Start Date: August 21, 2019
Christos Sevastopoulos
Traversability illustrates the difficulty of driving through a specific region and encompasses the suitability of the terrain for traverse based on its physical properties, such as slope and roughness, surface condition, etc. A key aspect of Self Supervised Learning that renders it as the contemporary most promising direction towards traversability estimation in unknown environments, is the ability to establish larger proportions of data efficiency in deep learning models that aim, as a consequence of reduced demand, for hand-labeled training data. This work’s primary focus is to present a unified indoor traversability estimation vision framework while leveraging the asset of transfer learning through contrastive pretraining.