Event Location: NSH 1109

Abstract: Maps are important for both human and robot navigation. Given a route, driving assistance systems brief maps to guide human drivers to their destinations. Similarly, topological maps of road network provide a robotic vehicle with information about where it can drive and what driving behaviors it should use. Maps simplify both manual and autonomous driving by only providing necessary information about the environment.

The majority of current map building techniques rely on human involvement. Human map building techniques require operator interactions and robot map building methods involve manual pre-driving target areas. Because human involvement is expensive and error prone, these map building techniques should be automated.

This thesis investigates a self-supervised framework for building maps from publicly available cartographic resources such as roadmaps and orthoimagery. Self-labeling methods analyze low-level image patterns to collect easy-to-detect parts (e.g., parking spots) of road structures (e.g., a parking lot) in orthoimagery. Part detectors learn object appearance models by using self-labeled examples and identify probable image regions of road structures. The framework seeks a configuration that optimally satisfies geometric and image constraints of part detection results in order to recover underlying road structures.

In initial work, we have developed heuristics for automatically collecting parking spot examples that are in turn used to train parking spot detectors. Due to an imperfect self-labeler, our parking spot detector was unable to correctly classify parking spots with unusual appearances. To improve performance, we manually collect a set of unusual parking spots that are incrementally used to reduce the false negative rate. While using the manually labeled examples, uncertainty sampling is employed to minimize the usage of manually labeled data.

We propose several improvements to our framework. Firstly, we will investigate a generative incremental learning method to further reduce the false negative rate by the minimal use of manually labeled data. Secondly, we will study the properties of self-labeling process to improve our self-labeler’s capability. Thirdly, we will devise an optimization method to effectively interpret part detection results. Lastly, we will validate the usefulness of our framework.

Committee:Christopher Urmson, Co-chair

David Wettergreen, Co-chair

Martial Hebert

John Krumm, Microsoft Research