Aprendizado de atravessabilidade em imagens aéreas usando redes totalmente convolucionais

Abstract

Traversability analysis is essential for ground robot operations because it allows the incorporation of knowledge about traversable and non-traversable terrain into the robot’s planning algorithm. It is possible to use aerial data to compute traversability maps for large regions and use them to assist in ground robot navigation. Most published methods to generate these maps from aerial images use terrain classification or heuristics to determine traversability. However, methods based on classification or handcrafted heuristics exhibit limitations with regard to output quality and processing time. To improve on these variables, this work explores a new method to compute a traversability map from an aerial image in one forward pass through a fully convolutional network. It is shown that this method can be employed to generate traversability maps in diverse environments and demonstrated that it provides useful information for a planner algorithm to compute safe paths.