-reinforcing models of the world, with algorithms that not only identify items in a well-calibrated system, but calibrate the system in real time based on items identified. My lab was really excited about replacing stereoscopic computer vision with monocular computer vision for 3D modelling, but the opposite direction is where you want to go if you care about as many 9's of reliability as possible and hardware is cheap. You want a few dozen cameras with fixed orientation, you want parallax laser rangefinders in a variety of bands, you want time of travel LIDAR units, you want a {GPS + 3DoF gyroscope + 3DoF magnetometer + 3 DoF accelerometer + barometer + thermometer}†, you want ultrasonic and radar rangefinders; You want everything you can get. You want enough lenses that if one gets obscured by water, hey, there are others that will pick up the slack, and do it without bothering the driver. You want to shrink the 'no coverage of object in environment' surface as small as possible. Nobody seems to respect this on the academic software side. Novelty, sure, but people don't seem to earn grants publishing extra 9's using a mashup of a bunch of pricy sensors. 'Less is more' is not how you get the data needed to make things reliable, it's how you get best-case detection rather than worst-case awareness. †This combination is now found in quadrotor IMUs. Each individual sensor is inaccurate enough that it would be useless for automation, but combine them to check each other's deficits and you get a remarkably elegant system.
- Animats 10y ago
  Where the solid object is a car or truck, yes. Phased array radar is good enough for that, and there are lots of deployed units on the road. Telephone poles, deer, and pedestrians - maybe.

beamatronic 10y ago

That raises an interesting question: It is 2016, do we ( or do we not ) currently have economically feasible technology to place into cars that can reliably and accurately determine when you are about to hit a solid object or not?

mapt 10y ago
Define "Solid object". In the lab, opaque textured objects in sufficient lighting are pretty easy to localize by a variety of methods. Reflective objects are quite a lot harder, but there are methods that work. Reflective objects through distortion like rain/fog... harder, but often doable. In low light? Depends. The weird thing with environment modelling in computer vision is that there are a lot of overlapping methods, but sensor fusion of these methods is rarely attempted. Most robots end up with a pair of laser rangefinders or a pair of cameras or something equally simple to program. Creating a high-reliability automated vehicle that can bypass all the edge cases where these things fail is another thing entirely. You want overlapping, inter-reinforcing models of the world, with algorithms that not only identify items in a well-calibrated system, but calibrate the system in real time based on items identified. My lab was really excited about replacing stereoscopic computer vision with monocular computer vision for 3D modelling, but the opposite direction is where you want to go if you care about as many 9's of reliability as possible and hardware is cheap. You want a few dozen cameras with fixed orientation, you want parallax laser rangefinders in a variety of bands, you want time of travel LIDAR units, you want a {GPS + 3DoF gyroscope + 3DoF magnetometer + 3 DoF accelerometer + barometer + thermometer}†, you want ultrasonic and radar rangefinders; You want everything you can get. You want enough lenses that if one gets obscured by water, hey, there are others that will pick up the slack, and do it without bothering the driver. You want to shrink the 'no coverage of object in environment' surface as small as possible. Nobody seems to respect this on the academic software side. Novelty, sure, but people don't seem to earn grants publishing extra 9's using a mashup of a bunch of pricy sensors. 'Less is more' is not how you get the data needed to make things reliable, it's how you get best-case detection rather than worst-case awareness. †This combination is now found in quadrotor IMUs. Each individual sensor is inaccurate enough that it would be useless for automation, but combine them to check each other's deficits and you get a remarkably elegant system.
Animats 10y ago
Where the solid object is a car or truck, yes. Phased array radar is good enough for that, and there are lots of deployed units on the road. Telephone poles, deer, and pedestrians - maybe.