Learning Semantic Traversability with Egocentric Video and Automated Annotation Strategy