Pointcloud placement

[SilaS]

Hi,

I am in searching for a simple answer! How can I place to each other with centimeter accuracy two Pointcloud?

Because till not found a working answer I made a TEST.

I recorded two sets of a cca. 3 acre of grain field 20 minutes between the records. During it I used an Emlid Reach L1 GPS. The base and the rover was full powered during record from beginning of first take-off till the landing of the 2nd flight.
During the postprocess I used the same WGS84+egm96 height ellipsoid and the same offset in alt.
Processing was perfect both model. It can reach here: https://drive.google.com/open?id=0B5zFToQ9qFweN0dYSUU1X1FVNjA

The result DEM and orto has been placed on Propeller Aero site to show the ERROR. I can send invitation to see it!

I made 3 version to align:
1. Marker placement (Align model with markers) - T1-repo, T2-repo
You can see the file in attachment. There is 3.173m3 difference! The marker was placed on object on the ground under 5 cm precision.
2. Point based Alignment - T1pp,T2pp
Also in attachment. They say 6.455m3 diff!
3. Meshlab placement (with Align tool) so exported the mesh in ply and imported it back
Here the diff is 6.248m3! So it seems the Meshlab wins over PS in this step =:-| The placement of Meshlab in attachment.

So till the marker based is the best one. But it seems to me, that detecting a small amount of change (1m3 - what we moved!) on the surface is TILL a DREAM and unreachable.
 
There will be ever 20-40cm of diff in altitude!

Did ANYBODY found a solution for it? Did I made something wrong?

SilaS

[SilaS]

  Did I asked somethig wrong? Nobody meet with this?!

[SAV]

Hi SilaS,

I assume you used RTKLIB to extract the position of your rover/drone in post processing. Have you checked the accuracy of each point (HRMS, VRMS, VDOP, HDOP)?
Maybe the difference in your 2 datasets is due to a variation in GPS accuracy. Depending on the position and availability of satellites you might not always get sub-dm or sub-m accuracy with the Emlid Reach. Moreover, magnetic noise from the drone might be an issue and could interfere with your GPS antenna, resulting in low accuracy measurements.

A simple way to check if that was the case is to have a look at the camera stations after you imported them into PhotoScan (the blue dots in the model pane). Change to Orthographic View (press 5 on keyboard) and then look at the camera stations from both flights. There should be a visible difference in altitude between them (assuming that you were flying both missions autonomously on a set altitude).

However, what you could still do is to compare both dense point clouds (from flight 1 and flight 2) using CloudCompare (free, open-source).

1. Align/register both point clouds using point pairs for rough alignment first and then ICP (see http://www.cloudcompare.org/doc/wiki/index.php?title=Alignment_and_Registration)

2. Compute cloud to cloud distance so assess the difference (see http://www.cloudcompare.org/doc/wiki/index.php?title=Cloud-to-Cloud_Distance)

DONE

You could also use your markers on the ground to register/align both point clouds.

If you need help PM me.

Regards
SAV

[SilaS]

Hi SAV,

Thanks for tips!

Emlid works perfect. It regular has Quality=1 results.
The problem is with Photoscan! Does not mother, how precise <10cm did I mark the markers on the photos. After Alignment of two chunk it has every time more then 40cm huge errors with bend and in both vertical and horizontal on pointclouds.

Now I try to made tiled models and marking markers on this instead photos and hope, that it will final do able to align with photoscan.

So periodical measurement of the same pointclouds is a nightmare! How did made it Pix4D? I am not enought brave to imagine it....

SilaS

[Alexey Pasumansky]

Hello SilaS,

Usually we do not recommend to perform the chunk alignment for the georeferneced projects and rely on the default alignment according to the referencing information.

When you are aligning chunks using any other approach (marker / camera / point based) you will loose georeferencing and scale for all chunks except the one used as a reference chunk. During such chunk alignment procedures PhotoScan will try to minimize the difference (the errors) for the points used for the alignment (marker or camera positions, or tie points found for the images from the different chunks, depending on the alignment method used).

To my mind the proper way would be precise georeferencing of each chunk in the same coordinate system (could be done using the same GCPs or different GCPs each time) and then subtracting the exported elevation models one from another.

For precise georeferencing you need fair amount of GCPs distributed more or less uniformly on the area of interest (or precise onboard GPS system). In your projects I see only four markers which lay almost on the same line, so optimization wouldn't improve the accuracy much.

Additionally I can suggest to pay attention to the image quality, since most of the images are blurred.