Vision-based Aerial 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.
We will be using the ARDrone quadcopter as the platform for our vision-based control and mapping efforts in 2012. In addition, we will be using ground platforms (the iRobot Create) with Microsoft's Kinect sensor. The aerial platforms are an example where the Kinect cannot completely replace vision: it is too heavy and power-hungry for the drone to support.
We certainly cannot hope to re-derive the amanzing progress in this field! Thus, the 2012 REU project will look into using the remarkable monocular slam work of Andrew Davison as a starting point and/or inspiration A concrete goal is to use an drone's fly-through of our indoor space to build a 3d model that, in turn, could support virtual fly-throughs of that environment.
Previous Work
In 2011, Brad Jensen '13, Lilian de Greef '12, Kim Sheely '12, Malen Sok '13, and Nick Berezny '12 developed software for our aerial platforms spanning from device drivers to vision-based localization routines using SURF features. They exhibited their work at GCER 2011 and AAAI 2011's robot exhibition and workshop. The team also attended RSS 2011 in Los Angeles and submitted their work to TePRA 2012 (we are still awaiting word on that submission).
In 2010, Nicole Lesperance and Michael Leece led a team of students in creating a top-to-bottom vision system that estimated a robot's local environment from its webcamera. Their vision system was published and presented in The 2010 International Symposium on Visual Computing, and a broader view of their perception work and its control of an iRobot Create was published and presented at the 2011's IEEE TePRA conference. Perhaps the highlight of the summer was the team's exhibit at the 2010 AAAI Robotics workshop and exhibition in Atlanta. In addition to exhibiting their autonomously navigating Create, the team took advantage of the venue to watch some of the first annual robot chess competition and several humanoid-robot obstacle course runs.
In 2009, Zeke Koziol, Sabreen Lakhani, and Anatole Paine implemented two systems that learned range from monocular images. One used feedforward neural networks trained by backpropagation; a second used a newer machine-learning technique known as Restricted Bolzmann Machines or Deep Belief Nets, depending on who's talking. Their work led to an exhibition at the 2009 IJCAI robot exhibition and a publication in its proceedings. In addition, their paper to IEEE's conference on Technologies for Practical Robot Applications was accepted and presented in November; Sabreen and Toli also competed in the 2009 Robotics Innovations Competition and Conference, winning the Neuron Robotics award (and cash prize) for the "Most elegant robotic solution."
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. (Full disclosure: all parties in this photograph are now much older!)
Required Background
While background in computer vision and/or robotics is wonderful, it's not required. The project will be tailored to the backgrounds and interests of the students involved. Particularly helpful are applicants who are confident with computation and eager to learn new technologies, languages, and interfaces as needed by the project. We will use Python a lot, interacting with Willow Garage's ROS infrastructure, along with whatever Andrew Davison's software requires, but our philosophy is that there's nothing like an interesting task to motivate learning new technologies!.
