Image alignment over forested area - any tips?

[jdyuen]

Hi, I have some images taken over a forest area and I am having trouble getting them to align. Does anyone have any tips for this situation?

I have GPS and yaw/pitch/roll data for each image (is orientation used in version 0.9.0?).  How are orientation angles defined, for a downlooking camera is it yaw 0 = north, pitch 0 = nadir (+ve nose up), roll +ve right wing down?

The position and orientation data I have for the images doesn't seem to be constraining the alignment very much as many cameras are often aligned in a very different location and orientation than the initial GPS/IMU reference point.

I have attached a few typical images showing the area and overlap.

Thanks!

[RalfH]

From the two very small images, it is difficult to say what could cause the problem. How many images do you have, what is the overlap between images (seems to be 80-90% for the two images - are there also several rows of images?), what is the image resolution? Do you want to create a model of the trees or of the underlying ground surface or of both? With the "naked" trees standing out over the landscape, I would assume that you need a very large overlap between images (>=90%in row and between rows) to minimize ground occlusion effects due to the branches.

I have not tried to work with image sets where I have accurate position and orientation of the cameras. My suggestion is to use the images without position and orientation data to create a 3D model, and then use the known camera positions as "ground" control points to rectify the model.

Also, I am not quite sure if Photoscan is the perfect tool if you want to get the ground, as "height field" will give you the top envelope of the vegetation and "arbitrary" will give you a model containing both ground and trees, requiring you to manually remove the trees or to filter the dense point cloud in another software.

To Agisoft: A third "object type" under "Build geomertry" would be great: base height field. Shouldn't be too hard to implement.

[Mohabon]

I dont know if i understand it right, but I want to give my contribute in this subjetc and see if it is possible to implement the "remove trees / vegetation" tool.

With my university we both work with LIDAR datas and Photogrammetric methodologies for MAPPING.

LIDAR is expensive, but it can easily remove vegetation/grass/trees as it uses an emitting source and interferometry criterias to find the "last echo return" that is assumed as the ground, even in very dense scenarios (forests, corn fields, ecc).

As it is said in this topic, the Build Geometry with Height field in PS PRO only gives the top of the envelope, so to manually remove trees / vegation from point cloud or DEM really takes lot of time.
This job is really hard and sometimes it is better to spend more money in LIDAR to save time in processing.

If there were a tool / procedure to find the ("BASE HEIGHT FIELD") = (ground cleaned by trees/vegetation) , we would save lot of money by using photogrammetric UAV instead of LIDAR aircraft.


This is an implementation that could really improved PS Pro market and i am sure many users would pay extra money for having it, my university department included!

Kind regards.

[jondandois]

We have done a lot of work using Photoscan, Bundler, and VSFM for reconstruction over forests with no additional camera parameters  (other than whatever the default EXIF tags report) and have not had the problems you describe.  I agree with user RalfH in asking about your overlap.  We are typically operating at > 80-90% endlap and have actually had success with almost no sidelap on a parallel grid path with a multirotor.  A big downside of no sidelap though is that the system is sensitive to slight path variation, leading to complete gaps in coverage with a little bit of deflection, so no sidelap is avoided. 

Interesting to see a lot of discussion about removing the vegetation - that is the part we are most interested in!

[MiguelVarela]

Any updates on "removing vegetation"  in Agisoft?

[bigben]

Classifying the point cloud helps remove unwanted parts.I'm starting to produce scenes for VR environments. To cut down the size of a single mesh I export a surface for the ground and then mesh individual buildings. The automatic classification of ground points saves a lot of work.

But back to the original question: the camera positions/orientations can drift during the alignment process. There are settings to prevent the cameras moving beyond a specified limit, but I've never used them as I just georeference afterwards using GCPs on the model... (but then I don't use/have a drone).