Safe and Topographic Path Planning with Deep Neural Networks

Path planning with the topographical characteristics of real-world terrains has been investigated in the solution of problems in simulation systems and computer games. Path planning in real-world terrains involves finding a path with a low cost in terms of distance traveled according to the terrain relief. This work investigates how the computations of the $A^{\ast}$ algorithm have to be performed for an agent to obtain a hidden path from the view of an observer in such terrains with topography. This indicates that terrain features are important for obtaining a path with reduced topographical cost that permits visually hiding an agent from an observer in a given terrain location. Moreover, this work evaluates the use of deep neural networks in the approximation of the heuristic used by the pathfinding algorithm, optimizing the execution time and number of expanded nodes required to compute safe routes with lower topographic costs that are generated with the proposed Deep Safe and Topographic $A^{\ast}$ (DSTA*) algorithm.