• Descent 302 is published on 15 February and it will soon be on its way to our subscribers.

    In the newsdesk, read a review of the underground events at Kendal Mountain Festival, plus tales of cannibalism and the Cavefish Asteroid.

    In regional news, we have three new connections in Ogof Agen Allwedd, a report on the iron mines of Anjou, an extension to Big Sink Cave in the Forest of Dean, a new dig in Yorkshire's Marble Steps Pot, student parties, an obituary for Tony Boycott, a tight find in the Peak District and a discovery in County Kerry with extensive formations.

    Click here for details of this edition

Handheld 3D scanning - move over disto?!

I want one, It's like a Leica geosystems MS60 but hand held, amazing technology, but the Leica also uses lasers, i've wanted to get one underground for ages.

 
How far away are we from having a caving lamp with’survey mode’ that plots a 3D model as you go.!?
Is there currently a sensible workflow for deriving a 2D paper survey from a 3D model? The thing about surveying currently, even with a Disto, is that the process is designed around getting a good 2D representation right from the start. I don't know how this would work for a 3D modelling workflow.
 
Er. Could you create the passage as a 3D solid and do a plan and elevation with the outline and have all the floor detail in faint perhaps run through some kind of detail filter to pick out only features of a certain size? It could easily find cylinders (stals) and apply gradient indicators. Cross sections would be a p of p.
 
You can do 3D stuff with your iPhone. All the latest ones and iPads have a LIDAR sensor built in.

Google Polycam app and you'll find it and other similar apps. Mainly aimed at building engineers for creating drawings and walk-throughs.
 
Is there currently a sensible workflow for deriving a 2D paper survey from a 3D model? The thing about surveying currently, even with a Disto, is that the process is designed around getting a good 2D representation right from the start. I don't know how this would work for a 3D modelling workflow.
We just need a skilled professional software developer, with experience in AI, to spend the next 10 years developing a 3D-2D interpreter. Train a neural network on 3D scans of caves with excellent 2D surveys - a few thousand should do (so maybe add a few teams of experienced surveyors to become full-time surveyors). If that all goes well, I reckon you'll save at least half the time of going from 3D lidar to 2D survey, once someone's gone over the resulting output and corrected all the mistakes.
 
My worry is that a good survey is, quite intentionally, not an accurate representation of reality. Instead it presents a cave as we experience and understand it.

Some of the best surveys that I've used to navigate are the least accurate, because they've distilled out much of the complexity that you don't actually observe yourself. In fact making a survey without extremely precise equipment or process makes you fill in the gaps with your intuition, which is formed by how you experience the cave.

Now I'm not saying that we shouldn't pursue easier and more accurate data capture techniques, but that I'm seeing great growth in capturing 3D models, and very little in translating these to surveys.

Logismos's tongue in cheek suggestion is a good one, because it would ideally capture the essence of that distillation of reality into understanding that survey makers do. Loki's suggestion is how I've understood it to be feasible, but I've not seen it ever actually done!

Addendum: which is why I don't think LIDAR or iPhones will be replacing Disto's any time soon!
 
We just need a skilled professional software developer, with experience in AI, to spend the next 10 years developing a 3D-2D interpreter. Train a neural network on 3D scans of caves with excellent 2D surveys - a few thousand should do (so maybe add a few teams of experienced surveyors to become full-time surveyors). If that all goes well, I reckon you'll save at least half the time of going from 3D lidar to 2D survey, once someone's gone over the resulting output and corrected all the mistakes.
That actually wouldn't be that hard to programme, the hard part is the getting the amount of processing power AI uses into something that can be carried by hand, people use AI on their phones but all the AI's processing is done on external servers, you can run models on your pc but you'd need an AM5 processor for it to be any good.
 
the hard part is the getting the amount of processing power AI uses into something that can be carried by hand
I don't know about that - sure training the model would be expensive, but once you've got the weights running them is far less computationally costly. We'd also expect to be doing this on a desktop outside the cave environment anyway, as processing LIDAR point clouds is horrendously computationally intensive.

The impractical bit is getting enough training data - you'd need loads of 3D scans and surveys of the same bits of passage. The problem framing would also be a complete nightmare, as you'd have to condense a point cloud into a lower dimensional space that is feasible for the model to fit, which would be a challenging problem.

Ultimately this isn't a good problem for a data driven model, as it should be fairly straightforward to determine a reasonable algorithm for deriving a plan and EE from a 3D model (with centerline would help a lot). A cartographer could then adjust, add symbols etc., in the usual way.
 
I don't know about that - sure training the model would be expensive, but once you've got the weights running them is far less computationally costly. We'd also expect to be doing this on a desktop outside the cave environment anyway, as processing LIDAR point clouds is horrendously computationally intensive.

The impractical bit is getting enough training data - you'd need loads of 3D scans and surveys of the same bits of passage. The problem framing would also be a complete nightmare, as you'd have to condense a point cloud into a lower dimensional space that is feasible for the model to fit, which would be a challenging problem.

Ultimately this isn't a good problem for a data driven model, as it should be fairly straightforward to determine a reasonable algorithm for deriving a plan and EE from a 3D model (with centerline would help a lot). A cartographer could then adjust, add symbols etc., in the usual way.
Luckily Leica Geo systems has software that does exactly that, but as far as I know it's only usable with their surveying systems, which are excessively expensive.
But even Leica use base data sets which usually consist of straight lines, and points of reference, using laser and crystal so even with the scanning hardware you still need to collect that data, for it to build the laser read 3D projection.

I'm very used to the Leica Geo systems MS60, because as a game developer it's easier to scan buildings to put them in a game than to recreate all the meshes by photogrammetry like we used to, which took hundreds of photos to do.
 
From a 'sport' caving perspective, what I'd really like to see is a hybrid of really good, human drawn, 2d line surveys , combined with a 3d perspective in line art. Create like a old-school climbing topo for the cave trip. I think a 3d model from the likes of iphone polycam would make it much easier to draw this at home, it just needs willing and capable folks to draw up the psuedo-3d line sketch. Eventually, with enough data to learn from, perhaps AI could do some of the legwork for this.

I had a go at sketching the entrance series to County Pot, from memory. An artist or illustrator with some actual skill could make some amazing things in this style.

County pot.jpg
 
What about one of these things? It does sound interesting.
 
Moasure is a double-integration solution to turn acceleration and digital compass bearing into plan displacement. It is very sensitive to technique. You need to move the puck horizontally, smoothly and quickly between stations. It must stop and be stationary at each station to find another boundary condition to prevent non-zero velocity drift. Might work in big flat chambers but probably not in any kind of awkward or vertical cave (it cannot distinguish between vertical acceleration and gravity and very prone to error vertically)

Of all the accelerometer based LIDAR things, Zebedee is my favorite one. Also likely to be the most practical to use in a cave environment and unaffected, or possibly improved, by having a skin-full the night before.


This video is 12 years old. The technology has existed for a while and it will all come together into something useful eventually. Perhaps just a pair of glasses worn by the caver.
 
I'm actually working on something like this at the moment, using a low light camera, a lidar camera and an IMU.

But there are some significant challenges, which I'm not sure can be overcome:

  • Lighting: photogrammetry (using photos to generate 3d shapes) expects things to look roughly the same - so having consistent lighting is vital. However, if the main source of light is your head torch, then this is going to move around and cause problems recognising two bits of cave as the same place
  • Lighting(2): we get a lot of blurring when recording video in low light conditions - this again is not good for point recognition
  • Range: most lidar units (which avoid the two above problems) have a very limited range - up to 4m for reasonably priced and portable units, up to 10m for devices up to £1000 - this isn't good enough for proper big chambers (eg gaping gill main chamber)
  • Hard maths: The maths for this is really complex, and I'm struggling to get this working in a reasonable timescale, even on a powerful laptop with GPU acceleration
  • Missed bits/disconnections: Because your just taking a recording, you don't get immediate feedback as to whether your data is good enough, or if there are missed connections etc. Solving the above point enough that it can all be done on eg a mobile phone could let you check your scan "in-cave"
  • Scrotty bits: how would these techiques deal with tight squeezes where you have to thrash your way through it?
  • Water - a very reflective water surface will behave like a mirror and confuse things. Also lots of spray/falling water will interfere with any kind of light based surveying

I'll not address the 3d->2d argument, but just note that there are some people (Julian Todd) who would argue you don't need the 2d bit at all.

Phil
 
This is a cave in Australia scanned using Zebedee in 2014. The detail captured from someone just walking about on the surface and then underground is amazing :

 
There is a scanning mode for the BRIC5 in Topodroid which at 1 Hz is about the same speed as most totalstations. Slow but more or less affordable.
Next step up is a handheld scanner. The FJD Trion P1 performes quite well underground. It can be used with camera if there is enough light for colouring and can be upgraded with RTK. Mapping works at walking speed from 0.5 to 70 metres range.
The alternative to the Moasure is the foot-mounted pedestrian navigation sensor of Inertial Elements. A bit more experimental but you can achive good results.
Have a look at CloudCompare for working with pointclouds.

 

Attachments

  • temple.jpeg
    temple.jpeg
    65.4 KB · Views: 36
Back
Top