| Abstract: |
| In this talk, we consider the problem of exploring and reconstructing an a priori unknown environment based on range (visual) data from a single, moving observer.
The observer is to roam around a piece of unknown domain and reconstruct it, using as few observations of the environment as possible.
We present a greedy and iterative algorithm for designing the observation locations based on observations made in the previous iteration. The choice of each new observation location is aided by a convolution neural network, which is trained with a suitable database as a prior to the unknown environment to be explored.
We will demonstrate the performance of the new algorithm for exploring realistic 2D and 3D settings. |
|