RoboGrads Student Lunch Seminar: Chris Beall


12:00 PM-1:00 PM on November 20th, 2014

Location: MiRC 102A

Title: Appearance-based Localization across Seasons in a Metric Map



Robust vehicle localization is a key enabling technology to make self-driving cars a reality. In this talk I will present ongoing work about appearance-based long-term outdoor localization, which is inherently robust to seasonal changes. This is a difficult task due to the changing appearance of visual landmarks across seasons and time of day. Our approach operates based on the premise that combining visual landmarks observed at different times of the year into a single metric map will yield better localization results than a map created from a single sequence alone. We integrate stereo imagery collected at two different times of the year into a unified 3D map, and use this as the basis for localization. A landmark visibility prediction framework is utilized to efficiently retrieve a small subset of landmarks and their feature descriptors from a database of millions of landmarks. The proposed approach is experimentally validated on a challenging sequence collected a year earlier.