Traversability Learning from Aerial Images with Fully Convolutional Neural Networks

Traversability analysis is essential for ground robot operations because it allows the incorporation of knowledge about traversable and non-traversable terrains into the robot’s path 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. In the literature that concerns path planning from aerial images, we found two main approaches to generate traversability maps: terrain classification or heuristics to determine traversability. However, methods based on classification or handcrafted heuristics exhibit limitations with regard to the accuracy and precision of the traversability maps, quality of calculated paths and processing time. This work explores a new method to compute a traversability map from an aerial image in one forward pass through a fully convolutional neural network. The objective is to quickly generate good quality traversability maps for use in the path planning phase of navigation algorithms. We evaluate our proposal in comparison to other approaches available in the literature to demonstrate that it provides valuable information for a planner algorithm to compute safe and cost-efficient paths using only images.