Event Location: NSH 1507

Abstract: Maps are important for both human and robot navigation. Given a route, driving assistance systems consult maps to guide human drivers to their destinations. Similarly, topological maps of a 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 providing the necessary information about the driving environment.

The majority of existing cartographic databases are primarily built, through manual surveys and operator interactions, to assist human navigation. Because human involvement is expensive and error prone, such manual endeavors should be minimized. In addition, the resolution of maps is insufficient for use in robotics applications and their coverage fails to reach places where robotics applications require detailed geometric information.

This thesis investigates computer vision algorithms for automatically building lane-level detail maps of highways and parking lots by analyzing publicly available cartographic resources such as orthoimagery.

To successfully build highway and parking lot maps from orthoimagery analysis, it is imperative to extract image cues tightly coupled with the structure of the underlying road network. Extracting relevant image cues is challenging because the seemingly salient image patterns, such as high intensity whitish lane-markings along parallel lines or rectangular grids, are not readily available for image processing due to variations in their appearances. To address these variations, we scrutinize the prior information and local image patterns (e.g., lines and their spatial relations) to produce task-specific image cues (e.g., parking spots, road image-region segmentation). These task-specific image cues provide sufficient information to build a map of a given image whereas acquiring global information from hand-labeled examples is, most of the time, uncertain and expensive. We demonstrate the effectiveness and robustness of our bootstrapping approach in tackling the problem of building roadway maps through experiments.

Due to expected abnormal events on highways (e.g., such as road-work) the geometry and traffic rules of highways that appear on maps can occasionally change. This thesis also addresses the problem of updating the resulting map with temporary changes by analyzing perspective imagery acquired from a vision sensor installed on a vehicle.

To robustly recognize highway workzones, our sign recognizer focuses on handling variations of signs’ color and shape. Since errors in sign recognition will cause our system to misread temporary highway changes, to handle potential recognition errors, our method utilizes the temporal redundancy of sign occurrences and their corresponding classification decisions. Through testing, using video data recorded under various weather conditions, our approach was able to perfectly identify the boundaries of workzones and robustly detect a majority of driving condition changes.

Committee:Chris Urmson, Co-chair

David Wettergreen, Co-chair

Martial Hebert

John Krumm, Microsoft Research