Visibility optimization for autonomous exploration and surveillance-evasion games

This dissertation considers problems involving line-of-sight visibility. In the exploration problem, the agent must efficiently map out a previously unknown environment, using as few sensor measurements as possible. For the surveillance-evasion game, the agent must always maintain visibility of a moving adversary. We first derive algorithms from a theoretical perspective. Although the resulting solutions are optimal, they are expensive to compute. We propose efficient approximations to the optimal solutions. At the expense of optimality, these approximations provide reasonable solutions that enable near real-time performance. We leverage state-of-the-art machine learning techniques to scale to scenarios that were not previously feasible. Level set functions allow for efficient computation of visibility and are a natural representation for input to convolutional neural networks.

Lastly, we consider the notion of the visibility of point clouds along rays. Using nearest neighbors along a set of randomly generated rays, we compute a signature tensor which encodes geometric and statistical information about the point cloud. The signature, in combination with a convolutional neural network, enables efficient and robust classification of point clouds. We present promising results in 2D and 3D.