Preliminary Exam Talks
Uncertainty in Vision-based SLAM
|
Abstract
Taking a sequence of images and computing from it the path of the camera and a map of the world is a problem that has been extensively studied by both the robotics and computer vision communities. Though the goals and priorities of camera-based SLAM (Simultaneous Localization and Mapping) on the robotics side and motion-from-video with 3D Reconstruction on the vision side were once different, as the study is maturing, the approaches are converging, and any solution to either problem borrows heavily from the other side.
Within the problem of SLAM, there are many sources of error, and the inability to track and correct this error is the primary cause of failure in reconstructing an accurate map of the world. Cameras in particular are problematic sensors. They are noisy, the data they capture requires significant processing to convert it into range data, and the uncertainty in the now-converted data varies significantly with sensor position. As a result, maps created using cameras for SLAM often have pockets of both low and high uncertainty. For my thesis work, I will be exploring methods for identifying uncertain regions and planning actions to improve them. In particular, I will be looking at the case where a robot has made a pass through the world, built a map, and then has the opportunity to plan and take a second pass through the world with the intent of maximizing utility with respect to entropy reduction and sensor cost.