In this work we propose a technique to compute optimal camera poses for image acquisition to be used for drone assisted automated surveys. We use this technique to create highly detailed 3D models of architectural buildings from multi-view reconstruction. Our reconstructed models are great for use in virtual reality (VR) environments since they exhibit good amount of detail that is useful for creating realistic virtual walkthroughs. Creating a good 3D reconstruction with a set of nadir looking images is difficult since the vertical surfaces of buildings are not captured very well and are therefore not reconstructed accurately. Acquisition of non-nadir images require avoiding obstacles around the structure. Our technique is based on mathematical optimisation, and is capable of calculating camera positions and orientations to maximally cover horizontal as well as vertical surface patches while avoiding obstacles around the building. We present a complete pipeline for a mostly automated and robust approach via a camera mounted quadcopter drone.
The video below shows our automated 3D reconstruction algorithm in action:
Code
https://github.com/ojaswa/DroneAutoCapture
Publication
O. Sharma, N. Arora, and H. Sagar, “Image acquisition for high quality architectural reconstruction”, in Graphics Interface, 2019.