SLAM and LiDAR Cave Mapping

[Addison Snyder]

I've been holding off on this post for a couple of months now, but I've been working on mapping caves using visual-inertial SLAM. There's still a ton of work to do, but I'm getting there with the process and it's only getting better over time. I could practically write a thesis on it all, but to simplify, my goal is to:

1. Use VI-SLAM to map out an accurate trajectory underwater (where I'm at now, refining lots of stuff)
2. Add an instrument tube to the front of my DPV. This tube will house a 360-degree blue-light LiDAR (ya'know, the spinning kind) to get accurate cave cross-sections. It will also read depth and house a much better IMU.
3. Refine the processing of all collected data into a 3D map, and make the data and derivatives available for free (as in beer!) under a strong copyleft license.
4. Spread the process/methods/documentation/data - this is really supposed to be happening throughout. I want to make this reproducible for all. This means how-to's, repositories of processed info, configuration, and maps. Ideally some random person in Mexico could buy a couple thousand dollars in equipment and map caves there, put in a PR, and contribute.

Maybe my vision is lofty, but I want for everyone to be able to access and contribute to crowdfunded/free cave maps, whether digitally produced or conventional. I hope that 10 years from now we won't be taking phone-pics of cavemaps on the walls of dive shops or referencing a grainy/outdated image of a more obscure cave, but rather accessing freely available/modifiable digital maps. I'm posting this as a sort of milestone marker that I got this far (I stand on the shoulders of giants of course), but soon I'm hoping to start a little gofundme (~3k or so) to cover the equipment involved and pilot a map of Ginnie's DPV-able routes. I'm highlighting the free/opensource parts because I want that to be the hard standard, and don't want to ever go back on that. Lots of people have done similar projects (Sunfish, Maria Concordia Project, etc), but the equipment and software costs are prohibitive, and/or not practical to imitate. I'm hoping to change that, both in method and in scale. The attached video/images are made with completely free/opensourced software. The hardware (so far) costs ~1k, and can be carried by hand or attached to a DPV (or presumably a superman'd tank).

My DPV (Dorothy I is a Twister reference):

 

[Addison Snyder]

Another pic of my DPV setup. Used 80/20 to attach two GoPro Hero 11s and 60-ish dollar video lights.

 

[Smache]

Looks like it's coming along nicely, that's really cool

 

[wetb4igetinthewater]

This is awesome. I will share with some of my cave diving friends in Greece (which has A LOT of caves that are waiting to be known outside of the country).

 

[berndo]

Great project!

 

[sprockjohnson]

This is really cool. Are you using the GoPro's for video or time-lapse photos?

 

[Addison Snyder]

This is really cool. Are you using the GoPro's for video or time-lapse photos?

240hz 2.7k video of course! I downscale it, but the raw high-end video is always there to reprocess later if I change methods, algorithms, configurations, etc.

 

[mariosx]

Hey, truly great work!

Interesting results and OrbSLAM2 seems like a solid option. My colleagues at the University of South Carolina (Dr. Ioannis Rekleitis' team) are working for at least 6-7 years on the problem, in case you would like to check some of the work or reach out. Here you can see a 4-5 years old work. Things have moved a lot the past years, and last time I checked they were providing (semi-)dense photorealistic reconstructions, along with superb quality. Map accuracy has been improved a lot even with using only a single gopro. There are some other cool videos I could show, but I am not sure how much I can share. They are funded almost entirely by NSF, so there is emphasis on using and producing open source packages. If you are truly interested on this problem I would recommend checking their published work, and/or with them directly, given that they are internationally the leading academic group in the domain. More and free data are always appreciated given the challenging logistics.

Disclaimer:
I am aware of the technology they are using, but I might not be the best option to answer all questions. SLAM is not taking the first position in my list of expertise.

P/S: We hope that soon we would be able to map some shipwrecks and potentially caves using fully autonomous robots. The technology, from our side, is there for the past 2 years, but we had major issues with hardware.

 

[Addison Snyder]

Here you can see a 4-5 years old work.

Very well done! Looks pretty incredible!
One of the things I've found with SLAM is that the individual feature points aren't particularly accurate, at least not on a single pass from a single angle. And while visual SLAM gives pretty good low-latency data on position changes, it drifts modestly over time without other sensor input. I'm hoping to eventually incorporate more "ground truth"-ish inputs to make up for visual SLAMs weaknesses. A single IMU will pretty much guarantee that gravity always points down within some margin. A (3d) compass, well, is a compass. A depth sensor takes care of an entire axis. Silt aside, it's pretty hard to fool a ToF laser for cross-sections!

What's that tubular instrument at the very top? Does it record flow past the DPV?
At 2:51, I see some Charuco markers. Are those to compare the results to scale?

I'm visiting South Carolina this weekend, maybe I'll pop in his office!

 

[mariosx]

One of the things I've found with SLAM is that the individual feature points aren't particularly accurate, at least not on a single pass from a single angle. And while visual SLAM gives pretty good low-latency data on position changes, it drifts modestly over time without other sensor input. I'm hoping to eventually incorporate more "ground truth"-ish inputs to make up for visual SLAMs weaknesses. A single IMU will pretty much guarantee that gravity always points down within some margin. A (3d) compass, well, is a compass. A depth sensor takes care of an entire axis. Silt aside, it's pretty hard to fool a ToF laser for cross-sections!

Yes, that's a yet to be solved problem, BUT there are some techniques that could help, such as loop closure. If you check the second link I shared this is on a single gopro, and the scale matches "perfectly" the manual measurements of the the cave. The drift exists for some time until loop-closure is performed (i.e. we detect the same place again) and the map is collapsed due to that point.
IMUs, magnetometers, etc are easy to understand, difficult to master. few degrees of errors over time can translate to many meters of error. I would recommend checking with Dr Ioannis Rekleitis before discussing more details. It would be optimal if you don't have to reinvent the wheel.

What's that tubular instrument at the very top? Does it record flow past the DPV?

No. It's a 2D sonar. It is mounted perpendicular to sensor rig. We have not yet see a reason to record the flow for measuring distance. This looks very unreliable for creating the high details map we aim for.

At 2:51, I see some Charuco markers. Are those to compare the results to scale?

Yes. It is only for validation. We just place them and measured the distance manually to see if the scale is reasonable. Also for comparing different techniques by reprojecting them and measuring the differences.

I'm visiting South Carolina this weekend, maybe I'll pop in his office!

I recommend sending an email first.