Monocular Robot Mapping
With this project we hope to push the limits of what can be done with monocular vision, a rich and ubiquitous sensor for mobile robots. This work combines principles of robotics, particularly probabilistic robotics, with well-established results in computer vision, especially the subfield known as "structure from motion."
Mapping with a single camera is a compelling problem because we humans have no trouble wandering a new environment (even with one eye closed). What's more, we can subsequently use that experience to get around and perform tasks. In contrast to systems that use laser range finders, robots with a webcamera — in principle, at least — have the ability to create and use richer environmental maps at a fraction of the cost. The tradeoff, however, comes in processing power: the computational work of analyzing the input image stream is substantial. This is particularly true because our map representation is task-independent: it is simply a 3d texture-mapped model of the environment.
To handle the inevitable difficulties in the "low-level" visual processing—image segmentation, feature extraction, and matching—the 2009 REU project will investigate machine-learning approaches to these problems. The REU participants will likely create and compare both supervised and unsupervised algorithms. Our goal is to instantiate those algorithms in time for the robot exhibition and competition at IJCAI 2009 in Pasadena in July.
Previous Work
In 2008, Devin Smith (HMC '09) investigated image profiles, which are simply sums of pixel columns. They turn out to be surprisingly powerful, enabling a compass-from-camera application, as well as full visual odometry, without correspondence or point-based reconstruction. His results generalized and improved upon those in RatSLAM [2008]. Hanna Hoersting (HMC '09) and Lesley Bilitchenko (Cal Poly Pomona '10) built maps from image-profile sequences. The day-to-day journals of all three give a sense of REU progress: Hannah Devin Lesley. The resulting submissions (Lesley and Hannah Devin) have been accepted as a paper presentation and poster, respectively, at ACM's Symposium on Applied Computing 2009.
In 2007, S. Cord Melton (U. Chicago '09) and Lilia Markham (HMC '08) developed an end-to-end system for region-based visual reconstruction. Their wiki site contains many details, with Cord's and Lilia's journal entries giving a sense of what day-to-day REU experience is like. In addition, their paper was accepted and presented at the ISC '07 conference in Cambridge, MA.
Web summaries of previous HMC work in this area prior to 2007 are linked here: Summer '05 robotics, Summer '05 vision and 2004-5. Related papers with student authors include ICARA '04, AAAI Robot Competition '05, AAAI Robot Exhibition '06, ICRA 2007, ISC 2007.
Mentor: Professor Zach Dodds
Zach has been a professor
at Harvey Mudd since 1999. His general research interests are in
computer vision and robotics. He received his B.A. in mathematics
and M.S. and Ph.D. in computer science from Yale University. In
addition to research and teaching, he likes to play in foam pits
with his children.
Required Background
While background in computer vision and robotics/robotics algorithms is wonderful, it's not required. Students interested in working on this project should have completed at least a course in C++ programming and data structures and should feel confident in applying their programming ability through new frameworks and for new problems (e.g., vision & robotics...). We will use Python a lot, too, but it is so easy to learn that it's not worth worrying about beforehand.
