9 ms·

I would think camera-based navigation would be fairly successful at this point. We can easily store large amounts of image and terrain data, and deep learning

I would think camera-based navigation would be fairly successful at this point. We can easily store large amounts of image and terrain data, and deep learning based image recognition systems are quite powerful. I would think that by limiting your image search space based on your likely position would make the problem even easier to solve. Terrain based navigation is certainly nothing new(1950s?), so there is probably a large pool of research into it, but I wonder what's been done with recent pattern recognition advances?

upbeat_general 5y ago
There’s still the issue of compute power required for a lot of the CV algorithms you’d want to use. There’s also the question of what data you are using; Satellite imagery is much harder to match with a ground perspective and ground imagery(e.g street view) is very tedious to collect and keep updated. In areas with unique features (dense urban) localizing to within ~50ft has been demonstrated pretty robustly but more accurate methods of localization are a lot more variable and depend on the datasets, cameras (monocular vs stereo), compute, etc.
Dunedan 5y ago
> I would think camera-based navigation would be fairly successful at this point. We can easily store large amounts of image and terrain data, and deep learning based image recognition systems are quite powerful. Aside from the fact that the amount of data required for that would be orders of magnitudes larger than you can reasonably store, that wouldn't work at all in areas with little terrain features like deserts or the open sea.
- Daniel_sk 5y ago
  Or just a dense forest.
- upbeat_general 5y ago
  It’s certainly conceivable that you could store a compressed representation of the image features. Especially if you’re operating in a more limited environment (e.g. a large metro area), it would be very feasible to only store extracted features from the image data that you could use for matching.
- 01100011 5y ago
  We have literally been doing it since the 50s. Seems like technology has come quite far since then. https://en.wikipedia.org/wiki/TERCOM https://en.wikipedia.org/wiki/TERCOM Here's a modern take on it: https://www.youtube.com/watch?v=U5Kr0YI3sec https://www.youtube.com/watch?v=U5Kr0YI3sec
- Melkman 5y ago
  Don't look down. Look up. That's been a navigation standard for years. Funnily enough replaced by GPS. https://airandspace.si.edu/collection-objects/astro-tracker-md-1-b-52/nasm_A20110297000 https://airandspace.si.edu/collection-objects/astro-tracker-... That doesn't really work in a sub though. Also compasses (especially of the 3D variety) can be a great help.
  - ac29 5y ago
    Does that work during day time?

upbeat_general 5y ago

There’s still the issue of compute power required for a lot of the CV algorithms you’d want to use. There’s also the question of what data you are using; Satellite imagery is much harder to match with a ground perspective and ground imagery(e.g street view) is very tedious to collect and keep updated. In areas with unique features (dense urban) localizing to within ~50ft has been demonstrated pretty robustly but more accurate methods of localization are a lot more variable and depend on the datasets, cameras (monocular vs stereo), compute, etc.

Dunedan 5y ago

> I would think camera-based navigation would be fairly successful at this point. We can easily store large amounts of image and terrain data, and deep learning based image recognition systems are quite powerful. Aside from the fact that the amount of data required for that would be orders of magnitudes larger than you can reasonably store, that wouldn't work at all in areas with little terrain features like deserts or the open sea.

Daniel_sk 5y ago
Or just a dense forest.
upbeat_general 5y ago
It’s certainly conceivable that you could store a compressed representation of the image features. Especially if you’re operating in a more limited environment (e.g. a large metro area), it would be very feasible to only store extracted features from the image data that you could use for matching.
01100011 5y ago
We have literally been doing it since the 50s. Seems like technology has come quite far since then. https://en.wikipedia.org/wiki/TERCOM https://en.wikipedia.org/wiki/TERCOM Here's a modern take on it: https://www.youtube.com/watch?v=U5Kr0YI3sec https://www.youtube.com/watch?v=U5Kr0YI3sec
Melkman 5y ago
Don't look down. Look up. That's been a navigation standard for years. Funnily enough replaced by GPS. https://airandspace.si.edu/collection-objects/astro-tracker-md-1-b-52/nasm_A20110297000 https://airandspace.si.edu/collection-objects/astro-tracker-... That doesn't really work in a sub though. Also compasses (especially of the 3D variety) can be a great help.
- ac29 5y ago
  Does that work during day time?