give a cheap&dirty attempt like a simple photogrammetry software a try! Something like Photoscan Professional. Widely used, but as far as I know few attempts have been made to determine accuracy properly.

It really depends on the ground and extent of the scan. Cities have a lot of 3D information (with easy feature extraction) over a relatively flat ground, natural landscapes are the opposite.

I would generally try small baseline techniques for stable frame rate/ air speed and if that is not possible then would try some local feature based 3D estimation (I never did drone reconstruction, only normal cameras so I may be saying something stupid here...)

Anyway, look at the work from the guys from http://www.sensefly.com/ an EPFL spin-off, their results are great.

Evaluations of depth-from-stereo algorithms based on the middleburry dataset are listed at http://vision.middlebury.edu/~schar/stereo/web/results.php

The selection of the algorithm strongly depends on your data: occlusion, texture, ...

It is interesting for you to read the following paper:

Scharstein, D., Szeliski, R., 2002. A taxonomy and evaluation of dense two-frame

stereo sorrespondence algorithms. Internat. J. Comput. Vision 47 (1–3), 7–42.

Certainly, It will you help to get more on different approaches proposed for 3D dense stereo methods.

The link to get to the paper:

http://www.google.co.ma/url?sa=t&rct=j&q=a%20taxonomy%20and%20evaluation%20of%20dense%20two-frame&source=web&cd=1&cad=rja&ved=0CC8QFjAA&url=http%3A%2F%2Fvision.middlebury.edu%2Fstereo%2Ftaxonomy-IJCV.pdf&ei=10KNUdfKCML40gXm8IDIBg&usg=AFQjCNHjzjN5VAmGMl9oafwlDxGDBjshcQ&bvm=bv.46340616,d.d2k

There are several algorithms and software packages that may work for this, but as mentioned above it depends on factors such as the terrain you're working with (for example urban versus fields and other landscapes) and the baseline-to-depth ratio of your imagery. I assume you need a dense stereo method, for which occlusions, moving objects, etc. may cause issues. You can give a general system like VisualSfM a try (http://homes.cs.washington.edu/~ccwu/vsfm/), whose output is a quasi-dense reconstruction, but that may leave holes in areas with less texture where patch matching may become difficult. In my experience, it seems to work pretty well on many different types of terrain and it's freely available, so I would try that first. I've also seen some great results and dense point clouds obtained with Pix4D (http://pix4d.com/). Also, take a look at some of the work being done by the Aerial Vision Group at TU Graz (http://aerial.icg.tugraz.at/). I hope this helps.

You can use conventional 3D Stereo Vision Matching image-video processing Algorithms based on your 3D Stereo Vision system by knowing your stereo focus, and build your own 3D Stereo Geometry model by using triangular 3D Reconstruction or polyfit of multipoints-pixels for 3D Surface Reconstruction for your aerial video sequences. I believe you are using a UAV with a 3D Stereo Vision Camera-sensor and you would like to apply stereo image-video interpolation to generate virtual 3D surface reconstruction (something similar to our project) with multi-sensors.

Sincere thanks to all friends for their feedback!