Unmanned Aerial Vehicles (UAVs) are already broadly employed for 3D modeling of large objects such as trees and monuments via photogrammetry. The usual workflow includes two distinct steps: image acquisition with UAV and computationally demanding post-flight image processing. Insufficient feature overlaps across images is a common shortcoming in post-flight image processing resulting in the failure of 3D reconstruction. Here we propose a real-time control system that overcomes this limitation by targeting specific spatial locations for image acquisition thereby providing sufficient feature overlap. We initially benchmark several implementations of the Scale-Invariant Feature Transform (SIFT) feature identification algorithm to determine whether they allow real-time execution on the low-cost processing hardware embedded on the UAV. We then experimentally test our UAV platform in virtual and real-life environments. The presented architecture consistently decreases failures and improves the overall quality of 3D reconstructions.