Automatic Ground Classification in ReCap Pro 23.1.0 Update

[Pointdump]

Trying out the new >>>Automatic Ground Classification<<< in ReCap. Not getting the expected result, which would be Ground Points Only.
Here's the Point Classifications of my LAS file:
2109533 unclassified (1)
2958253 ground (2)
2677 noise (7)
723837 water (9)
2527 rail (10)
1778 bridge deck (17)
2 Reserved for ASPRS Definition (18)
When I use the new tool and create a surface in Civil 3D I get 5,795,275 points. What am I doing wrong? Point Cloud (EPSG:2967) >>>Link<<<.
Dave

[TerryDotson]

This file has already been classified.  I think the intent of the new ReCap tool would be to process lidar that has not been classified, as in from a drone, etc.  A good test would be to remove the classifications from this file then run ReCap to see how close to the known "2958253 ground (2)" it returns.  Do you have a method to strip the classes, as in export to X,Y,Z (no class) and reimport?

[Pointdump]

Hi Terry,
That didn't occur to me. I could probably strip the classification with LASTools, but I've never done that before, so it might be easier to grab some LiDAR that is unclassified.
Dave

[TerryDotson]

I could probably strip the classification with LASTools, but I've never done that before, so it might be easier to grab some LiDAR that is unclassified.

---

Here is a download of the same lidar in LAZ format with the classifications stripped.  That would make it similar to drone lidar I've seen with an exception, the drone output is usually colorized.

Also, there is at least one point in the data reporting a negative elevation -513.08, so ReCap isn't to blame for that.   Anyway, run that through the new ReCap and see how close to the 2,958,253 ground points it returns.  I'll try our ground tool to see what it shows, it works differently than RapidLasso's LasTools and ReCap.

[Pointdump]

Thanks, Terry. I'll try it right now.

[TerryDotson]

---

I'll try our ground tool to see what it shows ... how close to 2,958,253:

---

DS> Elevation Tolerance <1.0000>:
DS> Initializing Projection Data ... 0.4278 Seconds
DS> Obtaining Web Elevations ... 1.0014 Seconds
DS> Total of (167684) Reference Surface Points Obtained.
DS> FastMesh Triangulation ... 0.845 Seconds
DS> Comparing Lidar Elevations ... 31.4420 Seconds
DS> File: [IndianaDeClass] Ground: 2680080, Other: 3118527

DS> Processing [IndianaDeClass] ... 48.40 Seconds

[Pointdump]

Terry,
Thanks for de-classifying my cloud.
OK, with las2las I filtered out the 2 points below 0 elevation, leaving 5798605 never classified (0). After using Automatic Ground Classification in ReCap I ended up with 5797046 points. So that's a reduction of only 1559 points, and I can still see every tree in the cloud. So Autodesk has some 'splainin' to do.
Dave

[TerryDotson]

.... I can still see every tree in the cloud. So Autodesk has some 'splainin' to do.

It would be best if you could get your hands on a drone produced lidar set to test with (that has trees, buildings, vehicles) before coming to a conclusion.  That way you could go straight from an unchanged data set to results.  Hopefully, someone who can provide one will chime in with some downloadable test data.

[Mike.M.Carlson]

Hello Gentlemen,

I hope all is well.  Please see the link below:

BRD_NAD83-2011-OH-N_usft_NAVD88.las

It is a small swatch from a recent UAS LiDAR data survey I did.  It is a raw unclassified .las that has a mix of trees, ground, and structures.  There's about 75MM points in this set due to 32 channel LiDAR system.

I hope it helps.

[Pointdump]

Thanks, Mike! Wow, 3.5GB! Downloading now.
Dave

[Pointdump]

Mike,
Thanks for that nice data. It took about 1-1/2 hours for ReCap to crunch through its Automatic Ground Classification routine. I ended up in Civil 3D with 75899901 Points. AGC only lost 9292 Points, and I still see trees. But it did take out buildings. Until I see the trees gone I remain unimpressed. I might be missing something. I wish Autodesk would add to the Help on this new function.
Dave

[Pointdump]

OK, I just had to try LASGround in the LASTools. Not free, that single tool runs about 1500 Euros. I figured they wouldn't mind me evaluating their software once. Took about 14 minutes:

Please note that LAStools is not "free" (see http://lastools.org/LICENSE.txt)
contact 'info@rapidlasso.de' to clarify licensing terms if needed.
WARNING: unlicensed. over 1.5 million points. output slightly distorted.
tiny xyz noise. points permuted. intensity, gps_time, user_data & point_source_ID zeroed.
done with 'D:\Download\BRD_NAD83-2011-OH-N_usft_NAVD88_Gnd.laz'. total time 843.926 sec.

Here's the results using LASInfo:

histogram of classification of points:
47805395 unclassified (1)
28103798 ground (2)

[Mike.M.Carlson]

Hi Dave!

Looks like you're using a good option to tackle this.  I'm using LP360 advanced and they have a recommended 2 step process for classifying ground.  Create an initial seed surface to use as a reference with the auto settings below (I also filter to last return only and note my subjective seed grid size): 

Next step is to rerun algo using the Step 1 seed surface with a larger grid spacing but more stringent filtering values:

The results still aren't perfect and require some manual intervention to classify points using LP360 profile view.  In the end even if you get a dyno-supreme point cloud with noise factored out and true ground points classified, you will be left with a ton of points that Civil 3D will choke on.  

Terry's Mass Importer has very slick tools to help with the large number of points.  However, the best way to define drainage features (like channels) or sharp terrain breaks is via 3D breaklines.  To me that is the holy grail if you can find an algo that will auto draw 3D breaklines for these sharp features and build a surface from that infrastructure.  So that would give you a lean and mean surface that is decimated but accurate.  Now you can use grading tools, etc and not bog down the engineering design process or drawing production workflows.  I have to do this 3D breakline conflation manually now in LP360 and it is a HUGE time suck.  However, that is part of the game if you want something defensible and accurate for engineering design.  

Anyway, I wish there was an easy answer to all of this.

Have a good weekend gents.

[Pointdump]

Hey Mike,
LP360 looks pretty rigorous.
LASGround also looks for last returns and also considers height differences in various settings, like city or wilderness.
Automatic breaklines is definitely the holy grail.

Dave

[Mike.M.Carlson]

You ever try cloud compare?  That is open source and has a decent ground classification engine.  Worth a look.  Have a good weekend!

[Pointdump]

I use CloudCompare quite a bit. I didn't know it could do anything but filter already classified points. I'll look into that. Thanks.
Dave

[neilyj666]

@Pointdump Dave - can Recap do the filtering? I tend to use Cloud Compare to create a cloud of ground points (filter by 2 - it is pretty quick) then into Recap and finally Civil which is a bit of a long way round.

[Pointdump]

Hi Neil,
I think LAS2LAS in LASTools edges out CloudCompare for Ground Point Filtering.

For classifying Ground Points from an unclassified Point Cloud, Cloud Compare looks like the bomb. It's waaaay faster and easier than LASGround. Whether it works better, I just don't know.
The latest ReCap update now shows the classifications, and by clicking all the other classifications off, it seems to filter ground points. But when you import the RCP into Civil 3D, all the points are still there. So unless I'm missing something it just doesn't work.
Dave

[Mike.M.Carlson]

Hi Dave,

It makes sense to me that the RCP insert will import all classes into C3D.  Think how long it took for them to have classification view filters as an option 😅

So I’m sure you’re aware that what you’re seeing in ReCAP is just a View Filter, so unless you can export a new RCP out of ReCAP without the unwanted classes, I think it is easier to use LASTools or Cloud Compare (or a combination) to do the classifying / filtering and export only the ground class as a .lzw or .las.  If you have high confidence in your ground class, then the Dot Soft Mass importer tool can import the .las directly and allow further filtering if needed to create the surface.  Super easy.  

Thanks.

[rl_jackson]

I've been taking a serious dive into Lidar data in the past 6 months, and I've thus far used IW and CC to classify cloud data. I'm with @Mike.M.Carlson in search of that "Holy Grail" at the moment, I've been using IW (the Recap-IW-C3D workflow) for the automatic linework features in terms of pavement markings and some other items that it is capable of generating. I would say that my take thus far is that Lidar has reduced the field crew time but increased the office time. I'm seeing some benefits to IW with respects to the linework & transect data that can be generated but the court of opinion is currently still in session, so I won't give a verdict at this time.

Currently I'm working on a project that has Field Topo, and Lidar for ground data, and I've got bathymetric data to add into the mix once we get out there and collect that data. Due to the existing bridge, I'm considering using the cloud data as a baseline, and then creating the bridge model in IW using the supplemental data from the field survey and Lidar. Should be an interesting stretch A1A (US 1).