UC Davis

Autonomous driving is a challenging task given the variety and complexity of drive sce- narios. Meta-learning, particularly few-shot learning, is a common approach to tackle a classification task with limited number of training data for each classes. In this project, we focus on a specific scenario of visually navigating vehicles along an unseen route through recognizing the waypoints with only a few available waypoint images and no need to re-train for a new course. To attack the targeted task, we proposed a hierarchical framework with two deep models in which the highlight is the StampNet, a few-shot learning architecture with a covariance estimator for recognizing waypoints. To train the models, we collected an associated indoor dataset with images along the routes in various buildings on UC Davis campus, and quantified the performance of the tests to explore the best hyperparameters. Additionally, to demonstrate the effectiveness of the proposed approach, the framework was implemented on a customized mini vehicle for an online test that run the models in real time to navigate the vehicle in unseen courses.