How is the forest you are trying to analyse?
If it's a planted forest (regular tree distribution) you can try a fourier analysis to see the density of the trees and the direction of the rows. Then try to segment the forest area using a texture analysis (Haralick descriptors, wavelet analysis, etc...) and estimate the number of trees from the area and density. (see the articles of Olivier Regniers)
Another method I know for tree counting is the marked point processes. It gives good results, but it is harder to implement. See the work of Guillaume Perrin at INRIA Nice.
If the trees are similar, you might also try the template matching algorithm; it's already implemented in OpenCV, but it's a less robust approach.
Probably by searching on the internet, you can find more articles on this topic, forest detection and tree counting is an important application of the remote sensing techniques. I know well the work of the two researchers I mentioned above, so I can recommend their approaches.
may be you can find some idea too in https://varcity.ethz.ch/index.html and http://www.vision.ee.ethz.ch/~rhayko/...
How do these methods come here? He isn't about reconstructing 3D models from multi-view images, but tree counting...
"We process images of real cities automatically and efficiently to create parameterized and semantic 3D models, in which streets, buildings and vegetation are discriminated, and the number of floors, positions ..." right vegetation is not only tree