eCognition

As Javier said, e-cognition is by far the best tool, but it comes with a rather high price-tag... same for Lastool (a bit cheaper though).

If you are not allergic to C++ (there is also a Java interface), the Point cloud library (PCL) is rather good: http://pointclouds.org/

Finally, the scan of a gravel pointbar for instance, regardless of how accurate and precise is the scan, is not going to give you A, B and C, because of imbrications and the absence of the bottom part of your gravel.... there is a lot of work on it, but nothing has been really cutting it so far.

Please find attached a poster describing an old and simple approach that starts from photographic gravel bed river sampling.. not sure it fit with your needs.

Thanks for all of your answers. It seems I may have to look in to eCognition a bit further. 

Christopher: I understand that the c-axis is going to be mostly buried, but I want to measure the exposed part of the axis to account for protrusion and quantify surface roughness, and use the a-axis and b-axis to look at alignment and size fractions. There are quite a few methods getting close to what I want but as you say nothing is quite there, and my programming isn't too sharp.

Ivan: I am looking at using the photographic method to compare to my method, but I ideally want to derive clast metrics from laser scan point clouds. The issue I am having is in defining the boundaries of the clasts from a 3D point cloud, and then I will use the same watershed type methods once the boundaries are identified.

I agree almost all the comments, I would try to start with CANUPO, e-Cognition is a great option but is oriented to 2.5D data and is commercial, althoug I agree is a good option, however, if you are using point clouds I would use a specific software to edit and classify point clouds in a 3D environment.